Merge remote-tracking branch 'origin/main' into bb/contrib-areas

# Conflicts:
#	apps/desktop/src/app/session/hooks/use-prompt-actions/index.ts

agent/auxiliary_client.py

@@ -102,7 +102,6 @@ OpenAI = _OpenAIProxy()  # module-level name, resolves lazily on call/isinstance
 
 from agent.credential_pool import load_pool
 from agent.model_metadata import MINIMUM_CONTEXT_LENGTH, get_model_context_length
-from agent.process_bootstrap import build_keepalive_http_client
 from hermes_cli.config import get_hermes_home
 from hermes_constants import OPENROUTER_BASE_URL
 from utils import base_url_host_matches, base_url_hostname, env_float, model_forces_max_completion_tokens, normalize_proxy_env_vars
@@ -156,22 +155,47 @@ def _resolve_aux_verify(base_url: Optional[str]) -> Any:
         return True
 
 
+_WARNED_KEEPALIVE_IMPORT_SKEW = False
+
+
 def _openai_http_client_kwargs(
     base_url: Optional[str],
     *,
     async_mode: bool = False,
 ) -> Dict[str, Any]:
     """Inject keepalive httpx client with env-only proxy (not macOS system proxy)."""
-    client = build_keepalive_http_client(
-        str(base_url or ""),
-        async_mode=async_mode,
-        verify=_resolve_aux_verify(base_url),
-    )
+    try:
+        from agent.process_bootstrap import build_keepalive_http_client
+        client = build_keepalive_http_client(
+            str(base_url or ""),
+            async_mode=async_mode,
+            verify=_resolve_aux_verify(base_url),
+        )
+    except (ImportError, AttributeError):
+        # Version-skewed installs (#64333): a process whose sys.path resolves
+        # an older agent/process_bootstrap.py without this helper — seen when
+        # the Desktop app's bundled runtime lags a git-installed source tree
+        # that newer callers (cron scheduler) were written against. Every cron
+        # job died on this ImportError before any agent logic ran. Degrade
+        # gracefully to the OpenAI SDK's default httpx client (respects macOS
+        # system proxy, no pool-level keepalive expiry) instead of failing the
+        # whole job, and say so once — silent version skew is how this bug
+        # went unnoticed until jobs were already dead on arrival.
+        global _WARNED_KEEPALIVE_IMPORT_SKEW
+        if not _WARNED_KEEPALIVE_IMPORT_SKEW:
+            _WARNED_KEEPALIVE_IMPORT_SKEW = True
+            logger.warning(
+                "agent.process_bootstrap.build_keepalive_http_client is "
+                "unavailable — mixed/stale install detected (#64333). Falling "
+                "back to the SDK default HTTP client. Run `hermes update` (or "
+                "reinstall the Desktop app) to resync the runtime."
+            )
+        client = None
+
     if client is None:
         return {}
     return {"http_client": client}
 
-
 def _create_openai_client(*, api_key: str, base_url: str, **kwargs: Any) -> Any:
     kwargs = {**_openai_http_client_kwargs(base_url), **kwargs}
     # Hermes owns auxiliary retry + provider/model fallback policy (the
@@ -1337,6 +1361,31 @@ class _AnthropicCompletionsAdapter:
             if not _forbids_sampling_params(model):
                 anthropic_kwargs["temperature"] = temperature
 
+        # Pass through caller-supplied extra_body so providers behind
+        # Anthropic-compatible gateways receive their per-vendor request
+        # fields (thinking control, metadata, portal tags, ...). The dict
+        # form is the documented Anthropic SDK passthrough for non-standard
+        # request body keys; merge on top of whatever build_anthropic_kwargs
+        # already produced (e.g. fast-mode ``speed``) so call-time settings
+        # survive. Two exclusions:
+        #   - ``reasoning``: the OpenAI-shaped config dict is TRANSLATED into
+        #     the native ``thinking`` field above (build_anthropic_kwargs);
+        #     forwarding the raw field alongside would double-specify
+        #     reasoning and 400 on strict gateways.
+        #   - ``_``-prefixed keys: private Hermes plumbing (_reasoning_config
+        #     et al.), never wire fields.
+        caller_extra_body = kwargs.get("extra_body")
+        if caller_extra_body and isinstance(caller_extra_body, dict):
+            passthrough = {
+                k: v for k, v in caller_extra_body.items()
+                if k != "reasoning" and not str(k).startswith("_")
+            }
+            if passthrough:
+                existing = anthropic_kwargs.get("extra_body") or {}
+                if not isinstance(existing, dict):
+                    existing = {}
+                anthropic_kwargs["extra_body"] = {**existing, **passthrough}
+
         response = create_anthropic_message(self._client, anthropic_kwargs)
         _transport = get_transport("anthropic_messages")
         _nr = _transport.normalize_response(

agent/bedrock_adapter.py

@@ -528,10 +528,19 @@ def _convert_content_to_converse(content) -> List[Dict]:
                         mime_part = header[5:].split(";")[0]
                         if mime_part:
                             media_type = mime_part
+                    # Decode base64 to raw bytes — boto3 re-encodes at the
+                    # wire layer, so passing the base64 string directly
+                    # results in double-encoding and Bedrock rejects it with
+                    # "Failed to sanitize image".  Ref: #33317.
+                    import base64
+                    try:
+                        raw_bytes = base64.b64decode(data)
+                    except Exception:
+                        raw_bytes = data.encode("utf-8")
                     blocks.append({
                         "image": {
                             "format": media_type.split("/")[-1] if "/" in media_type else "jpeg",
-                            "source": {"bytes": data},
+                            "source": {"bytes": raw_bytes},
                         }
                     })
                 else:

agent/chat_completion_helpers.py

@@ -504,6 +504,29 @@ def interruptible_api_call(agent, api_kwargs: dict):
         if _codex_floor:
             _stale_timeout = max(_stale_timeout, _codex_floor)
 
+    # ── Codex absolute hard ceiling (#64507) ──────────────────────────
+    # ``openai_codex_stale_timeout_floor`` *raises* the stale timeout (up to
+    # 1200s at >100k tokens) so healthy gateway-scale payloads aren't aborted.
+    # The scaled no-byte TTFB watchdog catches dead streams that never emit a
+    # first byte, but a request that emits SOME bytes and then wedges (the
+    # issue-64507 symptom: vision-inflated request, worker idle, no ended_at)
+    # is only reclaimed at the (high) stale floor. Add a flat, finite hard
+    # ceiling on total request time that ALWAYS applies to openai-codex
+    # requests regardless of the TTFB/stale interaction, so a stalled request
+    # is recovered (retry loop / visible failure) instead of hanging
+    # indefinitely. The default sits ABOVE the maximum stale floor (1200s) so
+    # it never clamps an intentionally-raised timeout for healthy large
+    # requests — it is a backstop against unbounded growth, not a tighter
+    # limit. Tunable via HERMES_CODEX_HARD_TIMEOUT_SECONDS (set to 0 to
+    # disable the ceiling entirely; that restores the pre-fix behavior).
+    _codex_hard_timeout = _env_float("HERMES_CODEX_HARD_TIMEOUT_SECONDS", 1500.0)
+    if (
+        _codex_watchdog_enabled
+        and _openai_codex_backend
+        and _codex_hard_timeout > 0
+    ):
+        _stale_timeout = min(_stale_timeout, _codex_hard_timeout)
+
     if _est_tokens_for_codex_watchdog > 100_000:
         _codex_idle_timeout_default = 180.0
     elif _est_tokens_for_codex_watchdog > 50_000:

agent/codex_responses_adapter.py

@@ -1118,6 +1118,22 @@ def _normalize_codex_response(
     differs from the one that minted the encrypted_content blob and drop
     the item instead of triggering HTTP 400 invalid_encrypted_content.
     """
+    response_status = getattr(response, "status", None)
+    if isinstance(response_status, str):
+        response_status = response_status.strip().lower()
+    else:
+        response_status = None
+
+    incomplete_details = getattr(response, "incomplete_details", None)
+    incomplete_reason = ""
+    if isinstance(incomplete_details, dict):
+        incomplete_reason = str(incomplete_details.get("reason") or "").strip().lower()
+    elif incomplete_details is not None:
+        incomplete_reason = str(getattr(incomplete_details, "reason", "") or "").strip().lower()
+    response_incomplete_content_filter = (
+        response_status == "incomplete" and incomplete_reason == "content_filter"
+    )
+
     output = getattr(response, "output", None)
     if not isinstance(output, list) or not output:
         # The Codex backend can return empty output when the answer was
@@ -1134,15 +1150,18 @@ def _normalize_codex_response(
                 content=[SimpleNamespace(type="output_text", text=out_text.strip())],
             )]
             response.output = output
+        elif response_incomplete_content_filter:
+            # This is a deterministic provider safety block, not a partial
+            # answer. Synthesize an empty message so finish_reason below becomes
+            # content_filter and the conversation loop can fallback/surface it
+            # instead of burning three continuation attempts.
+            output = [SimpleNamespace(
+                type="message", role="assistant", status="completed", content=[]
+            )]
+            response.output = output
         else:
             raise RuntimeError("Responses API returned no output items")
 
-    response_status = getattr(response, "status", None)
-    if isinstance(response_status, str):
-        response_status = response_status.strip().lower()
-    else:
-        response_status = None
-
     if response_status in {"failed", "cancelled"}:
         error_obj = getattr(response, "error", None)
         error_msg = _format_responses_error(error_obj, response_status)
@@ -1411,6 +1430,8 @@ def _normalize_codex_response(
 
     if tool_calls:
         finish_reason = "tool_calls"
+    elif response_incomplete_content_filter:
+        finish_reason = "content_filter"
     elif leaked_tool_call_text:
         finish_reason = "incomplete"
     elif saw_streaming_or_item_incomplete:

agent/context_compressor.py

@@ -586,6 +586,27 @@ def _strip_historical_media(messages: List[Dict[str, Any]]) -> List[Dict[str, An
     return result if changed else messages
 
 
+def _image_part_label(part: Dict[str, Any]) -> str:
+    """Render a multimodal image part as a short text label for the summarizer.
+
+    Keeps a real, referenceable URL when the image lives at an http(s)
+    address — the summary can then preserve the handle so the agent (or a
+    later vision_analyze call) can still reach the image after compaction.
+    Base64 ``data:`` URLs carry no reusable reference and would flood the
+    summarizer input, so they collapse to ``[image]``.
+    """
+    url = ""
+    if isinstance(part.get("image_url"), dict):
+        url = str(part["image_url"].get("url") or "")
+    elif isinstance(part.get("image_url"), str):
+        url = part["image_url"]
+    elif isinstance(part.get("url"), str):
+        url = part["url"]
+    if url.startswith(("http://", "https://")):
+        return f"[image: {url}]"
+    return "[image]"
+
+
 def _str_arg(args: dict, key: str, default: str = "") -> str:
     """Safely get a string argument from parsed tool args.
 
@@ -686,7 +707,16 @@ def _summarize_tool_result_unguarded(tool_name: str, tool_args: str, tool_conten
 
     if tool_name == "web_extract":
         urls = args.get("urls", [])
-        url_desc = urls[0] if isinstance(urls, list) and urls else "?"
+        first = urls[0] if isinstance(urls, list) and urls else "?"
+        # web_search results are dicts ({"url"/"href": ...}) and models often
+        # forward them straight into web_extract. Unwrap to the URL string so
+        # the summary stays readable and the ``+=`` below never hits the
+        # ``dict + str`` TypeError that would abort pre-compression pruning.
+        if isinstance(first, dict):
+            first = first.get("url") or first.get("href") or "?"
+        elif not isinstance(first, str):
+            first = "?"
+        url_desc = first
         if isinstance(urls, list) and len(urls) > 1:
             url_desc += f" (+{len(urls) - 1} more)"
         return f"[web_extract] {url_desc} ({content_len:,} chars)"
@@ -1655,7 +1685,24 @@ class ContextCompressor(ContextEngine):
         parts = []
         for msg in turns:
             role = msg.get("role", "unknown")
-            content = redact_sensitive_text(msg.get("content") or "")
+            content = msg.get("content")
+            if isinstance(content, list):
+                text_parts: list[str] = []
+                for part in content:
+                    if isinstance(part, dict):
+                        ptype = part.get("type")
+                        if ptype == "text":
+                            text_parts.append(part.get("text", ""))
+                        elif ptype in {"image", "image_url", "input_image"}:
+                            text_parts.append(_image_part_label(part))
+                        else:
+                            # Unknown part type — keep a marker so the
+                            # summarizer knows content existed here.
+                            text_parts.append(f"[{ptype or 'attachment'}]")
+                    elif isinstance(part, str):
+                        text_parts.append(part)
+                content = "\n".join(text_parts)
+            content = redact_sensitive_text(content or "")
             content = _MEDIA_DIRECTIVE_RE.sub("[media attachment]", content)
             # Strip inline reasoning blocks (<think>, <reasoning>, etc.) from
             # assistant content before it reaches the summarizer. Reasoning

agent/conversation_loop.py

@@ -1642,12 +1642,16 @@ def run_conversation(
                 # Check finish_reason before proceeding
                 if agent.api_mode == "codex_responses":
                     status = getattr(response, "status", None)
+                    if isinstance(status, str):
+                        status = status.strip().lower()
                     incomplete_details = getattr(response, "incomplete_details", None)
                     incomplete_reason = None
                     if isinstance(incomplete_details, dict):
                         incomplete_reason = incomplete_details.get("reason")
                     else:
                         incomplete_reason = getattr(incomplete_details, "reason", None)
+                    if incomplete_reason is not None:
+                        incomplete_reason = str(incomplete_reason).strip().lower()
                     if status == "incomplete" and incomplete_reason in {"max_output_tokens", "length"}:
                         # Responses API max-output exhaustion is a normal
                         # Codex incomplete turn.  Let the Codex-specific
@@ -1657,6 +1661,8 @@ def run_conversation(
                         # emits "Response truncated due to output length
                         # limit" and stops gateway turns.
                         finish_reason = "incomplete"
+                    elif status == "incomplete" and incomplete_reason == "content_filter":
+                        finish_reason = "content_filter"
                     else:
                         finish_reason = "stop"
                 elif agent.api_mode == "anthropic_messages":
@@ -2619,6 +2625,31 @@ def run_conversation(
                     )
                     continue
 
+                # ── Bedrock AnthropicBedrock SDK streaming failure ──
+                # The Anthropic SDK's stream accumulator raises RuntimeError
+                # "Unexpected event order" when Bedrock returns an error event
+                # before message_start (throttling, overload, validation).
+                # Fall back to the native Converse API path for the rest of
+                # this session — it handles these errors gracefully.  Ref: #28156.
+                if (
+                    isinstance(api_error, RuntimeError)
+                    and "unexpected event order" in str(api_error).lower()
+                    and getattr(agent, "provider", "") == "bedrock"
+                    and agent.api_mode == "anthropic_messages"
+                    and not getattr(agent, "_bedrock_converse_fallback_attempted", False)
+                ):
+                    agent._bedrock_converse_fallback_attempted = True
+                    agent.api_mode = "bedrock_converse"
+                    agent._bedrock_region = getattr(agent, "_bedrock_region", None) or "us-east-1"
+                    agent.client = None  # Drop the AnthropicBedrock client
+                    agent._client_kwargs = {}
+                    agent._vprint(
+                        f"{agent.log_prefix}⚠️  AnthropicBedrock SDK streaming failed — "
+                        f"falling back to native Converse API for this session.",
+                        force=True,
+                    )
+                    continue
+
                 status_code = getattr(api_error, "status_code", None)
                 error_context = agent._extract_api_error_context(api_error)
 
@@ -4459,27 +4490,28 @@ def run_conversation(
                     or interim_has_codex_message_items
                 ):
                     last_msg = messages[-1] if messages else None
-                    # Duplicate detection: two consecutive incomplete assistant
-                    # messages with identical content AND reasoning are collapsed.
-                    # For provider-state-only changes (encrypted reasoning
-                    # items or replayable message ids/phases/statuses differ
-                    # while visible content/reasoning are unchanged), compare
-                    # those opaque payloads too so we don't silently drop the
-                    # newer continuation state.
-                    last_codex_items = last_msg.get("codex_reasoning_items") if isinstance(last_msg, dict) else None
-                    interim_codex_items = interim_msg.get("codex_reasoning_items")
-                    last_codex_message_items = last_msg.get("codex_message_items") if isinstance(last_msg, dict) else None
-                    interim_codex_message_items = interim_msg.get("codex_message_items")
-                    duplicate_interim = (
+                    # Duplicate detection: compare only visible content
+                    # (content + reasoning).  Opaque provider state
+                    # (encrypted reasoning items, message item ids/phases)
+                    # drifts per continuation even when the visible output
+                    # is identical, so including it in the comparison defeats
+                    # dedup and causes message storms (#52711).
+                    visible_duplicate = (
                         isinstance(last_msg, dict)
                         and last_msg.get("role") == "assistant"
                         and last_msg.get("finish_reason") == "incomplete"
                         and (last_msg.get("content") or "") == (interim_msg.get("content") or "")
                         and (last_msg.get("reasoning") or "") == (interim_msg.get("reasoning") or "")
-                        and last_codex_items == interim_codex_items
-                        and last_codex_message_items == interim_codex_message_items
                     )
-                    if not duplicate_interim:
+                    if visible_duplicate:
+                        # Update opaque state in-place so the latest
+                        # provider payload is preserved without emitting
+                        # a duplicate visible message.
+                        for _key in ("codex_reasoning_items", "codex_message_items"):
+                            _new_val = interim_msg.get(_key)
+                            if _new_val is not None:
+                                last_msg[_key] = _new_val
+                    else:
                         messages.append(interim_msg)
                         agent._emit_interim_assistant_message(interim_msg)
 

agent/transports/codex.py

@@ -435,15 +435,27 @@ class ResponsesApiTransport(ProviderTransport):
     def validate_response(self, response: Any) -> bool:
         """Check Codex Responses API response has valid output structure.
 
-        Returns True only if response.output is a non-empty list.
-        Does NOT check output_text fallback — the caller handles that
-        with diagnostic logging for stream backfill recovery.
+        Returns True only if response.output is a non-empty list. Also treats
+        terminal content-filter incomplete responses as valid: the Responses API
+        may return status=incomplete with incomplete_details.reason='content_filter'
+        and no output items. That is a provider refusal signal, not a malformed
+        response, and must reach normalization so the agent loop can use the
+        content-policy / fallback path instead of invalid-response retries.
+
+        Does NOT check output_text fallback — the caller handles that with
+        diagnostic logging for stream backfill recovery.
         """
         if response is None:
             return False
         output = getattr(response, "output", None)
         if not isinstance(output, list) or not output:
-            return False
+            status = str(getattr(response, "status", "") or "").strip().lower()
+            incomplete_details = getattr(response, "incomplete_details", None)
+            if isinstance(incomplete_details, dict):
+                reason = str(incomplete_details.get("reason") or "").strip().lower()
+            else:
+                reason = str(getattr(incomplete_details, "reason", "") or "").strip().lower()
+            return status == "incomplete" and reason == "content_filter"
         return True
 
     def preflight_kwargs(

apps/desktop/electron/git-worktree-ops.test.ts

@@ -262,3 +262,40 @@ test('addWorktree: base param branches off a specified local branch', async () =
     fs.rmSync(dir, { recursive: true, force: true })
   }
 })
+
+test('addWorktree: base origin/main does not set up upstream tracking', async () => {
+  // Two repos: a bare "remote" and a clone, so origin/main resolves as a
+  // remote-tracking ref — the condition that triggers auto-tracking.
+  const remoteDir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-remote-'))
+  const cloneDir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-clone-'))
+  const git = (...args) => execFileSync('git', args, { cwd: cloneDir }).toString().trim()
+
+  try {
+    // Seed the remote with a commit on main. Inline identity so it works
+    // on CI runners with no global git config.
+    execFileSync('git', ['init', '-b', 'main', remoteDir])
+    execFileSync('git', ['-C', remoteDir, '-c', 'user.email=hermes@localhost', '-c', 'user.name=Hermes', 'commit', '--allow-empty', '-m', 'root'])
+
+    // Clone so origin/main exists as a remote-tracking ref.
+    execFileSync('git', ['clone', remoteDir, cloneDir])
+
+    const result = await addWorktree(cloneDir, { base: 'origin/main', branch: 'feature-branch', name: 'feature-branch' }, 'git')
+
+    assert.equal(result.branch, 'feature-branch')
+
+    // The new branch must NOT have an upstream — like `git checkout origin/main
+    // && git checkout -b feature-branch`, not `git worktree add -b … origin/main`.
+    let hasUpstream = true
+
+    try {
+      execFileSync('git', ['-C', result.path, 'rev-parse', '--abbrev-ref', '--symbolic-full-name', '@{u}'])
+    } catch {
+      hasUpstream = false
+    }
+
+    assert.equal(hasUpstream, false)
+  } finally {
+    fs.rmSync(remoteDir, { recursive: true, force: true })
+    fs.rmSync(cloneDir, { recursive: true, force: true })
+  }
+})

apps/desktop/electron/git-worktree-ops.ts

@@ -265,8 +265,15 @@ async function addWorktree(repoPath, options, gitBin) {
       } catch {
         // The fetch isn't mandatory, but it would be nice to do if possible.
         // If it's not possible, just use the local ref of the remote branch.
-        // If it doesn't exist locally, we'll get an error anyways 
+        // If it doesn't exist locally, we'll get an error
       }
+
+      // When branching off a remote-tracking ref, git auto-sets up tracking
+      // (e.g. `new-branch` → tracks `origin/main`). The user almost certainly
+      // wants a standalone local branch — like `git checkout origin/main &&
+      // git checkout -b new-branch` — not a branch silently wired to the
+      // remote's upstream. `--no-track` prevents that.
+      args.push('--no-track')
     }
 
     args.push(base)

apps/desktop/src/app/chat/composer/index.tsx

@@ -7,6 +7,7 @@ import { Button } from '@/components/ui/button'
 import { Slot as ContribSlot } from '@/contrib/react/slot'
 import { useI18n } from '@/i18n'
 import { chatMessageText } from '@/lib/chat-messages'
+import { sanitizeComposerInput } from '@/lib/composer-input-sanitize'
 import { DATA_IMAGE_URL_RE } from '@/lib/embedded-images'
 import { triggerHaptic } from '@/lib/haptics'
 import { cn } from '@/lib/utils'
@@ -285,7 +286,7 @@ export function ChatBar({
 
     normalizeComposerEditorDom(editor)
 
-    const nextDraft = composerPlainText(editor)
+    const nextDraft = sanitizeComposerInput(composerPlainText(editor))
 
     if (nextDraft !== draftRef.current) {
       draftRef.current = nextDraft
@@ -350,7 +351,7 @@ export function ChatBar({
     // blank lines (common when selecting from terminals, code blocks, web pages)
     // doesn't dump multiline padding into the composer. Internal newlines are
     // preserved — only the edges are cleaned up.
-    const pastedText = event.clipboardData.getData('text').trim()
+    const pastedText = sanitizeComposerInput(event.clipboardData.getData('text').trim())
 
     if (!pastedText) {
       event.preventDefault()

apps/desktop/src/app/right-sidebar/review/file-tree.tsx

@@ -218,6 +218,7 @@ function ReviewDirRow({
         <span className="min-w-0 flex-1 truncate" title={node.name}>
           {node.name}
         </span>
+        {!open && <DiffCount added={node.added} className="text-[0.64rem] leading-4" removed={node.removed} />}
       </div>
       {open && node.children && (
         <ReviewNodeList animate={animate} depth={depth + 1} motion={useMotion} nodes={node.children} />

apps/desktop/src/app/session/hooks/use-prompt-actions/index.ts

@@ -7,6 +7,7 @@ import { useI18n } from '@/i18n'
 import { stripAnsi } from '@/lib/ansi'
 import { type ChatMessage, textPart } from '@/lib/chat-messages'
 import { pathLabel, SLASH_COMMAND_RE } from '@/lib/chat-runtime'
+import { sanitizeComposerInput } from '@/lib/composer-input-sanitize'
 import { triggerHaptic } from '@/lib/haptics'
 import { setMutableRef } from '@/lib/mutable-ref'
 import { normalize } from '@/lib/text'
@@ -477,7 +478,7 @@ export function usePromptActions({
 
   const submitText = useCallback(
     async (rawText: string, options?: SubmitTextOptions) => {
-      const visibleText = rawText.trim()
+      const visibleText = sanitizeComposerInput(rawText).trim()
       const attachments = options?.attachments ?? $composerAttachments.get()
 
       if (!attachments.length && SLASH_COMMAND_RE.test(visibleText)) {
@@ -597,7 +598,7 @@ export function usePromptActions({
   // window) so the caller can fall back to queueing the words for the next turn.
   const steerPrompt = useCallback(
     async (rawText: string): Promise<boolean> => {
-      const text = rawText.trim()
+      const text = sanitizeComposerInput(rawText).trim()
       const sessionId = activeSessionId || activeSessionIdRef.current
 
       if (!text || !sessionId) {

apps/desktop/src/app/session/hooks/use-prompt-actions/submit.ts

@@ -4,6 +4,7 @@ import { PROMPT_SUBMIT_REQUEST_TIMEOUT_MS } from '@/hermes'
 import type { Translations } from '@/i18n'
 import { type ChatMessage, textPart } from '@/lib/chat-messages'
 import { optimisticAttachmentRef } from '@/lib/chat-runtime'
+import { sanitizeComposerInput } from '@/lib/composer-input-sanitize'
 import { setMutableRef } from '@/lib/mutable-ref'
 import {
   $composerAttachments,
@@ -88,7 +89,7 @@ export function useSubmitPrompt(deps: SubmitPromptDeps) {
 
   return useCallback(
     async (rawText: string, options?: SubmitTextOptions) => {
-      const visibleText = rawText.trim()
+      const visibleText = sanitizeComposerInput(rawText).trim()
       const usingComposerAttachments = !options?.attachments
 
       // Drop undefined/null holes a session switch or draft restore can leave in

apps/desktop/src/app/settings/model-settings.test.tsx

@@ -186,3 +186,139 @@ describe('ModelSettings', () => {
     expect(await screen.findByText(/still run on/)).toBeTruthy()
   })
 })
+
+describe('ModelSettings MoA preset editor', () => {
+  const moaConfig = () => ({
+    default_preset: 'default',
+    active_preset: '',
+    presets: {
+      default: {
+        reference_models: [
+          { provider: 'nous', model: 'hermes-4' },
+          { provider: 'openrouter', model: 'deepseek/deepseek-v4-pro' }
+        ],
+        aggregator: { provider: 'openrouter', model: 'anthropic/claude-opus-4.8' },
+        reference_temperature: 0,
+        aggregator_temperature: 0,
+        max_tokens: 4096,
+        enabled: true
+      }
+    },
+    reference_models: [
+      { provider: 'nous', model: 'hermes-4' },
+      { provider: 'openrouter', model: 'deepseek/deepseek-v4-pro' }
+    ],
+    aggregator: { provider: 'openrouter', model: 'anthropic/claude-opus-4.8' },
+    reference_temperature: 0,
+    aggregator_temperature: 0,
+    max_tokens: 4096,
+    enabled: true
+  })
+
+  beforeEach(() => {
+    getGlobalModelOptions.mockResolvedValue({
+      providers: [
+        {
+          name: 'Nous',
+          slug: 'nous',
+          models: ['hermes-4', 'hermes-4-mini'],
+          authenticated: true,
+          capabilities: { 'hermes-4': { reasoning: true, fast: true } }
+        },
+        {
+          name: 'OpenRouter',
+          slug: 'openrouter',
+          models: ['deepseek/deepseek-v4-pro', 'anthropic/claude-opus-4.8'],
+          authenticated: true
+        }
+      ]
+    })
+    getMoaModels.mockResolvedValue(moaConfig())
+    saveMoaModels.mockImplementation((body: unknown) => Promise.resolve(body))
+  })
+
+  async function openReferenceEditor() {
+    await renderModelSettings()
+    expect(await screen.findByText('Reference 1')).toBeTruthy()
+  }
+
+  function slotSelects() {
+    // Combobox order in the MoA section (last 7 on the page): preset select,
+    // then provider+model per reference (2 refs), then aggregator
+    // provider+model. Reference 1's pair is therefore at -6 / -5.
+    const all = screen.getAllByRole('combobox')
+
+    return { ref1Provider: all.at(-6)!, ref1Model: all.at(-5)! }
+  }
+
+  it('holds the autosave while a slot is half-filled (provider changed, model pending)', async () => {
+    vi.useFakeTimers({ shouldAdvanceTime: true })
+
+    try {
+      await openReferenceEditor()
+
+      fireEvent.click(slotSelects().ref1Provider)
+      fireEvent.click(await screen.findByRole('option', { name: 'OpenRouter' }))
+
+      // Model was cleared by the provider change → config incomplete → the
+      // debounced autosave must NOT fire, even well past the 600ms window.
+      await vi.advanceTimersByTimeAsync(2000)
+      expect(saveMoaModels).not.toHaveBeenCalled()
+    } finally {
+      vi.useRealTimers()
+    }
+  })
+
+  it('saves once the model pick completes the slot', async () => {
+    vi.useFakeTimers({ shouldAdvanceTime: true })
+
+    try {
+      await openReferenceEditor()
+
+      fireEvent.click(slotSelects().ref1Provider)
+      fireEvent.click(await screen.findByRole('option', { name: 'OpenRouter' }))
+      await vi.advanceTimersByTimeAsync(700)
+
+      fireEvent.click(slotSelects().ref1Model)
+      fireEvent.click(await screen.findByRole('option', { name: 'anthropic/claude-opus-4.8' }))
+      await vi.advanceTimersByTimeAsync(700)
+
+      expect(saveMoaModels).toHaveBeenCalledTimes(1)
+      const sent = saveMoaModels.mock.calls[0][0] as ReturnType<typeof moaConfig>
+      expect(sent.presets.default.reference_models[0]).toEqual({
+        provider: 'openrouter',
+        model: 'anthropic/claude-opus-4.8'
+      })
+      // The untouched slots ride along unchanged — nothing reverts to defaults.
+      expect(sent.presets.default.reference_models[1]).toEqual({
+        provider: 'openrouter',
+        model: 'deepseek/deepseek-v4-pro'
+      })
+      expect(sent.presets.default.aggregator).toEqual({
+        provider: 'openrouter',
+        model: 'anthropic/claude-opus-4.8'
+      })
+    } finally {
+      vi.useRealTimers()
+    }
+  })
+
+  it('does not clear the model or save when the same provider is re-selected', async () => {
+    vi.useFakeTimers({ shouldAdvanceTime: true })
+
+    try {
+      await openReferenceEditor()
+
+      fireEvent.click(slotSelects().ref1Provider)
+      fireEvent.click(await screen.findByRole('option', { name: 'Nous' }))
+      await vi.advanceTimersByTimeAsync(700)
+
+      // Radix treats re-picking the current value as a no-op (no
+      // onValueChange), so nothing changes: no save, model still shown.
+      expect(saveMoaModels).not.toHaveBeenCalled()
+      expect(screen.getByText('nous · hermes-4')).toBeTruthy()
+    } finally {
+      vi.useRealTimers()
+    }
+  })
+})

apps/desktop/src/app/settings/model-settings.tsx

@@ -130,6 +130,24 @@ const NO_PROVIDERS: readonly ModelOptionProvider[] = [{ name: '—', slug: '', m
 export const withActive = (models: readonly string[], active: string): readonly string[] =>
   active && !models.includes(active) ? [active, ...models] : models
 
+// A slot is complete when both halves are chosen. Changing a slot's provider
+// intentionally clears its model (see updateMoaSlot), so every provider change
+// passes through an incomplete state while the user picks the new model.
+export const moaSlotComplete = (slot: MoaModelSlot): boolean => !!(slot.provider.trim() && slot.model.trim())
+
+// True when every slot in every preset is fully specified — the only state
+// that is safe to persist. The backend rejects configs with half-filled slots
+// (HTTP 422) instead of silently swapping the preset for hardcoded defaults
+// (#64156), so the autosave must simply wait for the edit to finish rather
+// than trying to "repair" the payload.
+export const moaConfigComplete = (config: MoaConfigResponse): boolean =>
+  Object.values(config.presets).every(
+    preset =>
+      preset.reference_models.length > 0 &&
+      preset.reference_models.every(moaSlotComplete) &&
+      moaSlotComplete(preset.aggregator)
+  )
+
 interface StaleAuxWarningProps {
   applying: boolean
   onReset: () => void
@@ -317,19 +335,44 @@ export function ModelSettings({ onMainModelChanged }: ModelSettingsProps) {
     []
   )
 
+  // Guard against stale save responses overwriting newer state.
+  const moaSaveGeneration = useRef(0)
+
   // Quiet debounced persist for inline MoA edits — mirrors the config page's
   // autosave so slot/aggregator tweaks save themselves, matching the
   // preset-level ops (set default / add / delete) that already persist on
   // click. No `applying` spinner, so selecting stays responsive.
+  //
+  // While any slot is half-filled (provider picked, model pending) the save is
+  // HELD, not sent: the previous complete config stays on disk and the next
+  // edit that completes the slot flushes the whole preset. Every edit bumps
+  // the generation so an in-flight response from an older save can never
+  // repaint over the user's mid-edit state.
   const scheduleMoaSave = useCallback((next: MoaConfigResponse) => {
     if (moaSaveTimer.current) {
       window.clearTimeout(moaSaveTimer.current)
+      moaSaveTimer.current = null
+    }
+
+    const generation = moaSaveGeneration.current + 1
+    moaSaveGeneration.current = generation
+
+    if (!moaConfigComplete(next)) {
+      return
     }
 
     moaSaveTimer.current = window.setTimeout(() => {
       void saveMoaModels(next)
-        .then(setMoa)
-        .catch(err => setError(err instanceof Error ? err.message : String(err)))
+        .then(saved => {
+          if (moaSaveGeneration.current === generation) {
+            setMoa(saved)
+          }
+        })
+        .catch(err => {
+          if (moaSaveGeneration.current === generation) {
+            setError(err instanceof Error ? err.message : String(err))
+          }
+        })
     }, 600)
   }, [])
 
@@ -359,7 +402,10 @@ export function ModelSettings({ onMainModelChanged }: ModelSettingsProps) {
   const updateMoaSlot = useCallback((slot: MoaModelSlot, patch: Partial<MoaModelSlot>): MoaModelSlot => {
     const next = { ...slot, ...patch }
 
-    if (patch.provider) {
+    // Picking a new provider invalidates the model choice (models are
+    // per-provider). A same-provider update must not wipe the model — Radix
+    // filters same-value changes, but programmatic callers may not.
+    if (patch.provider && patch.provider !== slot.provider) {
       next.model = ''
     }
 
@@ -368,6 +414,16 @@ export function ModelSettings({ onMainModelChanged }: ModelSettingsProps) {
 
   const saveMoa = useCallback(async (next: MoaConfigResponse) => {
     const epoch = profileEpoch.current
+
+    // Explicit preset ops (set default / add / delete) supersede any pending
+    // debounced slot autosave — cancel it and invalidate in-flight responses
+    // so the two writers can't race each other's state.
+    if (moaSaveTimer.current) {
+      window.clearTimeout(moaSaveTimer.current)
+      moaSaveTimer.current = null
+    }
+
+    moaSaveGeneration.current += 1
     setApplying(true)
     setError('')
 
@@ -991,11 +1047,15 @@ export function ModelSettings({ onMainModelChanged }: ModelSettingsProps) {
                         <SelectValue placeholder={m.provider} />
                       </SelectTrigger>
                       <SelectContent>
-                        {moaSlotProviderOptions.map(provider => (
-                          <SelectItem key={provider.slug || 'none'} value={provider.slug || 'none'}>
-                            {provider.name}
-                          </SelectItem>
-                        ))}
+                        {withActive(moaSlotProviderOptions.map(p => p.slug || 'none'), slot.provider).map(slug => {
+                          const provider = moaSlotProviderOptions.find(p => (p.slug || 'none') === slug)
+
+                          return (
+                            <SelectItem key={slug} value={slug}>
+                              {provider?.name || slug}
+                            </SelectItem>
+                          )
+                        })}
                       </SelectContent>
                     </Select>
                     <Select
@@ -1037,10 +1097,10 @@ export function ModelSettings({ onMainModelChanged }: ModelSettingsProps) {
                 }
                 description={
                   <span className="font-mono text-[0.68rem]">
-                    {slot.provider} · {slot.model}
+                    {slot.provider} · {slot.model || m.model}
                   </span>
                 }
-                key={`${selectedMoaPreset}-${slot.provider}-${slot.model}-${index}`}
+                key={`${selectedMoaPreset}-${index}`}
                 title={`Reference ${index + 1}`}
               />
             ))}
@@ -1070,11 +1130,15 @@ export function ModelSettings({ onMainModelChanged }: ModelSettingsProps) {
                       <SelectValue placeholder={m.provider} />
                     </SelectTrigger>
                     <SelectContent>
-                      {moaSlotProviderOptions.map(provider => (
-                        <SelectItem key={provider.slug || 'none'} value={provider.slug || 'none'}>
-                          {provider.name}
-                        </SelectItem>
-                      ))}
+                      {withActive(moaSlotProviderOptions.map(p => p.slug || 'none'), currentMoaPreset.aggregator.provider).map(slug => {
+                        const provider = moaSlotProviderOptions.find(p => (p.slug || 'none') === slug)
+
+                        return (
+                          <SelectItem key={slug} value={slug}>
+                            {provider?.name || slug}
+                          </SelectItem>
+                        )
+                      })}
                     </SelectContent>
                   </Select>
                   <Select

apps/desktop/src/components/assistant-ui/thread/user-edit-composer.tsx

@@ -57,6 +57,7 @@ import { Codicon } from '@/components/ui/codicon'
 import type { HermesGateway } from '@/hermes'
 import { useI18n } from '@/i18n'
 import { attachmentDisplayText, attachmentId, pathLabel } from '@/lib/chat-runtime'
+import { sanitizeComposerInput } from '@/lib/composer-input-sanitize'
 import { DATA_IMAGE_URL_RE } from '@/lib/embedded-images'
 import { triggerHaptic } from '@/lib/haptics'
 import { Loader2Icon } from '@/lib/icons'
@@ -188,7 +189,7 @@ export const UserEditComposer: FC<UserEditComposerProps> = ({ cwd, gateway, sess
 
   const syncDraftFromEditor = useCallback(
     (editor: HTMLDivElement) => {
-      const nextDraft = composerPlainText(editor)
+      const nextDraft = sanitizeComposerInput(composerPlainText(editor))
 
       if (nextDraft !== draftRef.current) {
         draftRef.current = nextDraft
@@ -477,7 +478,7 @@ export const UserEditComposer: FC<UserEditComposerProps> = ({ cwd, gateway, sess
   }
 
   const handlePaste = (event: ClipboardEvent<HTMLDivElement>) => {
-    const pastedText = event.clipboardData.getData('text')
+    const pastedText = sanitizeComposerInput(event.clipboardData.getData('text'))
 
     if (!pastedText || DATA_IMAGE_URL_RE.test(pastedText.trim())) {
       event.preventDefault()

apps/desktop/src/lib/composer-input-sanitize.test.ts

@@ -0,0 +1,51 @@
+import { describe, expect, it } from 'vitest'
+
+import {
+  collapseRepeatedInputArtifacts,
+  sanitizeComposerInput,
+  stripLeakedBracketedPasteWrappers
+} from './composer-input-sanitize'
+
+describe('stripLeakedBracketedPasteWrappers', () => {
+  it('leaves plain text unchanged', () => {
+    expect(stripLeakedBracketedPasteWrappers('hello world')).toBe('hello world')
+  })
+
+  it('strips canonical escape wrappers', () => {
+    expect(stripLeakedBracketedPasteWrappers('\x1b[200~hello\x1b[201~')).toBe('hello')
+  })
+
+  it('keeps embedded literal bracket forms', () => {
+    const text = 'literal[200~tag and literal[201~tag should stay'
+    expect(stripLeakedBracketedPasteWrappers(text)).toBe(text)
+  })
+})
+
+describe('collapseRepeatedInputArtifacts', () => {
+  it('removes the desktop corruption tail from #62557', () => {
+    const prefix = '需要时随时叫我。'
+    const tail = '[e~[[e' + '~[[e'.repeat(20)
+    expect(collapseRepeatedInputArtifacts(prefix + tail)).toBe(prefix)
+  })
+
+  it('preserves a mid-string marker followed by valid suffix', () => {
+    const text = 'notes ~[[e more text here'
+    expect(collapseRepeatedInputArtifacts(text)).toBe(text)
+  })
+
+  it('preserves trailing punctuation that is not the corruption signature', () => {
+    expect(collapseRepeatedInputArtifacts('wait....')).toBe('wait....')
+  })
+
+  it('does not strip when fewer than minRepeats markers appear at the tail', () => {
+    const text = 'hello~[[e~[[e'
+    expect(collapseRepeatedInputArtifacts(text)).toBe(text)
+  })
+})
+
+describe('sanitizeComposerInput', () => {
+  it('normalizes wrappers and repeated artifact tails together', () => {
+    const corrupted = 'hello[' + '~[[e'.repeat(8)
+    expect(sanitizeComposerInput(corrupted)).toBe('hello')
+  })
+})

apps/desktop/src/lib/composer-input-sanitize.ts

@@ -0,0 +1,71 @@
+/**
+ * Strip terminal bracketed-paste leaks and repeated artifact tails from composer
+ * text before it is shown in the UI or sent to the gateway.
+ *
+ * Mirrors hermes_cli/input_sanitize.py (CLI/TUI gateway defensive path).
+ */
+
+const BRACKETED_PASTE_BOUNDARY_START = /(^|[\s\n>:\]\)])\[200~/g
+const BRACKETED_PASTE_BOUNDARY_END = /\[201~(?=$|[\s\n<[():;.,!?])/g
+const BRACKETED_PASTE_DEGRADED_START = /(^|[\s\n>:\]\)])00~/g
+const BRACKETED_PASTE_DEGRADED_END = /01~(?=$|[\s\n<[():;.,!?])/g
+
+const DESKTOP_PASTE_ARTIFACT = '~[[e'
+
+/** Strip leaked bracketed-paste wrapper markers from user-visible text. */
+export function stripLeakedBracketedPasteWrappers(text: string): string {
+  if (!text) {
+    return text
+  }
+
+  let cleaned = text
+    .replace(/\x1b\[200~/g, '')
+    .replace(/\x1b\[201~/g, '')
+    .replace(/\^\[\[200~/g, '')
+    .replace(/\^\[\[201~/g, '')
+
+  cleaned = cleaned.replace(BRACKETED_PASTE_BOUNDARY_START, '$1')
+  cleaned = cleaned.replace(BRACKETED_PASTE_BOUNDARY_END, '')
+  cleaned = cleaned.replace(BRACKETED_PASTE_DEGRADED_START, '$1')
+  cleaned = cleaned.replace(BRACKETED_PASTE_DEGRADED_END, '')
+
+  return cleaned
+}
+
+/** Drop a trailing run of the desktop ~[[e corruption signature (#62557). */
+export function collapseRepeatedInputArtifacts(text: string, minRepeats = 4): string {
+  if (!text) {
+    return text
+  }
+
+  const marker = DESKTOP_PASTE_ARTIFACT
+  let index = text.length
+  let repeatCount = 0
+
+  while (index >= marker.length && text.slice(index - marker.length, index) === marker) {
+    repeatCount += 1
+    index -= marker.length
+  }
+
+  if (repeatCount < minRepeats) {
+    return text
+  }
+
+  let start = index
+  if (start >= 2 && text.slice(start - 2, start) === '[e') {
+    start -= 2
+  } else if (start >= 1 && text[start - 1] === '[') {
+    start -= 1
+  }
+
+  return text.slice(0, start)
+}
+
+/** Normalize composer text before submit or draft persistence. */
+export function sanitizeComposerInput(text: string): string {
+  if (!text) {
+    return text
+  }
+
+  return collapseRepeatedInputArtifacts(stripLeakedBracketedPasteWrappers(text))
+}

apps/desktop/src/types/hermes.ts

@@ -857,6 +857,8 @@ export interface AuxiliaryModelsResponse {
 export interface MoaModelSlot {
   provider: string
   model: string
+  /** Optional per-slot reasoning effort — round-tripped, not edited here. */
+  reasoning_effort?: string
 }
 
 export interface MoaConfigResponse {
@@ -871,6 +873,10 @@ export interface MoaConfigResponse {
       max_tokens: number
       reference_models: MoaModelSlot[]
       reference_temperature: number
+      /** Optional advisor output cap — round-tripped, not edited here. */
+      reference_max_tokens?: number | null
+      /** Fan-out cadence (per_iteration | user_turn) — round-tripped. */
+      fanout?: string
     }
   >
   aggregator: MoaModelSlot

cli.py

@@ -2995,29 +2995,9 @@ def _should_auto_attach_clipboard_image_on_paste(pasted_text: str) -> bool:
 
 
 def _strip_leaked_bracketed_paste_wrappers(text: str) -> str:
-    """Strip leaked bracketed-paste wrapper markers from user-visible text.
+    from hermes_cli.input_sanitize import strip_leaked_bracketed_paste_wrappers
 
-    Defensive normalization for cases where terminal/prompt_toolkit parsing
-    fails and bracketed-paste markers end up in the buffer as literal text.
-
-    We strip canonical wrappers unconditionally and also handle degraded
-    visible forms like ``[200~`` / ``[201~`` and ``00~`` / ``01~`` when they
-    look like wrapper boundaries, not arbitrary user content.
-    """
-    if not text:
-        return text
-
-    text = (
-        text.replace("\x1b[200~", "")
-        .replace("\x1b[201~", "")
-        .replace("^[[200~", "")
-        .replace("^[[201~", "")
-    )
-    text = re.sub(r"(^|[\s\n>:\]\)])\[200~", r"\1", text)
-    text = re.sub(r"\[201~(?=$|[\s\n<\[\(\):;.,!?])", "", text)
-    text = re.sub(r"(^|[\s\n>:\]\)])00~", r"\1", text)
-    text = re.sub(r"01~(?=$|[\s\n<\[\(\):;.,!?])", "", text)
-    return text
+    return strip_leaked_bracketed_paste_wrappers(text)
 
 
 def _apply_bracketed_paste_timeout_patch() -> None:

docs/profile-routing.md

@@ -0,0 +1,117 @@
+# Profile-Based Routing for Inbound Messages
+
+> **Audience:** Gateway operators and contributors
+> **Source files:** `gateway/profile_routing.py`, `gateway/run.py` (`_profile_name_for_source`), `gateway/platforms/base.py` (`build_source`), `gateway/config.py`
+> **Related:** [Session Lifecycle](session-lifecycle.md), `docs/design/profile-builder.md`
+
+## Overview
+
+By default a single gateway run uses one profile (memory, persona, tools). **Profile-based
+routing** lets one gateway instance serve **multiple isolated profiles**, selecting which
+profile handles an inbound message based on *where the message came from* — the platform,
+server (`guild_id`), channel (`chat_id`), and/or thread (`thread_id`).
+
+This is the inbound counterpart to multiplexing: instead of running N gateways, run one
+gateway and route per-community / per-channel / per-thread to a dedicated profile. Each
+profile keeps fully isolated state (`MEMORY.md`, `USER.md`, `SOUL.md`, sessions, tools).
+
+Routing is **platform-generic**: it works for Discord, Telegram, Feishu, Slack, and every
+adapter — not just Discord.
+
+## Configuring routes
+
+Routes live under `profile_routes` in `config.yaml`. Both the top-level and the nested
+`gateway.profile_routes` forms are accepted (the nested form is what
+`hermes config set gateway.profile_routes ...` writes).
+
+```yaml
+profile_routes:
+  # Route an entire Discord server (guild) to one profile.
+  - name: server-default
+    platform: discord
+    guild_id: "1234567890"
+    profile: server-profile
+
+  # Override a specific channel within that server with a different profile.
+  - name: support-channel
+    platform: discord
+    guild_id: "1234567890"
+    chat_id: "9876543210"
+    profile: support-profile
+
+  # Pin a Telegram group to a profile (Telegram has no guild_id — chat_id only).
+  - name: tg-group
+    platform: telegram
+    chat_id: "-1001234567890"
+    profile: tg-profile
+
+  # Route a single Discord thread.
+  - name: standup-thread
+    platform: discord
+    guild_id: "1234567890"
+    chat_id: "9876543210"
+    thread_id: "1111111111"
+    profile: standup
+```
+
+### Fields
+
+| Field | Required | Description |
+|---|---|---|
+| `name` | yes | Human-readable route identifier (used in logs). |
+| `platform` | yes | Adapter platform: `discord`, `telegram`, `feishu`, `slack`, … |
+| `profile` | yes | Target profile name (must exist under `~/.hermes/profiles/<name>`). |
+| `guild_id` | no | Server/guild (Discord). |
+| `chat_id` | no | Channel/group/DM id. |
+| `thread_id` | no | Thread id within a channel. |
+| `enabled` | no | Default `true`; set `false` to disable a route without removing it. |
+
+## Matching rules
+
+A route matches an inbound source when **every discriminator the route declares is satisfied**
+(conjunctive / AND). A field the route leaves unset is ignored.
+
+- **`platform`** must equal the source platform exactly.
+- **`thread_id`** (if set) must equal the source thread id.
+- **`chat_id`** (if set) must match the source channel **or** its parent — a thread in a
+  channel matches the channel's route (hierarchical match for Discord forums/threads).
+- **`guild_id`** (if set) must equal the source guild.
+
+> A route declaring **both** `guild_id` and `chat_id` requires both to hold. A channel match
+> alone does not satisfy a guild constraint — this is intentional and tested.
+
+When multiple routes match, the **most specific** one wins. Specificity is additive:
+
+| Discriminator | Weight |
+|---|---|
+| `thread_id` | 8 |
+| `chat_id` | 4 |
+| `guild_id` | 2 |
+| (platform only) | 0 |
+
+So a thread route (8) beats a channel route (4) beats a guild route (2) within the same server.
+If no route matches, the message uses the default/active profile.
+
+## How it works at runtime
+
+1. An inbound message arrives at a platform adapter.
+2. `BasePlatformAdapter.build_source` builds the `SessionSource` for the message. Every
+   adapter carries a back-reference to the running `GatewayRunner`
+   (`gateway_runner`, injected in `gateway/run.py`), so it asks the runner to resolve the
+   target profile via `_profile_name_for_source`.
+3. `_profile_name_for_source` runs the configured routes through `match_profile_route` and
+   stamps `source.profile` with the winning route's profile (or leaves it unset).
+4. Downstream, `_resolve_profile_home_for_source` chooses the profile home directory
+   (`source.profile` → active profile → `default`) and the session is scoped per-profile, so
+   each routed community gets isolated memory and conversation state.
+
+Because `gateway_runner` is injected for **all** adapters (declared on `BasePlatformAdapter`),
+every platform goes through this path — not just Discord.
+
+## Relationship to multiplexing
+
+`profile_routes` requires `gateway.multiplex_profiles: true`. Multiplexing is what
+activates the per-profile runtime scope (per-profile `HERMES_HOME`, secret scope, and
+profile-namespaced session keys); routing is the decision layer that picks *which*
+profile a given guild/channel/thread lands in. With multiplexing off, `profile_routes`
+is ignored entirely — behavior is byte-identical to a single-profile gateway.

gateway/config.py

@@ -12,7 +12,7 @@ import logging
 import os
 import json
 from pathlib import Path
-from dataclasses import dataclass, field
+from dataclasses import asdict, dataclass, field, is_dataclass
 from typing import Dict, List, Optional, Any, Callable
 from enum import Enum
 
@@ -721,6 +721,11 @@ class GatewayConfig:
     # fresh session exactly as if the reset policy had fired.  0 = disabled.
     session_store_max_age_days: int = 90
 
+    # Profile-based routing: route specific guilds/channels/threads to
+    # different profiles. See gateway/profile_routing.py. Each entry is a
+    # dict with: name, platform, profile, and optional guild_id/chat_id/thread_id.
+    profile_routes: list = field(default_factory=list)
+
     def get_connected_platforms(self) -> List[Platform]:
         """Return list of platforms that are enabled and configured."""
         connected = []
@@ -827,6 +832,10 @@ class GatewayConfig:
             "unauthorized_dm_behavior": self.unauthorized_dm_behavior,
             "streaming": self.streaming.to_dict(),
             "session_store_max_age_days": self.session_store_max_age_days,
+            "profile_routes": [
+                asdict(r) if is_dataclass(r) and not isinstance(r, type) else r
+                for r in self.profile_routes
+            ],
         }
     
     @classmethod
@@ -919,6 +928,10 @@ class GatewayConfig:
         except (TypeError, ValueError):
             session_store_max_age_days = 90
 
+        # Parse profile routes (validated by gateway.profile_routing)
+        from gateway.profile_routing import parse_profile_routes
+        profile_routes = parse_profile_routes(data.get("profile_routes") or [])
+
         return cls(
             platforms=platforms,
             default_reset_policy=default_policy,
@@ -941,6 +954,7 @@ class GatewayConfig:
             unauthorized_dm_behavior=unauthorized_dm_behavior,
             streaming=StreamingConfig.from_dict(data.get("streaming", {})),
             session_store_max_age_days=session_store_max_age_days,
+            profile_routes=profile_routes,
         )
 
     def get_unauthorized_dm_behavior(self, platform: Optional[Platform] = None) -> str:
@@ -1053,6 +1067,17 @@ def load_gateway_config() -> GatewayConfig:
             if "multiplex_profiles" in yaml_cfg:
                 gw_data["multiplex_profiles"] = yaml_cfg["multiplex_profiles"]
 
+            # Profile-based routing rules: accept either top-level
+            # ``profile_routes`` or the nested ``gateway.profile_routes`` form
+            # (matching the multiplex_profiles parity above).
+            _pr = yaml_cfg.get("profile_routes")
+            if _pr is None:
+                _gw_section = yaml_cfg.get("gateway")
+                if isinstance(_gw_section, dict):
+                    _pr = _gw_section.get("profile_routes")
+            if isinstance(_pr, list):
+                gw_data["profile_routes"] = _pr
+
             gateway_section = yaml_cfg.get("gateway")
             if isinstance(gateway_section, dict):
                 if "multiplex_profiles" in gateway_section and "multiplex_profiles" not in gw_data:

gateway/platforms/base.py

@@ -2353,6 +2353,16 @@ class BasePlatformAdapter(ABC):
     # generic seam; Slack is merely the first consumer).
     supports_inchannel_continuable: bool = False
 
+    # Back-reference to the running ``GatewayRunner``, injected by
+    # ``gateway/run.py`` after the adapter is created. Adapters consume it via
+    # ``getattr(self, "gateway_runner", None)`` for cross-platform delivery and
+    # — critically — for inbound profile routing: ``build_source`` resolves the
+    # target profile through ``runner._profile_name_for_source(...)``. Declaring
+    # it on the base (rather than only on adapters that happen to pre-declare
+    # it) means EVERY platform adapter receives the injection, so profile
+    # routing is platform-generic instead of Discord-only.
+    gateway_runner = None  # type: ignore[assignment]  # set by gateway/run.py
+
     def __init__(self, config: PlatformConfig, platform: Platform):
         self.config = config
         self.platform = platform
@@ -5507,10 +5517,47 @@ class BasePlatformAdapter(ABC):
         auto_thread_created: bool = False,
         auto_thread_initial_name: Optional[str] = None,
     ) -> SessionSource:
-        """Helper to build a SessionSource for this platform."""
+        """Helper to build a SessionSource for this platform.
+
+        When ``gateway.profile_routes`` is configured, the routing engine
+        resolves the matching profile from guild/chat/thread and stamps it on
+        ``source.profile``. Downstream code (``_resolve_profile_home_for_source``
+        in run.py) reads that field to enter ``_profile_runtime_scope`` for
+        per-profile HERMES_HOME isolation.
+        """
         # Normalize empty topic to None
         if chat_topic is not None and not chat_topic.strip():
             chat_topic = None
+
+        # Resolve profile from configured routes (None when no match / no routes)
+        profile = None
+        runner = getattr(self, "gateway_runner", None)
+        if runner is not None:
+            try:
+                profile = runner._profile_name_for_source(
+                    SessionSource(
+                        platform=self.platform,
+                        chat_id=str(chat_id),
+                        chat_name=chat_name,
+                        chat_type=chat_type,
+                        user_id=str(user_id) if user_id else None,
+                        user_name=user_name,
+                        thread_id=str(thread_id) if thread_id else None,
+                        chat_topic=chat_topic.strip() if chat_topic else None,
+                        user_id_alt=user_id_alt,
+                        chat_id_alt=chat_id_alt,
+                        is_bot=is_bot,
+                        guild_id=str(guild_id) if guild_id else None,
+                        parent_chat_id=str(parent_chat_id) if parent_chat_id else None,
+                        message_id=str(message_id) if message_id else None,
+                    )
+                )
+            except Exception:
+                logger.warning(
+                    "Profile resolution failed for %s/%s, defaulting to active profile",
+                    self.platform, chat_id, exc_info=True,
+                )
+
         return SessionSource(
             platform=self.platform,
             chat_id=str(chat_id),
@@ -5527,6 +5574,7 @@ class BasePlatformAdapter(ABC):
             guild_id=str(guild_id) if guild_id else None,
             parent_chat_id=str(parent_chat_id) if parent_chat_id else None,
             message_id=str(message_id) if message_id else None,
+            profile=profile,
             role_authorized=role_authorized,
             auto_thread_created=auto_thread_created,
             auto_thread_initial_name=auto_thread_initial_name,

gateway/platforms/weixin.py

@@ -1515,6 +1515,7 @@ class WeixinAdapter(BasePlatformAdapter):
             event.source,
             group_sessions_per_user=self.config.extra.get("group_sessions_per_user", True),
             thread_sessions_per_user=self.config.extra.get("thread_sessions_per_user", False),
+            profile=event.source.profile,
         )
 
     def _enqueue_text_event(self, event: MessageEvent) -> None:

gateway/profile_routing.py

@@ -0,0 +1,166 @@
+"""Profile-based routing for the gateway with hierarchical matching.
+
+Allows a single Hermes instance to route specific Discord guilds/channels/threads
+to different profiles — each with their own model, tools, memory, and persona.
+
+Matching priority (most specific first):
+  1. platform + chat_id + thread_id (exact thread)  — specificity 14
+  2. platform + chat_id (channel route)             — specificity 6
+  3. platform + guild_id (guild/server route)       — specificity 2
+  4. No match                                       → default profile
+
+Parent-chain matching:
+For Discord threads and forum posts, ``parent_chat_id`` carries the
+direct parent (the channel for a thread, the forum channel for a post).
+Routes keyed on a channel match both direct messages and messages in
+any thread/post whose parent is that channel.
+
+Configuration (config.yaml):
+
+    gateway:
+      profile_routes:
+        - name: server-default
+          platform: discord
+          guild_id: "YOUR_GUILD_ID"
+          profile: server-profile
+
+        - name: special-channel
+          platform: discord
+          guild_id: "YOUR_GUILD_ID"
+          chat_id: "YOUR_CHANNEL_ID"
+          profile: channel-profile
+
+        - name: thread-route
+          platform: discord
+          chat_id: "YOUR_CHANNEL_ID"
+          thread_id: "YOUR_THREAD_ID"
+          profile: thread-profile
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class ProfileRoute:
+    """A single routing rule that maps a platform scope to a profile."""
+
+    name: str
+    platform: str
+    profile: str
+    guild_id: Optional[str] = None
+    chat_id: Optional[str] = None
+    thread_id: Optional[str] = None
+    enabled: bool = True
+
+    @property
+    def specificity(self) -> int:
+        """Higher value = more specific match."""
+        s = 0
+        if self.guild_id:
+            s += 2
+        if self.chat_id:
+            s += 4
+        if self.thread_id:
+            s += 8
+        return s
+
+    def matches(
+        self,
+        platform: str,
+        guild_id: Optional[str] = None,
+        chat_id: Optional[str] = None,
+        thread_id: Optional[str] = None,
+        parent_chat_id: Optional[str] = None,
+    ) -> bool:
+        """Return True if this route matches the given source fields.
+
+        All configured discriminators are matched conjunctively (AND): every
+        discriminator that the route declares must hold. ``chat_id`` supports
+        hierarchical matching for Discord forums/threads:
+        - Direct channel match: chat_id == route.chat_id
+        - Thread in channel: parent_chat_id == route.chat_id
+        A route declaring both ``guild_id`` and ``chat_id`` requires both to
+        match (a chat match alone does not satisfy a guild constraint).
+        """
+        if not self.enabled:
+            return False
+        if self.platform != platform:
+            return False
+        if self.thread_id and self.thread_id != thread_id:
+            return False
+        if self.chat_id and self.chat_id != chat_id and self.chat_id != parent_chat_id:
+            return False
+        if self.guild_id and self.guild_id != guild_id:
+            return False
+        return True
+
+
+def parse_profile_routes(raw: Optional[List[Dict[str, Any]]]) -> List[ProfileRoute]:
+    """Parse profile_routes from config.yaml into ProfileRoute objects.
+
+    Returns routes sorted by specificity (most specific first).
+    """
+    if not raw:
+        return []
+    routes: List[ProfileRoute] = []
+    for entry in raw:
+        if not isinstance(entry, dict):
+            continue
+        name = entry.get("name", "")
+        platform = entry.get("platform", "")
+        profile = entry.get("profile", "")
+        if not platform or not profile:
+            logger.warning(
+                "Skipping profile route %s: missing platform or profile",
+                name,
+            )
+            continue
+        # Validate profile name to prevent path traversal. Lazy import avoids a
+        # circular dependency at module load time.
+        try:
+            from hermes_cli.profiles import (
+                normalize_profile_name,
+                validate_profile_name,
+            )
+            profile = normalize_profile_name(profile)
+            validate_profile_name(profile)
+        except (ValueError, ImportError):
+            logger.warning("Skipping profile route %s: invalid profile name %r", name, profile)
+            continue
+        routes.append(
+            ProfileRoute(
+                name=name,
+                platform=platform,
+                profile=profile,
+                guild_id=entry.get("guild_id"),
+                chat_id=entry.get("chat_id"),
+                thread_id=entry.get("thread_id"),
+                enabled=entry.get("enabled", True),
+            )
+        )
+    # Sort: most specific first so the first match wins.
+    routes.sort(key=lambda r: r.specificity, reverse=True)
+    logger.debug("Loaded %d profile routes (most-specific-first)", len(routes))
+    return routes
+
+
+def match_profile_route(
+    routes: List[ProfileRoute],
+    platform: str,
+    guild_id: Optional[str] = None,
+    chat_id: Optional[str] = None,
+    thread_id: Optional[str] = None,
+    parent_chat_id: Optional[str] = None,
+) -> Optional[ProfileRoute]:
+    """Return the best-matching route, or None for no match."""
+    for route in routes:
+        if route.matches(platform, guild_id=guild_id, chat_id=chat_id, thread_id=thread_id, parent_chat_id=parent_chat_id):
+            return route
+    return None

gateway/run.py

@@ -2667,6 +2667,8 @@ def _normalize_empty_agent_response(
             )
         return response
     if api_calls > 0:
+        if _is_gateway_hidden_reasoning_incomplete_turn(agent_result):
+            return ""
         if agent_result.get("partial"):
             err = agent_result.get("error", "processing incomplete")
             return f"⚠️ Processing stopped: {str(err)[:200]}. Try again."
@@ -2694,6 +2696,30 @@ def _normalize_empty_agent_response(
     return response
 
 
+def _is_gateway_hidden_reasoning_incomplete_turn(agent_result: dict) -> bool:
+    """Detect retry-exhausted turns with hidden reasoning but no visible answer.
+
+    The conversation loop returns the retry-exhaustion sentinel as BOTH
+    ``final_response`` and ``error`` ("Codex response remained incomplete
+    after 3 continuation attempts"), so ``final_response`` being non-empty
+    does not mean the model produced a visible answer. Treat the turn as
+    hidden when the error sentinel is present and ``final_response`` is
+    either empty or merely echoes that sentinel — any genuinely different
+    final text means the model DID answer and must be delivered.
+    """
+    if not isinstance(agent_result, dict):
+        return False
+    if agent_result.get("failed") or agent_result.get("interrupted"):
+        return False
+    if not agent_result.get("partial"):
+        return False
+    error_text = str(agent_result.get("error", "") or "").strip()
+    if "remained incomplete after" not in error_text.lower():
+        return False
+    final_response = str(agent_result.get("final_response") or "").strip()
+    return not final_response or final_response == error_text
+
+
 def _should_clear_resume_pending_after_turn(agent_result: dict) -> bool:
     """Return True only when a gateway turn really completed successfully.
 
@@ -8759,6 +8785,20 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
             if isinstance(val, str) and val.strip():
                 token = val.strip()
                 break
+        # Many adapters (e.g. Discord) store the token on their `config`
+        # sub-object rather than directly on the adapter. Without this lookup
+        # those adapters all return None here, the same-token conflict check
+        # is silently skipped, and every profile's adapter for that platform
+        # starts polling the same bot token — producing a per-message race
+        # for which adapter answers. See test_reads_config_token.
+        if not token:
+            cfg = getattr(adapter, "config", None)
+            if cfg is not None:
+                for attr in ("token", "bot_token"):
+                    val = getattr(cfg, attr, None)
+                    if isinstance(val, str) and val.strip():
+                        token = val.strip()
+                        break
         if not token:
             config = getattr(adapter, "config", None)
             val = getattr(config, "token", None)
@@ -8795,12 +8835,14 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
             if platform_registry.is_registered(platform.value):
                 adapter = platform_registry.create_adapter(platform.value, config)
                 if adapter is not None:
-                    # Adapters that need a back-reference to the gateway runner
-                    # (e.g. for cross-platform admin alerts) declare a
-                    # ``gateway_runner`` attribute. Inject it after creation so
-                    # plugin adapters don't need a custom factory signature.
-                    if hasattr(adapter, "gateway_runner"):
-                        adapter.gateway_runner = self
+                    # Inject a back-reference to the gateway runner so every
+                    # adapter can (a) deliver cross-platform admin alerts and
+                    # (b) resolve inbound profile routing through
+                    # ``runner._profile_name_for_source``. Unconditional:
+                    # ``BasePlatformAdapter`` declares ``gateway_runner``, so
+                    # this reaches ALL platforms (not just the ones that
+                    # pre-declared it), making profile routing platform-generic.
+                    adapter.gateway_runner = self
                     return adapter
                 # Registered but failed to instantiate — don't silently fall
                 # through to built-ins (there are none for plugin platforms).
@@ -11809,6 +11851,14 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                 return None
 
             response = agent_result.get("final_response") or ""
+            # Hidden-reasoning-only retry exhaustion: the loop's sentinel text
+            # ("Codex response remained incomplete after 3 continuation
+            # attempts") doubles as final_response, so it would be delivered
+            # verbatim into the channel — where peer agents can ingest it as a
+            # completed assistant turn (#51628). Blank it here so the normal
+            # empty-response handling (and the suppression below) applies.
+            if _is_gateway_hidden_reasoning_incomplete_turn(agent_result):
+                response = ""
             try:
                 from gateway.response_filters import is_intentional_silence_agent_result
                 _intentional_silence = is_intentional_silence_agent_result(
@@ -12039,6 +12089,9 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
             # forgets what was just asked.  Persist the user turn so the
             # conversation is preserved. (#7100)
             agent_failed_early = bool(agent_result.get("failed"))
+            hidden_reasoning_incomplete = _is_gateway_hidden_reasoning_incomplete_turn(
+                agent_result
+            )
             _err_str_for_classify = str(agent_result.get("error", "")).lower()
             # Use specific multi-word phrases (not bare "exceed" or "token")
             # to avoid false positives on transient errors like "rate limit
@@ -12067,6 +12120,13 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                     "message so conversation context is preserved on retry.",
                     session_entry.session_id,
                 )
+            elif hidden_reasoning_incomplete:
+                logger.warning(
+                    "Suppressing hidden-reasoning-only incomplete gateway turn "
+                    "for session %s: %s",
+                    session_entry.session_id,
+                    agent_result.get("error", "processing incomplete"),
+                )
 
             # When compression is exhausted, the session is permanently too
             # large to process.  Auto-reset it so the next message starts
@@ -12150,11 +12210,14 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
             # entries that were stripped before the agent saw them.
             if is_context_overflow_failure:
                 pass  # handled above — skip all transcript writes
-            elif agent_failed_early:
+            elif agent_failed_early or hidden_reasoning_incomplete:
                 # Transient failure (429/timeout/5xx): persist only the user
                 # message so the next message can load a transcript that
                 # reflects what was said.  Skip the assistant error text since
-                # it's a gateway-generated hint, not model output. (#7100)
+                # it's a gateway-generated hint, not model output. Hidden-
+                # reasoning-only incomplete turns follow the same persistence
+                # rule so peer-agent channels don't ingest them as completed
+                # assistant turns. (#7100, #51628)
                 _user_entry = {
                     "role": "user",
                     "content": (
@@ -17246,19 +17309,108 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                 persist_user_timestamp=persist_user_timestamp,
             )
 
+    def _profile_name_for_source(self, source: SessionSource) -> Optional[str]:
+        """Resolve the profile name for an inbound source via configured routes.
+
+        Returns ``None`` when multiplexing is off, no routes are configured, or
+        no route matches. Callers (``build_source``,
+        ``_resolve_profile_home_for_source``) treat ``None`` as "use the
+        default/active profile". When ``gateway.profile_routes`` is configured,
+        the most specific matching route wins (guild < channel < thread). See
+        :mod:`gateway.profile_routing` for matching rules.
+
+        Gated on ``gateway.multiplex_profiles``: routing stamps
+        ``source.profile``, which selects the session-key namespace and batch
+        keys — but the profile-scoped agent run only activates under
+        multiplexing. Without this gate, a configured route with multiplexing
+        off would namespace batch/session keys by profile while the agent
+        still runs in ``agent:main``, splitting the two out of agreement.
+        """
+        config = getattr(self, "config", None)
+        if not getattr(config, "multiplex_profiles", False):
+            return None
+        routes = getattr(config, "profile_routes", None)
+        if not routes:
+            return None
+        from gateway.profile_routing import match_profile_route
+        try:
+            matched = match_profile_route(
+                routes,
+                platform=source.platform.value,
+                guild_id=getattr(source, "guild_id", None),
+                chat_id=source.chat_id,
+                thread_id=getattr(source, "thread_id", None),
+                parent_chat_id=getattr(source, "parent_chat_id", None),
+            )
+        except Exception:
+            logger.warning(
+                "Profile route matching failed for %s/%s, falling back to default",
+                source.platform, source.chat_id, exc_info=True,
+            )
+            return None
+        if matched:
+            return matched.profile
+        logger.debug(
+            "No profile route matched: platform=%s chat_id=%s thread_id=%s parent_chat_id=%s",
+            source.platform.value, source.chat_id,
+            getattr(source, "thread_id", None), getattr(source, "parent_chat_id", None),
+        )
+        return None
+
     def _resolve_profile_home_for_source(self, source: SessionSource) -> "Path":
         """Resolve which profile's HERMES_HOME should serve this inbound source.
 
-        Prefers the profile the source was routed to (``source.profile`` — set
-        by the /p/<profile>/ URL prefix or a per-credential adapter), falling
-        back to the active profile (the multiplexer's own home).
+        Resolution order:
+          1. ``source.profile`` — set by /p/<profile>/ URL prefix, per-credential
+             adapter ownership, OR profile_routes matching at ``build_source`` time.
+          2. ``_profile_name_for_source`` — re-run routing here as a defensive
+             fallback for sources that bypass ``build_source``.
+          3. The active profile (the multiplexer's own home).
         """
-        from hermes_cli.profiles import get_active_profile_name, get_profile_dir
+        from hermes_cli.profiles import (
+            get_active_profile_name,
+            get_profile_dir,
+            profile_exists,
+        )
+        from hermes_constants import get_hermes_home
+        
+        # Track whether a profile was explicitly requested (vs. falling back to default)
+        explicit_profile = None
         try:
-            name = (source.profile or "").strip() or get_active_profile_name() or "default"
-            return get_profile_dir(name)
+            name = (source.profile or "").strip()
+            if name:
+                explicit_profile = name  # User explicitly set this profile
+            if not name:
+                name = self._profile_name_for_source(source)
+                if name:
+                    explicit_profile = name  # Routing explicitly set this profile
+            if not name:
+                name = get_active_profile_name() or "default"
+            
+            profile_dir = get_profile_dir(name)
+            # Warn if an explicit profile doesn't exist on disk
+            if explicit_profile and not profile_exists(name):
+                logger.warning(
+                    "Profile %r does not exist for source %s/%s (guild_id=%s), "
+                    "falling back to global HERMES_HOME",
+                    explicit_profile,
+                    source.platform.value,
+                    source.chat_id,
+                    getattr(source, "guild_id", None),
+                )
+                return get_hermes_home()
+            return profile_dir
         except Exception:
-            from hermes_constants import get_hermes_home
+            # Catch normalization errors, path errors, etc.
+            logger.warning(
+                "Failed to resolve profile directory for source %s/%s (guild_id=%s), "
+                "falling back to global HERMES_HOME: %s",
+                source.platform.value,
+                source.chat_id,
+                getattr(source, "guild_id", None),
+                explicit_profile or "(no profile)",
+                exc_info=True,
+            )
             return get_hermes_home()
 
     async def _run_agent_inner(

gateway/session.py

@@ -1993,6 +1993,7 @@ class SessionStore:
                     "chat_id": source.chat_id,
                     "chat_type": source.chat_type,
                     "thread_id": source.thread_id,
+                    "profile_name": source.profile,
                 }
 
         if _needs_save:
@@ -2273,6 +2274,7 @@ class SessionStore:
                 "chat_id": old_entry.origin.chat_id if old_entry.origin else None,
                 "chat_type": old_entry.origin.chat_type if old_entry.origin else None,
                 "thread_id": old_entry.origin.thread_id if old_entry.origin else None,
+                "profile_name": old_entry.origin.profile if old_entry.origin else None,
             }
 
         if self._db and db_end_session_id:

hermes_cli/config.py

@@ -1700,6 +1700,18 @@ DEFAULT_CONFIG = {
             "extra_body": {},
             "reasoning_effort": "",  # per-task thinking level: none|minimal|low|medium|high|xhigh|max|ultra (empty = provider default)
         },
+        # Goal judge — evaluates whether a /goal run's latest response
+        # satisfies the goal/contract, and drafts goal contracts. Short
+        # structured-JSON calls; a fast cheap model is fine.
+        "goal_judge": {
+            "provider": "auto",
+            "model": "",
+            "base_url": "",
+            "api_key": "",
+            "timeout": 60,
+            "extra_body": {},
+            "reasoning_effort": "",  # per-task thinking level: none|minimal|low|medium|high|xhigh|max|ultra (empty = provider default)
+        },
         # Curator — skill-usage review fork. Timeout is generous because the
         # review pass can take several minutes on reasoning models (umbrella
         # building over hundreds of candidate skills). "auto" = use main chat
@@ -5524,10 +5536,13 @@ def _persist_migration(config: Dict[str, Any]) -> None:
 
     Every migration step MUST route its write through this helper instead of
     calling ``save_config`` directly. It is a thin wrapper over
-    ``save_config(config)`` (default-stripping ON); centralising the call makes
-    the invariant impossible to regress one migration at a time. Correctness
-    across seeds, non-default values, behaviour flips, and data transforms is
-    verified by the migration parity tests.
+    ``save_config(config)`` (default-stripping ON, no ``merge_existing``);
+    centralising the call makes the invariant impossible to regress one
+    migration at a time. Callers must pass the full raw config returned by
+    ``read_raw_config()`` after in-place mutations (including key removals);
+    deep-merging the on-disk file back in would resurrect keys the migration
+    just deleted. Partial-save preservation for unrelated top-level sections
+    belongs on ``save_config(..., merge_existing=True)``, not here.
     """
     save_config(config)
 
@@ -6302,6 +6317,25 @@ def migrate_config(interactive: bool = True, quiet: bool = False) -> Dict[str, A
     return results
 
 
+def _merge_partial_save(raw: dict, override: dict) -> dict:
+    """Merge *override* over *raw* for partial ``save_config`` writes.
+
+    Top-level sections omitted from *override* are preserved from *raw*.
+    Shared top-level dict sections are deep-merged so a caller can update one
+    nested key without dropping sibling keys from disk. Intentional key
+    removals within a section are not supported here — migration writes must
+    route through ``_persist_migration`` with a full ``read_raw_config()`` dict
+    instead.
+    """
+    result = copy.deepcopy(override)
+    for key, value in raw.items():
+        if key not in result:
+            result[key] = copy.deepcopy(value)
+        elif isinstance(result.get(key), dict) and isinstance(value, dict):
+            result[key] = _deep_merge(value, result[key])
+    return result
+
+
 def _deep_merge(base: dict, override: dict) -> dict:
     """Recursively merge *override* into *base*, preserving nested defaults.
 
@@ -7184,6 +7218,7 @@ def save_config(
     *,
     strip_defaults: bool = True,
     preserve_keys: Optional[Set[Tuple[str, ...]]] = None,
+    merge_existing: bool = False,
 ):
     """Save configuration to ~/.hermes/config.yaml.\n
 
@@ -7192,6 +7227,12 @@ def save_config(
     before any normalisation).  This prevents config.yaml from being
     contaminated with schema defaults on every save, which makes future
     default changes invisible to users.
+
+    When ``merge_existing`` is True, the on-disk raw config is deep-merged
+    under *config* before writing so partial callers (migration steps via
+    ``_persist_migration``) cannot drop unrelated sections the caller omitted.
+    Full-document replacement callers (dashboard raw YAML editor, callers that
+    already deep-merge) must leave this False so intentional deletions survive.
     """
     with _CONFIG_LOCK:
         if is_managed():
@@ -7226,11 +7267,17 @@ def save_config(
         explicit_raw_paths: Optional[Set[Tuple[str, ...]]] = (
             _explicit_config_paths(_raw_for_paths) if _raw_for_paths else None
         )
+        if merge_existing and _raw_for_paths:
+            config = _merge_partial_save(_raw_for_paths, config)
         # ----------------------------------------------------------------
 
         current_normalized = _normalize_root_model_keys(_normalize_max_turns_config(config))
         normalized = current_normalized
-        raw_existing = _normalize_root_model_keys(_normalize_max_turns_config(read_raw_config()))
+        raw_existing = (
+            _normalize_root_model_keys(_normalize_max_turns_config(_raw_for_paths))
+            if _raw_for_paths
+            else {}
+        )
         if raw_existing:
             normalized = _preserve_env_ref_templates(
                 normalized,
@@ -7278,6 +7325,7 @@ def save_config(
             extra_content="".join(parts) if parts else None,
         )
         _secure_file(config_path)
+        _RAW_CONFIG_CACHE.pop(str(config_path), None)
         _LAST_EXPANDED_CONFIG_BY_PATH[str(config_path)] = copy.deepcopy(current_normalized)
 
 

hermes_cli/goals.py

@@ -878,20 +878,11 @@ def judge_goal(
         return "continue", "empty response (nothing to evaluate)", False, None
 
     try:
-        from agent.auxiliary_client import get_auxiliary_extra_body, get_text_auxiliary_client
+        from agent.auxiliary_client import call_llm
     except Exception as exc:
         logger.debug("goal judge: auxiliary client import failed: %s", exc)
         return "continue", "auxiliary client unavailable", False, None
 
-    try:
-        client, model = get_text_auxiliary_client("goal_judge")
-    except Exception as exc:
-        logger.debug("goal judge: get_text_auxiliary_client failed: %s", exc)
-        return "continue", "auxiliary client unavailable", False, None
-
-    if client is None or not model:
-        return "continue", "no auxiliary client configured", False, None
-
     # Build the prompt. Priority: contract > subgoals > plain. When both a
     # contract and subgoals exist, the subgoals are appended into the
     # contract block as extra criteria so the judge sees a single source of
@@ -935,8 +926,11 @@ def judge_goal(
         )
 
     try:
-        resp = client.chat.completions.create(
-            model=model,
+        # Route through call_llm so auxiliary.goal_judge.* config
+        # (provider/model/base_url, extra_body, reasoning_effort, retries)
+        # all apply — the direct-create path dropped extra_body (#35566).
+        resp = call_llm(
+            task="goal_judge",
             messages=[
                 {"role": "system", "content": JUDGE_SYSTEM_PROMPT},
                 {"role": "user", "content": prompt},
@@ -944,7 +938,6 @@ def judge_goal(
             temperature=0,
             max_tokens=_goal_judge_max_tokens(),
             timeout=timeout,
-            extra_body=get_auxiliary_extra_body() or None,
         )
     except Exception as exc:
         logger.info("goal judge: API call failed (%s) — falling through to continue", exc)
@@ -999,23 +992,15 @@ def draft_contract(objective: str, *, timeout: float = DEFAULT_JUDGE_TIMEOUT) ->
         return None
 
     try:
-        from agent.auxiliary_client import get_auxiliary_extra_body, get_text_auxiliary_client
+        from agent.auxiliary_client import call_llm
     except Exception as exc:
         logger.debug("goal draft: auxiliary client import failed: %s", exc)
         return None
 
     try:
-        client, model = get_text_auxiliary_client("goal_judge")
-    except Exception as exc:
-        logger.debug("goal draft: get_text_auxiliary_client failed: %s", exc)
-        return None
-
-    if client is None or not model:
-        return None
-
-    try:
-        resp = client.chat.completions.create(
-            model=model,
+        # Route through call_llm — same #35566 fix as the judge call above.
+        resp = call_llm(
+            task="goal_judge",
             messages=[
                 {"role": "system", "content": DRAFT_CONTRACT_SYSTEM_PROMPT},
                 {"role": "user", "content": f"Objective:\n{_truncate(objective, 4000)}"},
@@ -1023,7 +1008,6 @@ def draft_contract(objective: str, *, timeout: float = DEFAULT_JUDGE_TIMEOUT) ->
             temperature=0,
             max_tokens=_goal_judge_max_tokens(),
             timeout=timeout,
-            extra_body=get_auxiliary_extra_body() or None,
         )
     except Exception as exc:
         logger.info("goal draft: API call failed (%s)", exc)

hermes_cli/input_sanitize.py

@@ -0,0 +1,70 @@
+"""Sanitize user prompt text leaked from terminal / paste control sequences."""
+
+from __future__ import annotations
+
+import re
+
+_BRACKETED_PASTE_BOUNDARY_START = re.compile(r"(^|[\s\n>:\]\)])\[200~")
+_BRACKETED_PASTE_BOUNDARY_END = re.compile(r"\[201~(?=$|[\s\n<\[\(\):;.,!?])")
+_BRACKETED_PASTE_DEGRADED_START = re.compile(r"(^|[\s\n>:\]\)])00~")
+_BRACKETED_PASTE_DEGRADED_END = re.compile(r"01~(?=$|[\s\n<\[\(\):;.,!?])")
+
+# Corruption signature from desktop bracketed-paste leaks (#62557).
+_DESKTOP_PASTE_ARTIFACT = "~[[e"
+
+
+def strip_leaked_bracketed_paste_wrappers(text: str) -> str:
+    """Strip leaked bracketed-paste wrapper markers from user-visible text.
+
+    Defensive normalization for cases where terminal/prompt_toolkit parsing
+    fails and bracketed-paste markers end up in the buffer as literal text.
+
+    Canonical wrappers are stripped unconditionally. Degraded visible forms like
+    ``[200~`` / ``[201~`` and ``00~`` / ``01~`` are removed only at boundaries
+    so embedded literals such as ``literal[200~tag`` stay intact.
+    """
+    if not text:
+        return text
+
+    text = (
+        text.replace("\x1b[200~", "")
+        .replace("\x1b[201~", "")
+        .replace("^[[200~", "")
+        .replace("^[[201~", "")
+    )
+    text = _BRACKETED_PASTE_BOUNDARY_START.sub(r"\1", text)
+    text = _BRACKETED_PASTE_BOUNDARY_END.sub("", text)
+    text = _BRACKETED_PASTE_DEGRADED_START.sub(r"\1", text)
+    text = _BRACKETED_PASTE_DEGRADED_END.sub("", text)
+    return text
+
+
+def collapse_repeated_input_artifacts(text: str, min_repeats: int = 4) -> str:
+    """Drop a trailing run of the desktop ~[[e corruption signature (#62557)."""
+    if not text:
+        return text
+
+    marker = _DESKTOP_PASTE_ARTIFACT
+    index = len(text)
+    repeat_count = 0
+    while index >= len(marker) and text[index - len(marker) : index] == marker:
+        repeat_count += 1
+        index -= len(marker)
+
+    if repeat_count < min_repeats:
+        return text
+
+    start = index
+    if start >= 2 and text[start - 2 : start] == "[e":
+        start -= 2
+    elif start >= 1 and text[start - 1] == "[":
+        start -= 1
+    return text[:start]
+
+
+def sanitize_user_prompt_text(text: str) -> str:
+    """Normalize user-authored prompt text before persistence or model input."""
+    if not isinstance(text, str) or not text:
+        return text
+    cleaned = strip_leaked_bracketed_paste_wrappers(text)
+    return collapse_repeated_input_artifacts(cleaned)

hermes_cli/kanban_decompose.py

@@ -298,23 +298,11 @@ def decompose_task(
     roster, valid_names = _build_roster()
 
     try:
-        from agent.auxiliary_client import (  # type: ignore
-            get_auxiliary_extra_body,
-            get_text_auxiliary_client,
-        )
+        from agent.auxiliary_client import call_llm  # type: ignore
     except Exception as exc:
         logger.debug("decompose: auxiliary client import failed: %s", exc)
         return DecomposeOutcome(task_id, False, "auxiliary client unavailable")
 
-    try:
-        client, model = get_text_auxiliary_client("kanban_decomposer")
-    except Exception as exc:
-        logger.debug("decompose: get_text_auxiliary_client failed: %s", exc)
-        return DecomposeOutcome(task_id, False, "auxiliary client unavailable")
-
-    if client is None or not model:
-        return DecomposeOutcome(task_id, False, "no auxiliary client configured")
-
     user_msg = _USER_TEMPLATE.format(
         task_id=task.id,
         title=_truncate(task.title or "", 400),
@@ -324,8 +312,12 @@ def decompose_task(
     )
 
     try:
-        resp = client.chat.completions.create(
-            model=model,
+        # Route through call_llm so auxiliary.kanban_decomposer.* config
+        # (provider/model/base_url, extra_body, reasoning_effort, retries)
+        # all apply — the previous direct client.chat.completions.create()
+        # path dropped auxiliary.<task>.extra_body entirely (#35566).
+        resp = call_llm(
+            task="kanban_decomposer",
             messages=[
                 {"role": "system", "content": _SYSTEM_PROMPT},
                 {"role": "user", "content": user_msg},
@@ -333,7 +325,6 @@ def decompose_task(
             temperature=0.3,
             max_tokens=4000,
             timeout=timeout or 180,
-            extra_body=get_auxiliary_extra_body() or None,
         )
     except Exception as exc:
         logger.info(

hermes_cli/kanban_specify.py

@@ -162,22 +162,11 @@ def specify_task(
         )
 
     try:
-        from agent.auxiliary_client import get_auxiliary_extra_body, get_text_auxiliary_client
+        from agent.auxiliary_client import call_llm
     except Exception as exc:  # pragma: no cover — import smoke test
         logger.debug("specify: auxiliary client import failed: %s", exc)
         return SpecifyOutcome(task_id, False, "auxiliary client unavailable")
 
-    try:
-        client, model = get_text_auxiliary_client("triage_specifier")
-    except Exception as exc:
-        logger.debug("specify: get_text_auxiliary_client failed: %s", exc)
-        return SpecifyOutcome(task_id, False, "auxiliary client unavailable")
-
-    if client is None or not model:
-        return SpecifyOutcome(
-            task_id, False, "no auxiliary client configured"
-        )
-
     user_msg = _USER_TEMPLATE.format(
         task_id=task.id,
         title=_truncate(task.title or "", 400),
@@ -185,8 +174,11 @@ def specify_task(
     )
 
     try:
-        resp = client.chat.completions.create(
-            model=model,
+        # Route through call_llm so auxiliary.triage_specifier.* config
+        # (provider/model/base_url, extra_body, reasoning_effort, retries)
+        # all apply — the direct-create path dropped extra_body (#35566).
+        resp = call_llm(
+            task="triage_specifier",
             messages=[
                 {"role": "system", "content": _SYSTEM_PROMPT},
                 {"role": "user", "content": user_msg},
@@ -194,7 +186,6 @@ def specify_task(
             temperature=0.3,
             max_tokens=HERMES_KANBAN_SPECIFY_MAX_TOKENS,
             timeout=timeout or 120,
-            extra_body=get_auxiliary_extra_body() or None,
         )
     except Exception as exc:
         logger.info(

hermes_cli/mcp_catalog.py

@@ -10,7 +10,11 @@ Catalog policy:
 - Entries are added only by merging a PR into hermes-agent. Presence in the
   ``optional-mcps/`` directory = Nous approval. No community tier, no trust
   signals beyond "it's in the catalog".
-- Manifests pin transport details (commands, args, refs). MCPs are never
+- Manifests pin transport details (commands, args, refs). Pins follow the
+  same supply-chain rules as pyproject dependencies: exact versions for
+  package launchers (``uvx pkg==X``, ``npx pkg@X``), full commit SHAs for
+  git installs, and the pinned release should be at least 2 weeks old at
+  pin time. MCPs are never
   auto-updated; users explicitly re-run ``hermes mcp install <name>`` to
   pull a new manifest version after a repo update.
 - Secrets prompted at install time go to ``~/.hermes/.env`` (the
@@ -77,6 +81,10 @@ class TransportSpec:
     args: List[str] = field(default_factory=list)
     url: Optional[str] = None
     version: Optional[str] = None  # informational, pinned
+    # Static environment variables for the stdio subprocess (e.g. telemetry
+    # opt-outs, mode flags). NOT for secrets — credentials go through
+    # auth.env so they are prompted for and land in ~/.hermes/.env.
+    env: Dict[str, str] = field(default_factory=dict)
 
 
 @dataclass
@@ -184,12 +192,20 @@ def _parse_manifest(path: Path) -> CatalogEntry:
     args = transport_raw.get("args") or []
     if not isinstance(args, list):
         raise CatalogError(f"{path}: transport.args must be a list")
+    env_raw = transport_raw.get("env") or {}
+    if not isinstance(env_raw, dict) or not all(
+        isinstance(k, str) and isinstance(v, str) for k, v in env_raw.items()
+    ):
+        raise CatalogError(
+            f"{path}: transport.env must be a mapping of string to string"
+        )
     transport = TransportSpec(
         type=t_type,
         command=transport_raw.get("command"),
         args=[str(a) for a in args],
         url=transport_raw.get("url"),
         version=transport_raw.get("version"),
+        env=dict(env_raw),
     )
     if t_type == "stdio" and not transport.command:
         raise CatalogError(f"{path}: stdio transport requires 'command'")
@@ -468,6 +484,8 @@ def _build_server_config(
         cfg["command"] = _expand_install_dir(t.command or "", install_dir)
         if t.args:
             cfg["args"] = [_expand_install_dir(a, install_dir) for a in t.args]
+        if t.env:
+            cfg["env"] = dict(t.env)
     elif t.type == "http":
         cfg["url"] = t.url
         if entry.auth.type == "oauth":

hermes_cli/moa_config.py

@@ -110,6 +110,77 @@ def _clean_slot(slot: Any) -> dict[str, Any] | None:
     return clean
 
 
+def _slot_problem(slot: Any) -> str | None:
+    """Return a human-readable problem for a slot ``_clean_slot`` would drop.
+
+    None means the slot is complete and valid. Mirrors ``_clean_slot`` exactly
+    so the write-boundary validator (``validate_moa_payload``) and the
+    tolerant runtime normalizer can never disagree about what is acceptable.
+    """
+    if not isinstance(slot, dict):
+        return "must be an object with 'provider' and 'model'"
+    provider = str(slot.get("provider") or "").strip()
+    model = str(slot.get("model") or "").strip()
+    if not provider and not model:
+        return "provider and model are required"
+    if not provider:
+        return "provider is required"
+    if not model:
+        return f"model is required (provider '{provider}' has no model selected)"
+    if provider.lower() == "moa":
+        return "the Mixture of Agents provider cannot be used inside a preset (recursive MoA)"
+    return None
+
+
+def validate_moa_payload(raw: Any) -> list[str]:
+    """Return the problems ``normalize_moa_config`` would silently paper over.
+
+    ``normalize_moa_config`` is deliberately tolerant: at *read* time a
+    hand-edited config must degrade to defaults rather than crash the agent.
+    That same tolerance at *write* time is a corruption engine — a client that
+    sends a half-filled slot gets its whole preset silently replaced with the
+    hardcoded defaults (#64156). API write paths call this first and reject
+    invalid payloads loudly instead of saving something the user never chose.
+
+    Returns a list of human-readable problems; empty means safe to save.
+    """
+    if not isinstance(raw, dict):
+        return ["MoA config must be an object"]
+
+    presets_raw = raw.get("presets")
+    if isinstance(presets_raw, dict) and presets_raw:
+        presets: dict[Any, Any] = presets_raw
+    else:
+        # Legacy flat payload: the top-level object is the default preset.
+        presets = {DEFAULT_MOA_PRESET_NAME: raw}
+
+    problems: list[str] = []
+    for name, preset in presets.items():
+        label = str(name or "").strip() or "(unnamed)"
+        if not isinstance(preset, dict):
+            problems.append(f"preset '{label}': must be an object")
+            continue
+
+        refs = preset.get("reference_models")
+        if not isinstance(refs, list):
+            refs = [refs] if isinstance(refs, dict) else []
+        complete_refs = 0
+        for index, slot in enumerate(refs):
+            issue = _slot_problem(slot)
+            if issue:
+                problems.append(f"preset '{label}' reference {index + 1}: {issue}")
+            else:
+                complete_refs += 1
+        if not complete_refs:
+            problems.append(f"preset '{label}': needs at least one complete reference model")
+
+        agg_issue = _slot_problem(preset.get("aggregator"))
+        if agg_issue:
+            problems.append(f"preset '{label}' aggregator: {agg_issue}")
+
+    return problems
+
+
 def _default_preset() -> dict[str, Any]:
     return {
         "reference_models": deepcopy(DEFAULT_MOA_REFERENCE_MODELS),

hermes_cli/model_setup_flows.py

@@ -27,6 +27,54 @@ import subprocess
 from hermes_cli.config import clear_model_endpoint_credentials
 
 
+# AWS cross-region inference profile prefixes. Any geo-prefixed profile only
+# routes from endpoints in its own geography, so the Bedrock picker must not
+# offer (e.g.) us.* profiles to an eu-central-2 endpoint — selecting one
+# produces a config AWS rejects regardless of credentials (#28156).
+# global.* routes from everywhere. Full set per the AWS cross-region
+# inference docs.
+BEDROCK_GEO_PREFIXES = (
+    "us.", "eu.", "ap.", "apac.", "jp.", "ca.", "sa.", "me.", "af.",
+)
+
+
+def bedrock_region_geo_prefix(region_name: str) -> str:
+    """Map an AWS region name to its inference-profile geo prefix ('' = unknown)."""
+    r = (region_name or "").lower()
+    for geo, region_prefixes in (
+        ("us.", ("us-", "us_gov")),
+        ("eu.", ("eu-",)),
+        ("ap.", ("ap-",)),
+        ("ca.", ("ca-",)),
+        ("sa.", ("sa-",)),
+        ("me.", ("me-",)),
+        ("af.", ("af-",)),
+    ):
+        if r.startswith(region_prefixes):
+            return geo
+    return ""
+
+
+def bedrock_model_routable_from_region(model_id: str, region_name: str) -> bool:
+    """True when *model_id* can be invoked from *region_name*'s endpoint.
+
+    Bare foundation-model ids and ``global.*`` profiles route from anywhere.
+    Geo-prefixed inference profiles (``us.*``, ``eu.*``, ...) only route from
+    endpoints in their own geography. Unknown region shapes hide nothing.
+    """
+    mid = (model_id or "").lower()
+    matched_geo = next((p for p in BEDROCK_GEO_PREFIXES if mid.startswith(p)), None)
+    if matched_geo is None or mid.startswith("global."):
+        return True
+    geo = bedrock_region_geo_prefix(region_name)
+    if not geo:
+        return True
+    if geo == "ap.":
+        # Asia-Pacific regions can carry ap./apac./jp. profile spellings.
+        return matched_geo in ("ap.", "apac.", "jp.")
+    return matched_geo == geo
+
+
 def _prune_replaced_custom_model_config_credentials(
     base_url: str,
     *,
@@ -2265,6 +2313,8 @@ def _model_flow_bedrock(config, current_model=""):
             "global.twelvelabs.",
         )
         _EXCLUDE_SUBSTRINGS = ("safeguard", "voxtral", "palmyra-vision")
+
+
         filtered = []
         for m in live_models:
             mid = m["id"]
@@ -2272,44 +2322,59 @@ def _model_flow_bedrock(config, current_model=""):
                 continue
             if any(s in mid.lower() for s in _EXCLUDE_SUBSTRINGS):
                 continue
+            if not bedrock_model_routable_from_region(mid, region):
+                continue
             filtered.append(m)
 
-        # Deduplicate: prefer inference profiles (us.*, global.*) over bare
-        # foundation model IDs.
+        # Deduplicate: prefer inference profiles (geo-prefixed or global.*)
+        # over bare foundation model IDs.
+        _PROFILE_PREFIXES = BEDROCK_GEO_PREFIXES + ("global.",)
         profile_base_ids = set()
         for m in filtered:
             mid = m["id"]
-            if mid.startswith(("us.", "global.")):
-                base = mid.split(".", 1)[1] if "." in mid[3:] else mid
-                profile_base_ids.add(base)
+            _pp = next((p for p in _PROFILE_PREFIXES if mid.startswith(p)), None)
+            if _pp:
+                profile_base_ids.add(mid[len(_pp):])
 
         deduped = []
         for m in filtered:
             mid = m["id"]
-            if not mid.startswith(("us.", "global.")) and mid in profile_base_ids:
+            if (
+                not mid.startswith(_PROFILE_PREFIXES)
+                and mid in profile_base_ids
+            ):
                 continue
             deduped.append(m)
 
-        _RECOMMENDED = [
-            "us.anthropic.claude-sonnet-4-6",
-            "us.anthropic.claude-opus-4-6",
-            "us.anthropic.claude-haiku-4-5",
-            "us.amazon.nova-pro",
-            "us.amazon.nova-lite",
-            "us.amazon.nova-micro",
+        # Recommended models, matched geo-agnostically so an EU (eu.*) or
+        # APAC (apac.*) picker pins its own region's profile of the same
+        # model rather than a us.* one it can't route to (#28156).
+        _RECOMMENDED_BASES = [
+            "anthropic.claude-sonnet-4-6",
+            "anthropic.claude-opus-4-6",
+            "anthropic.claude-haiku-4-5",
+            "amazon.nova-pro",
+            "amazon.nova-lite",
+            "amazon.nova-micro",
             "deepseek.v3",
-            "us.meta.llama4-maverick",
-            "us.meta.llama4-scout",
+            "meta.llama4-maverick",
+            "meta.llama4-scout",
         ]
 
+        def _base_id(mid: str) -> str:
+            _pp = next((p for p in _PROFILE_PREFIXES if mid.startswith(p)), None)
+            return mid[len(_pp):] if _pp else mid
+
         def _sort_key(m):
             mid = m["id"]
-            for i, rec in enumerate(_RECOMMENDED):
-                if mid.startswith(rec):
-                    return (0, i, mid)
+            base = _base_id(mid)
+            for i, rec in enumerate(_RECOMMENDED_BASES):
+                if base.startswith(rec):
+                    # In-region geo profile beats global.* for the same model
+                    return (0, i, 0 if not mid.startswith("global.") else 1, mid)
             if mid.startswith("global."):
-                return (1, 0, mid)
-            return (2, 0, mid)
+                return (1, 0, 0, mid)
+            return (2, 0, 0, mid)
 
         deduped.sort(key=_sort_key)
         model_list = [m["id"] for m in deduped]

hermes_cli/profile_describer.py

@@ -210,23 +210,11 @@ def describe_profile(
         model, provider = None, None
 
     try:
-        from agent.auxiliary_client import (  # type: ignore
-            get_auxiliary_extra_body,
-            get_text_auxiliary_client,
-        )
+        from agent.auxiliary_client import call_llm  # type: ignore
     except Exception as exc:
         logger.debug("describe: auxiliary client import failed: %s", exc)
         return DescribeOutcome(canon, False, "auxiliary client unavailable")
 
-    try:
-        client, aux_model = get_text_auxiliary_client("profile_describer")
-    except Exception as exc:
-        logger.debug("describe: get_text_auxiliary_client failed: %s", exc)
-        return DescribeOutcome(canon, False, "auxiliary client unavailable")
-
-    if client is None or not aux_model:
-        return DescribeOutcome(canon, False, "no auxiliary client configured")
-
     user_msg = _USER_TEMPLATE.format(
         name=canon,
         model=(model or "(unset)"),
@@ -237,8 +225,11 @@ def describe_profile(
     )
 
     try:
-        resp = client.chat.completions.create(
-            model=aux_model,
+        # Route through call_llm so auxiliary.profile_describer.* config
+        # (provider/model/base_url, extra_body, reasoning_effort, retries)
+        # all apply — the direct-create path dropped extra_body (#35566).
+        resp = call_llm(
+            task="profile_describer",
             messages=[
                 {"role": "system", "content": _SYSTEM_PROMPT},
                 {"role": "user", "content": user_msg},
@@ -246,7 +237,6 @@ def describe_profile(
             temperature=0.3,
             max_tokens=400,
             timeout=timeout or 60,
-            extra_body=get_auxiliary_extra_body() or None,
         )
     except Exception as exc:
         logger.info("describe: API call failed for %s (%s)", canon, exc)

hermes_cli/runtime_provider.py

@@ -1975,8 +1975,15 @@ def resolve_runtime_provider(
         # Dual-path routing: Claude models use AnthropicBedrock SDK for full
         # feature parity (prompt caching, thinking budgets, adaptive thinking).
         # Non-Claude models use the Converse API for multi-model support.
+        #
+        # Exception: Bearer Token auth (AWS_BEARER_TOKEN_BEDROCK) is NOT
+        # supported by the AnthropicBedrock SDK (it only does SigV4 signing —
+        # a bearer-only setup fails at runtime with "could not resolve
+        # credentials from session"). Route these users through the Converse
+        # API regardless of model. Ref: #28156.
         _current_model = str(target_model or model_cfg.get("default") or "").strip()
-        if is_anthropic_bedrock_model(_current_model):
+        _has_bearer_token = bool(os.environ.get("AWS_BEARER_TOKEN_BEDROCK", "").strip())
+        if is_anthropic_bedrock_model(_current_model) and not _has_bearer_token:
             # Claude on Bedrock → AnthropicBedrock SDK → anthropic_messages path
             runtime = {
                 "provider": "bedrock",

hermes_cli/web_server.py

@@ -996,6 +996,9 @@ class ModelAssignment(BaseModel):
 class MoaModelSlot(BaseModel):
     provider: str = ""
     model: str = ""
+    # Optional per-slot reasoning effort. Declared so a client round-tripping
+    # the GET payload doesn't have it stripped at parse time and wiped on save.
+    reasoning_effort: Optional[str] = None
 
 
 class MoaPresetPayload(BaseModel):
@@ -1006,6 +1009,11 @@ class MoaPresetPayload(BaseModel):
     reference_temperature: Optional[float] = None
     aggregator_temperature: Optional[float] = None
     max_tokens: int = 4096
+    # Newer per-preset knobs (see moa_config._normalize_preset). Optional so
+    # older clients that never send them keep working; declared so clients
+    # that round-trip the GET payload don't silently erase hand-set values.
+    reference_max_tokens: Optional[int] = None
+    fanout: Optional[str] = None
     enabled: bool = True
 
 
@@ -1020,6 +1028,8 @@ class MoaConfigPayload(BaseModel):
     reference_temperature: Optional[float] = None
     aggregator_temperature: Optional[float] = None
     max_tokens: int = 4096
+    reference_max_tokens: Optional[int] = None
+    fanout: Optional[str] = None
     enabled: bool = True
     profile: Optional[str] = None
 
@@ -5701,7 +5711,23 @@ def get_moa_models(profile: Optional[str] = None):
 def set_moa_models(body: MoaConfigPayload, profile: Optional[str] = None):
     """Persist the Mixture-of-Agents provider/model slots."""
     try:
-        from hermes_cli.moa_config import normalize_moa_config
+        from hermes_cli.moa_config import normalize_moa_config, validate_moa_payload
+
+        def _slot_dict(slot: MoaModelSlot) -> dict:
+            # Drop unset optionals so saved slots stay minimal ({provider, model}).
+            return {k: v for k, v in slot.dict().items() if v is not None}
+
+        def _preset_dict(preset: MoaPresetPayload) -> dict:
+            return {
+                "reference_models": [_slot_dict(slot) for slot in preset.reference_models],
+                "aggregator": _slot_dict(preset.aggregator),
+                "reference_temperature": preset.reference_temperature,
+                "aggregator_temperature": preset.aggregator_temperature,
+                "max_tokens": preset.max_tokens,
+                "reference_max_tokens": preset.reference_max_tokens,
+                "fanout": preset.fanout,
+                "enabled": preset.enabled,
+            }
 
         with _profile_scope(body.profile or profile):
             cfg = load_config()
@@ -5709,27 +5735,35 @@ def set_moa_models(body: MoaConfigPayload, profile: Optional[str] = None):
                 raw = {
                     "default_preset": body.default_preset,
                     "active_preset": body.active_preset,
-                    "presets": {
-                        name: {
-                            "reference_models": [slot.dict() for slot in preset.reference_models],
-                            "aggregator": preset.aggregator.dict(),
-                            "reference_temperature": preset.reference_temperature,
-                            "aggregator_temperature": preset.aggregator_temperature,
-                            "max_tokens": preset.max_tokens,
-                            "enabled": preset.enabled,
-                        }
-                        for name, preset in body.presets.items()
-                    },
+                    "presets": {name: _preset_dict(preset) for name, preset in body.presets.items()},
                 }
             else:
-                raw = {
-                    "reference_models": [slot.dict() for slot in body.reference_models],
-                    "aggregator": body.aggregator.dict(),
-                    "reference_temperature": body.reference_temperature,
-                    "aggregator_temperature": body.aggregator_temperature,
-                    "max_tokens": body.max_tokens,
-                    "enabled": body.enabled,
-                }
+                raw = _preset_dict(
+                    MoaPresetPayload(
+                        reference_models=body.reference_models,
+                        aggregator=body.aggregator,
+                        reference_temperature=body.reference_temperature,
+                        aggregator_temperature=body.aggregator_temperature,
+                        max_tokens=body.max_tokens,
+                        reference_max_tokens=body.reference_max_tokens,
+                        fanout=body.fanout,
+                        enabled=body.enabled,
+                    )
+                )
+
+            # Reject-don't-repair: normalize_moa_config() silently swaps any
+            # preset containing incomplete slots for the hardcoded defaults —
+            # correct tolerance for hand-edited configs at READ time, silent
+            # data loss at WRITE time (#64156: desktop autosave of a
+            # half-filled slot replaced the user's whole preset). Refuse the
+            # save loudly so no client can corrupt config through this route.
+            problems = validate_moa_payload(raw)
+            if problems:
+                raise HTTPException(
+                    status_code=422,
+                    detail="Invalid MoA config: " + "; ".join(problems),
+                )
+
             normalized = normalize_moa_config(raw)
             cfg["moa"] = normalized
             save_config(cfg)
@@ -14148,7 +14182,9 @@ async def update_config_raw(body: RawConfigUpdate, profile: Optional[str] = None
         if not isinstance(parsed, dict):
             raise HTTPException(status_code=400, detail="YAML must be a mapping")
         with _profile_scope(body.profile or profile):
-            save_config(parsed)
+            # Full-document replacement: the editor owns the whole file; do not
+            # merge omitted sections back from disk (#62723).
+            save_config(parsed, merge_existing=False)
         return {"ok": True}
     except yaml.YAMLError as e:
         raise HTTPException(status_code=400, detail=f"Invalid YAML: {e}")

hermes_state.py

@@ -791,6 +791,7 @@ CREATE TABLE IF NOT EXISTS sessions (
     compression_failure_cooldown_until REAL,
     compression_failure_error TEXT,
     compression_fallback_streak INTEGER NOT NULL DEFAULT 0,
+    profile_name TEXT,
     rewind_count INTEGER NOT NULL DEFAULT 0,
     archived INTEGER NOT NULL DEFAULT 0,
     FOREIGN KEY (parent_session_id) REFERENCES sessions(id)
@@ -986,7 +987,7 @@ class SessionDB:
     _WRITE_MAX_RETRIES = 15
     _WRITE_RETRY_MIN_S = 0.020   # 20ms
     _WRITE_RETRY_MAX_S = 0.150   # 150ms
-    # Attempt a PASSIVE WAL checkpoint every N successful writes.
+    # Attempt a WAL checkpoint every N successful writes (PASSIVE mode).
     _CHECKPOINT_EVERY_N_WRITES = 50
     # Merge fragmented FTS5 segments every N successful writes. The message
     # triggers append one segment per insert; left unmaintained these grow
@@ -1301,35 +1302,34 @@ class SessionDB:
         )
 
     def _try_wal_checkpoint(self) -> None:
-        """Best-effort TRUNCATE WAL checkpoint.  Never raises.
+        """Best-effort PASSIVE WAL checkpoint.  Never raises.
 
-        Flushes committed WAL frames back into the main DB file and
-        truncates the WAL file to zero bytes.  Keeps the WAL from
-        growing unbounded when many processes hold persistent
-        connections.
+        Flushes committed WAL frames back into the main DB file without
+        requiring an exclusive lock.  PASSIVE is safe for frequent
+        periodic use because it does not block concurrent writers and
+        cannot corrupt B-tree pages under I/O pressure.
 
-        PASSIVE checkpoint was previously used here, but it never
-        truncates the WAL file — the file stays at its high-water
-        mark until an explicit TRUNCATE is called (which only
-        happened inside the infrequent vacuum()).
+        PASSIVE does not truncate the WAL file — it stays at its
+        high-water mark.  WAL truncation happens in :meth:`close`
+        (TRUNCATE) and pre-VACUUM checkpoints, which run infrequently
+        under controlled conditions.
 
-        TRUNCATE may block writers briefly while checkpointing, but
-        _try_wal_checkpoint is called off the hot path (every 50
-        writes) and already runs under ``self._lock``, so the
-        additional hold time is negligible.
+        Previous TRUNCATE strategy caused B-tree corruption on large
+        databases (65K+ pages) due to the exclusive-lock I/O pressure
+        from checkpointing thousands of frames at once (issue #45383).
         """
         try:
             with self._lock:
                 result = self._conn.execute(
-                    "PRAGMA wal_checkpoint(TRUNCATE)"
+                    "PRAGMA wal_checkpoint(PASSIVE)"
                 ).fetchone()
                 if result and result[1] > 0:
                     logger.debug(
                         "WAL checkpoint: %d/%d pages checkpointed",
                         result[2], result[1],
                     )
-        except Exception:
-            pass  # Best effort — never fatal.
+        except Exception as exc:
+            logger.warning("WAL checkpoint (PASSIVE) failed: %s", exc)
 
     def _try_optimize_fts(self) -> None:
         """Best-effort FTS5 segment merge. Never raises.
@@ -1357,8 +1357,8 @@ class SessionDB:
             if self._conn:
                 try:
                     self._conn.execute("PRAGMA wal_checkpoint(TRUNCATE)")
-                except Exception:
-                    pass
+                except Exception as exc:
+                    logger.debug("WAL checkpoint (TRUNCATE) at close failed: %s", exc)
                 self._conn.close()
                 self._conn = None
 
@@ -1760,6 +1760,7 @@ class SessionDB:
         thread_id: str = None,
         parent_session_id: str = None,
         cwd: str = None,
+        profile_name: str = None,
     ) -> None:
         """Insert a session row, enriching NULL metadata on conflict.
 
@@ -1783,9 +1784,9 @@ class SessionDB:
             conn.execute(
                 """INSERT INTO sessions (
                    id, source, user_id, session_key, chat_id, chat_type, thread_id,
-                   model, model_config, system_prompt, parent_session_id, cwd, started_at
+                   model, model_config, system_prompt, parent_session_id, cwd, profile_name, started_at
                 )
-                   VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                   VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
                    ON CONFLICT(id) DO UPDATE SET
                        model = COALESCE(sessions.model, excluded.model),
                        model_config = COALESCE(sessions.model_config, excluded.model_config),
@@ -1795,7 +1796,8 @@ class SessionDB:
                        chat_type = COALESCE(sessions.chat_type, excluded.chat_type),
                        thread_id = COALESCE(sessions.thread_id, excluded.thread_id),
                        parent_session_id = COALESCE(sessions.parent_session_id, excluded.parent_session_id),
-                       cwd = COALESCE(sessions.cwd, excluded.cwd)""",
+                       cwd = COALESCE(sessions.cwd, excluded.cwd),
+                       profile_name = COALESCE(sessions.profile_name, excluded.profile_name)""",
                 (
                     session_id,
                     source,
@@ -1809,6 +1811,7 @@ class SessionDB:
                     system_prompt,
                     parent_session_id,
                     cwd,
+                    profile_name,
                     time.time(),
                 ),
             )
@@ -7015,8 +7018,8 @@ class SessionDB:
             # Best-effort WAL checkpoint first, then VACUUM.
             try:
                 self._conn.execute("PRAGMA wal_checkpoint(TRUNCATE)")
-            except Exception:
-                pass
+            except Exception as exc:
+                logger.debug("WAL checkpoint (TRUNCATE) before VACUUM failed: %s", exc)
             self._conn.execute("VACUUM")
         return optimized
 

optional-mcps/blender/manifest.yaml

@@ -0,0 +1,88 @@
+# Nous-approved MCP catalog entry.
+# Presence in this directory = approval. Merged via PR review.
+manifest_version: 1
+
+name: blender
+description: Drive a live Blender session — modeling, scenes, and renders.
+source: https://github.com/ahujasid/blender-mcp
+
+# ahujasid/blender-mcp (MIT, PyPI: blender-mcp) is the de-facto standard
+# Blender MCP. It is a bridge with two halves:
+#   1. This stdio server (uvx blender-mcp), which Hermes launches.
+#   2. An addon (addon.py) running inside Blender that opens a local command
+#      socket on 127.0.0.1:9876. The stdio server relays to it.
+# There is nothing to git-clone on the Hermes side — uvx resolves the package —
+# so no install block. The in-Blender addon setup is covered in post_install.
+#
+# Why not Blender's official MCP (blender.org/lab): it requires Blender 5.1+
+# and targets scene analysis/documentation rather than asset creation. This
+# server works across Blender 3.x/4.x and exposes full modeling control.
+transport:
+  type: stdio
+  command: "uvx"
+  # Pinned per catalog dependency policy: exact version, and the release must
+  # be at least 2 weeks old at pin time (supply-chain cooldown — same rule as
+  # pyproject.toml dependencies). 1.6.4 released 2026-06-11. The catalog never
+  # auto-updates; bumping the pin is a PR to this manifest.
+  args:
+    - "blender-mcp==1.6.4"
+  version: "1.6.4"
+  env:
+    # Upstream ships anonymous telemetry, on by default. Hermes policy is
+    # no outbound telemetry without explicit opt-in, so the catalog entry
+    # disables it. Remove this line only if you deliberately want it on.
+    DISABLE_TELEMETRY: "true"
+
+# The Blender addon binds 127.0.0.1 only and has no auth of its own; the
+# stdio server needs no credentials. Optional integrations (Sketchfab,
+# Hyper3D Rodin, Hunyuan3D) take API keys entered in the addon's N-panel
+# inside Blender, not via Hermes env.
+auth:
+  type: none
+
+# Tool selection at install time:
+# The server advertises 22 tools. 18 of them front optional third-party asset
+# services (PolyHaven, Sketchfab, Hyper3D Rodin, Hunyuan3D) that are dead
+# weight unless the matching integration is enabled in the addon panel — and
+# upstream has no server-side trim mechanism, so without a filter every one
+# of them costs schema tokens on every API call. Default to the core surface:
+# scene/object inspection, viewport screenshots, and code execution, which is
+# the complete modeling/render capability. Users who enable an asset service
+# in the addon can opt into its tools with `hermes mcp configure blender`.
+tools:
+  default_enabled:
+    - get_scene_info
+    - get_object_info
+    - get_viewport_screenshot
+    - execute_blender_code
+
+post_install: |
+  This entry launches the bridge server, but the bridge talks to an addon
+  INSIDE a running Blender (3.0+). One-time setup:
+
+    1. Download addon.py from https://github.com/ahujasid/blender-mcp
+       (raw file: https://raw.githubusercontent.com/ahujasid/blender-mcp/main/addon.py)
+    2. Blender > Edit > Preferences > Add-ons > Install... > select addon.py,
+       then enable "Interface: Blender MCP".
+    3. In the 3D viewport press N, open the "BlenderMCP" tab, click
+       "Connect to Claude" (starts the local socket on 127.0.0.1:9876).
+
+  Blender must be RUNNING with the addon connected before the tools work —
+  start Blender first, then your Hermes session. The addon refuses to start
+  under `blender -b` (background mode): its command queue runs on UI timers.
+  On a machine without a display, run Blender under a virtual one:
+    xvfb-run blender    (or Xvfb :99 & DISPLAY=:99 blender)
+  GPU rendering (Cycles/OptiX) works fine under Xvfb.
+
+  SECURITY: execute_blender_code runs arbitrary Python inside Blender with no
+  sandbox — same trust level as the terminal tool. Upstream telemetry is
+  disabled via DISABLE_TELEMETRY in this entry's env block.
+
+  The 18 asset-service tools (PolyHaven/Sketchfab/Hyper3D/Hunyuan3D) are off
+  by default. To use one: enable the service in the addon's BlenderMCP panel
+  (API key there if required), then `hermes mcp configure blender` and tick
+  its tools. PolyHaven is free and keyless; the others need accounts.
+
+  If Hermes and Blender run on different machines, note that file paths in
+  execute_blender_code (texture loads, exports) resolve on the BLENDER host's
+  filesystem, not where Hermes runs.

optional-mcps/n8n/manifest.yaml

@@ -27,8 +27,12 @@ transport:
 install:
   type: git
   url: https://github.com/CyberSamuraiX/hermes-n8n-mcp.git
-  # Pin to a commit/tag. Required — manifests do not float HEAD.
-  ref: main
+  # Pinned per catalog dependency policy: full commit SHA (branches and tags
+  # can be moved; SHAs cannot), and the pinned commit must be at least
+  # 2 weeks old at pin time — same supply-chain rules as pyproject
+  # dependencies. This SHA is "feat: add local n8n MCP bridge for Hermes"
+  # (2026-05-23). Bumping the pin is a PR to this manifest.
+  ref: 7a9ae00795593aa1fdb4e61ecd640e8bfd0c3841
   # Bootstrap commands run inside the cloned directory after clone.
   bootstrap:
     - "python3 -m venv .venv"

optional-skills/creative/blender-mcp/SKILL.md

@@ -1,94 +1,88 @@
 ---
 name: blender-mcp
-description: Control Blender directly from Hermes via socket connection to the blender-mcp addon. Create 3D objects, materials, animations, and run arbitrary Blender Python (bpy) code. Use when user wants to create or modify anything in Blender.
-version: 1.0.0
-requires: Blender 4.3+ (desktop instance required, headless not supported)
-author: alireza78a
+description: Drive Blender via the catalog blender MCP, with bpy recipes.
+version: 2.0.0
+requires: Blender 3.0+ desktop instance (headless via xvfb-run)
+author: alireza78a + Hermes Agent
 tags: [blender, 3d, animation, modeling, bpy, mcp]
 platforms: [linux, macos, windows]
 ---
 
-# Blender MCP
+# Blender MCP Skill
 
-Control a running Blender instance from Hermes via socket on TCP port 9876.
+Companion skill for the `blender` entry in the Hermes MCP catalog. The MCP
+server provides the connection to Blender; this skill teaches the bpy idioms
+and pitfalls for driving it well. It does not cover Blender UI workflows —
+everything here goes through the MCP tools against a live Blender session.
 
-## Setup (one-time)
+## When to Use
 
-### 1. Install the Blender addon
+Use when the user wants to create or modify anything in a running Blender
+instance: meshes, materials, animations, lighting, renders. Requires the
+blender MCP server installed and a Blender desktop session with the addon
+connected.
 
-    curl -sL https://raw.githubusercontent.com/ahujasid/blender-mcp/main/addon.py -o ~/Desktop/blender_mcp_addon.py
+## Prerequisites
 
-In Blender:
-    Edit > Preferences > Add-ons > Install > select blender_mcp_addon.py
-    Enable "Interface: Blender MCP"
+1. Install the MCP server from the Nous catalog (one-time):
 
-### 2. Start the socket server in Blender
+       hermes mcp install blender
 
-Press N in Blender viewport to open sidebar.
-Find "BlenderMCP" tab and click "Start Server".
+   This configures the pinned `blender-mcp` stdio server with the curated
+   tool set: `get_scene_info`, `get_object_info`, `get_viewport_screenshot`,
+   `execute_blender_code`.
 
-### 3. Verify connection
+2. Install the addon inside Blender (one-time — the catalog entry's
+   post-install notes cover this too):
+   - Download https://raw.githubusercontent.com/ahujasid/blender-mcp/main/addon.py
+   - Blender > Edit > Preferences > Add-ons > Install... > select addon.py,
+     enable "Interface: Blender MCP".
 
-    nc -z -w2 localhost 9876 && echo "OPEN" || echo "CLOSED"
+3. Every session: start Blender FIRST, press N in the viewport, open the
+   "BlenderMCP" tab, click "Connect to Claude" (starts the local bridge
+   socket). Then start your Hermes session so the MCP tools are loaded.
 
-## Protocol
+   The addon refuses to start under `blender -b` (background mode). On a
+   machine without a display, run Blender under a virtual one:
+   `xvfb-run blender`. GPU rendering works fine under Xvfb.
 
-Plain UTF-8 JSON over TCP -- no length prefix.
+## Quick Reference
 
-Send:     {"type": "<command>", "params": {<kwargs>}}
-Receive:  {"status": "success", "result": <value>}
-          {"status": "error",   "message": "<reason>"}
+| MCP tool                  | Use for                                    |
+|---------------------------|--------------------------------------------|
+| `get_scene_info`          | List objects before touching the scene     |
+| `get_object_info`         | Inspect one object (transform, materials)  |
+| `get_viewport_screenshot` | Visual check of what you built             |
+| `execute_blender_code`    | Everything else — arbitrary bpy Python     |
 
-## Available Commands
+Optional asset-service tools (PolyHaven, Sketchfab, Hyper3D, Hunyuan3D) are
+disabled by default. If the user has enabled a service in the addon panel,
+opt into its tools with `hermes mcp configure blender`.
 
-| type                    | params            | description                     |
-|-------------------------|-------------------|---------------------------------|
-| execute_code            | code (str)        | Run arbitrary bpy Python code   |
-| get_scene_info          | (none)            | List all objects in scene       |
-| get_object_info         | object_name (str) | Details on a specific object    |
-| get_viewport_screenshot | (none)            | Screenshot of current viewport  |
+## Procedure
 
-## Python Helper
+1. Call `get_scene_info` first — never assume the scene is empty.
+2. Build with `execute_blender_code`, in small focused calls (one logical
+   step per call: add objects, then materials, then animation). Large
+   monolithic scripts hit the bridge timeout.
+3. Verify visually with `get_viewport_screenshot` between major steps.
+4. Render to an absolute path and tell the user where the file is.
 
-Use this inside execute_code tool calls:
+### Common bpy Patterns
 
-    import socket, json
+Clear scene:
 
-    def blender_exec(code: str, host="localhost", port=9876, timeout=15):
-        s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
-        s.connect((host, port))
-        s.settimeout(timeout)
-        payload = json.dumps({"type": "execute_code", "params": {"code": code}})
-        s.sendall(payload.encode("utf-8"))
-        buf = b""
-        while True:
-            try:
-                chunk = s.recv(4096)
-                if not chunk:
-                    break
-                buf += chunk
-                try:
-                    json.loads(buf.decode("utf-8"))
-                    break
-                except json.JSONDecodeError:
-                    continue
-            except socket.timeout:
-                break
-        s.close()
-        return json.loads(buf.decode("utf-8"))
-
-## Common bpy Patterns
-
-### Clear scene
     bpy.ops.object.select_all(action='SELECT')
     bpy.ops.object.delete()
 
-### Add mesh objects
+Add mesh objects:
+
     bpy.ops.mesh.primitive_uv_sphere_add(radius=1, location=(0, 0, 0))
     bpy.ops.mesh.primitive_cube_add(size=2, location=(3, 0, 0))
     bpy.ops.mesh.primitive_cylinder_add(radius=0.5, depth=2, location=(-3, 0, 0))
 
-### Create and assign material
+Create and assign material:
+
     mat = bpy.data.materials.new(name="MyMat")
     mat.use_nodes = True
     bsdf = mat.node_tree.nodes.get("Principled BSDF")
@@ -97,21 +91,43 @@ Use this inside execute_code tool calls:
     bsdf.inputs["Metallic"].default_value = 0.0
     obj.data.materials.append(mat)
 
-### Keyframe animation
+Keyframe animation:
+
     obj.location = (0, 0, 0)
     obj.keyframe_insert(data_path="location", frame=1)
     obj.location = (0, 0, 3)
     obj.keyframe_insert(data_path="location", frame=60)
 
-### Render to file
+Render to file:
+
     bpy.context.scene.render.filepath = "/tmp/render.png"
     bpy.context.scene.render.engine = 'CYCLES'
     bpy.ops.render.render(write_still=True)
 
 ## Pitfalls
 
-- Must check socket is open before running (nc -z localhost 9876)
-- Addon server must be started inside Blender each session (N-panel > BlenderMCP > Connect)
-- Break complex scenes into multiple smaller execute_code calls to avoid timeouts
-- Render output path must be absolute (/tmp/...) not relative
-- shade_smooth() requires object to be selected and in object mode
+- The blender MCP tools only exist if the server is installed and the session
+  started after install. If they're missing, run `hermes mcp install blender`
+  and start a new session.
+- The addon bridge must be (re)connected inside Blender each Blender session
+  (N-panel > BlenderMCP > Connect). "Connection refused" from the tools means
+  Blender isn't running or the addon isn't connected — fix that, don't retry.
+- Break complex scenes into multiple smaller `execute_blender_code` calls to
+  avoid bridge timeouts.
+- Render output paths must be absolute (`/tmp/render.png`), not relative —
+  they resolve on the BLENDER host's filesystem, which matters if Hermes and
+  Blender run on different machines.
+- `shade_smooth()` requires the object to be selected and in object mode.
+- `execute_blender_code` runs arbitrary Python inside Blender with no sandbox
+  — same trust level as the `terminal` tool. Don't paste untrusted code into
+  it.
+- Do NOT hand-roll raw TCP JSON to port 9876 from `execute_code` — that was
+  this skill's pre-MCP workaround. It bypasses the catalog's version pinning
+  and tool curation. The MCP tools are the supported path.
+
+## Verification
+
+- `get_scene_info` returns the expected object list after each build step.
+- `get_viewport_screenshot` shows the scene you intended.
+- After a render, confirm the output file exists and report its absolute
+  path to the user.

plugins/model-providers/custom/__init__.py

@@ -4,7 +4,9 @@ Covers any endpoint registered as provider="custom", including local
 Ollama instances and OpenAI-compatible reasoning endpoints (GLM-5.2 on
 Volcengine ARK, vLLM, llama.cpp). Key quirks:
   - ollama_num_ctx → extra_body.options.num_ctx (local context window)
-  - reasoning_config disabled → extra_body.think = False
+  - reasoning_config disabled → top-level reasoning_effort="none"
+    (Ollama /v1/chat/completions ignores think=False — ollama#14820)
+    + extra_body.think = False for /api/chat and proxies
   - reasoning_config enabled + effort → top-level reasoning_effort
     (the native OpenAI-compatible format GLM/ARK expect; unset omits it
     so the endpoint's server default applies)
@@ -53,6 +55,13 @@ class CustomProfile(ProviderProfile):
             _effort = (reasoning_config.get("effort") or "").strip().lower()
             _enabled = reasoning_config.get("enabled", True)
             if _effort == "none" or _enabled is False:
+                # Ollama's /v1/chat/completions silently ignores
+                # extra_body.think (only /api/chat honours it — ollama#14820)
+                # but respects the top-level reasoning_effort field, so both
+                # are needed to actually stop a thinking-capable model from
+                # reasoning (#25758). Endpoints that recognize neither simply
+                # ignore them.
+                top_level["reasoning_effort"] = "none"
                 extra_body["think"] = False
             elif _effort:
                 top_level["reasoning_effort"] = _effort

plugins/platforms/discord/adapter.py

@@ -6612,12 +6612,20 @@ class DiscordAdapter(BasePlatformAdapter):
     # ------------------------------------------------------------------
 
     def _text_batch_key(self, event: MessageEvent) -> str:
-        """Session-scoped key for text message batching."""
+        """Session-scoped key for text message batching.
+
+        Passes ``event.source.profile`` through so routed messages batch
+        under the same namespace the agent run will use (e.g.
+        ``agent:crypto-trader`` instead of ``agent:main``). Without this,
+        the batch key would always land in ``agent:main`` even when the
+        routed profile differs.
+        """
         from gateway.session import build_session_key
         return build_session_key(
             event.source,
             group_sessions_per_user=self.config.extra.get("group_sessions_per_user", True),
             thread_sessions_per_user=self.config.extra.get("thread_sessions_per_user", False),
+            profile=event.source.profile,
         )
 
     def _enqueue_text_event(self, event: MessageEvent) -> None:

plugins/platforms/feishu/adapter.py

@@ -3633,6 +3633,7 @@ class FeishuAdapter(BasePlatformAdapter):
             event.source,
             group_sessions_per_user=self.config.extra.get("group_sessions_per_user", True),
             thread_sessions_per_user=self.config.extra.get("thread_sessions_per_user", False),
+            profile=event.source.profile,
         )
 
     @staticmethod

plugins/platforms/matrix/adapter.py

@@ -3455,6 +3455,7 @@ class MatrixAdapter(BasePlatformAdapter):
             thread_sessions_per_user=self.config.extra.get(
                 "thread_sessions_per_user", False
             ),
+            profile=event.source.profile,
         )
 
     def _enqueue_text_event(self, event: MessageEvent) -> None:

plugins/platforms/telegram/adapter.py

@@ -7925,6 +7925,7 @@ class TelegramAdapter(BasePlatformAdapter):
             event.source,
             group_sessions_per_user=self.config.extra.get("group_sessions_per_user", True),
             thread_sessions_per_user=self.config.extra.get("thread_sessions_per_user", False),
+            profile=event.source.profile,
         )
 
     def _enqueue_text_event(self, event: MessageEvent) -> None:

plugins/platforms/wecom/adapter.py

@@ -578,6 +578,7 @@ class WeComAdapter(BasePlatformAdapter):
             event.source,
             group_sessions_per_user=self.config.extra.get("group_sessions_per_user", True),
             thread_sessions_per_user=self.config.extra.get("thread_sessions_per_user", False),
+            profile=event.source.profile,
         )
 
     def _enqueue_text_event(self, event: MessageEvent) -> None:

plugins/platforms/whatsapp/adapter.py

@@ -1278,6 +1278,7 @@ class WhatsAppAdapter(WhatsAppBehaviorMixin, BasePlatformAdapter):
             event.source,
             group_sessions_per_user=self.config.extra.get("group_sessions_per_user", True),
             thread_sessions_per_user=self.config.extra.get("thread_sessions_per_user", False),
+            profile=event.source.profile,
         )
 
     def _enqueue_text_event(self, event: MessageEvent) -> None:

run_agent.py

@@ -599,6 +599,13 @@ class AIAgent:
             return
         source = _session_source_for_agent(self.platform)
         try:
+            try:
+                from hermes_cli.profiles import get_active_profile_name
+                _profile_for_session = get_active_profile_name()
+                if _profile_for_session == "default":
+                    _profile_for_session = None
+            except Exception:
+                _profile_for_session = None
             self._session_db.create_session(
                 session_id=self.session_id,
                 source=source,
@@ -608,6 +615,7 @@ class AIAgent:
                 user_id=None,
                 parent_session_id=self._parent_session_id,
                 cwd=_launch_cwd_for_session(source),
+                profile_name=_profile_for_session,
             )
             self._session_db_created = True
         except Exception as e:

scripts/release.py

@@ -45,7 +45,10 @@ ACP_REGISTRY_MANIFEST = REPO_ROOT / "acp_registry" / "agent.json"
 
 # Auto-extracted from noreply emails + manual overrides
 AUTHOR_MAP = {
+    "Burgunthy@users.noreply.github.com": "Burgunthy",  # PR #20096 salvage (gateway: profile-based routing for inbound messages)
     "75556242+webtecnica@users.noreply.github.com": "webtecnica",  # PR #63360 salvage (nous: restore inference-api base_url)
+    "skosarevivan@yandex.ru": "Epoxidex",  # PR #29820 salvage (ollama: top-level reasoning_effort=none; #25758)
+    "jdjiayou@163.com": "JiaDe-Wu",  # PR #34742 salvage (bedrock: bearer routing + streaming fallback + image decode; #28156)
     "changhyun.min@gmail.com": "minchang",  # PR #42231 salvage (providers: add Upstage Solar)
     "neo@neodeMac-mini.local": "neo-claw-bot",  # PR #58465 salvage (moa: drop empty user turns from advisory view)
     "2024104039@mails.szu.edu.cn": "pixel4039",  # PR #64420 salvage (streaming: retry zero-chunk streams)
@@ -334,6 +337,7 @@ AUTHOR_MAP = {
     "290859878+synapsesx@users.noreply.github.com": "synapsesx",
     "157689911+itsflownium@users.noreply.github.com": "itsflownium",
     "dirtyren@users.noreply.github.com": "dirtyren",
+    "dorokuma@users.noreply.github.com": "dorokuma",
     "liuwei666888@users.noreply.github.com": "liuwei666888",
     "527711370@qq.com": "liuwei666888",
     "217401759+justinschille@users.noreply.github.com": "justinschille",

tests/agent/test_auxiliary_client.py

@@ -3380,6 +3380,81 @@ class TestAuxiliaryTaskExtraBody:
         }
         mock_create.assert_called_once()
 
+    def _run_anthropic_adapter(self, *, call_extra_body=None, bak_result=None):
+        """Drive _AnthropicCompletionsAdapter.create() with mocked SDK layers;
+        return the api_kwargs handed to create_anthropic_message."""
+        from agent.auxiliary_client import _AnthropicCompletionsAdapter
+
+        adapter = _AnthropicCompletionsAdapter(MagicMock(), "claude-sonnet-4-6", is_oauth=False)
+        bak_result = bak_result or {
+            "model": "claude-sonnet-4-6", "messages": [], "max_tokens": 64,
+        }
+        with patch("agent.anthropic_adapter.build_anthropic_kwargs",
+                   return_value=dict(bak_result)), \
+             patch("agent.anthropic_adapter.create_anthropic_message") as mock_create, \
+             patch("agent.transports.get_transport") as mock_gt:
+            mock_gt.return_value.normalize_response.return_value = MagicMock(
+                content="ok", tool_calls=None, reasoning=None, finish_reason="stop",
+                usage=None, provider_data=None,
+            )
+            kwargs = {
+                "model": "claude-sonnet-4-6",
+                "messages": [{"role": "user", "content": "hi"}],
+                "max_tokens": 64,
+            }
+            if call_extra_body is not None:
+                kwargs["extra_body"] = call_extra_body
+            adapter.create(**kwargs)
+        return mock_create.call_args.args[1]
+
+    def test_anthropic_aux_extra_body_passthrough(self):
+        """Bug B (#37217): vendor fields in extra_body reach the Anthropic SDK."""
+        api_kwargs = self._run_anthropic_adapter(
+            call_extra_body={"thinking": {"type": "disabled"}, "metadata": {"user_id": "u1"}},
+        )
+        assert api_kwargs["extra_body"] == {
+            "thinking": {"type": "disabled"}, "metadata": {"user_id": "u1"},
+        }
+
+    def test_anthropic_aux_extra_body_excludes_reasoning_and_private_keys(self):
+        """The OpenAI-shaped reasoning dict is translated (not forwarded), and
+        private _-prefixed plumbing keys never reach the wire."""
+        api_kwargs = self._run_anthropic_adapter(
+            call_extra_body={
+                "reasoning": {"enabled": True, "effort": "low"},
+                "_internal": "plumbing",
+                "metadata": {"user_id": "u1"},
+            },
+        )
+        assert api_kwargs["extra_body"] == {"metadata": {"user_id": "u1"}}
+
+    def test_anthropic_aux_extra_body_merges_over_existing(self):
+        """Caller extra_body merges on top of what build_anthropic_kwargs
+        already emitted (fast-mode speed) instead of clobbering it."""
+        api_kwargs = self._run_anthropic_adapter(
+            call_extra_body={"metadata": {"user_id": "u1"}},
+            bak_result={
+                "model": "claude-sonnet-4-6", "messages": [], "max_tokens": 64,
+                "extra_body": {"speed": "fast"},
+            },
+        )
+        assert api_kwargs["extra_body"] == {
+            "speed": "fast", "metadata": {"user_id": "u1"},
+        }
+
+    def test_anthropic_aux_no_extra_body_unchanged(self):
+        """Regression guard: no caller extra_body -> kwargs identical to before."""
+        api_kwargs = self._run_anthropic_adapter(call_extra_body=None)
+        assert "extra_body" not in api_kwargs
+
+    def test_anthropic_aux_reasoning_only_extra_body_adds_nothing(self):
+        """extra_body containing ONLY the reasoning key must not create an
+        empty extra_body dict on the wire."""
+        api_kwargs = self._run_anthropic_adapter(
+            call_extra_body={"reasoning": {"enabled": False}},
+        )
+        assert "extra_body" not in api_kwargs
+
     def test_no_warning_when_provider_is_custom(self, monkeypatch, caplog):
         """No warning when the provider is 'custom' — OPENAI_BASE_URL is expected."""
         import agent.auxiliary_client as mod

tests/agent/test_auxiliary_client_bootstrap_skew.py

@@ -0,0 +1,47 @@
+"""Regression for #64333 — auxiliary client must survive a version-skewed
+agent.process_bootstrap that lacks build_keepalive_http_client.
+
+Desktop installs can run a runtime whose agent/process_bootstrap.py predates
+the helper while newer callers (cron scheduler → auxiliary_client) expect it.
+Before the fix the module-level import made every cron job die with
+ImportError before any agent logic ran.
+"""
+
+import builtins
+import logging
+
+import agent.auxiliary_client as aux
+
+
+def test_missing_bootstrap_helper_degrades_instead_of_raising(monkeypatch, caplog):
+    """ImportError from process_bootstrap → empty kwargs + one-time warning."""
+    real_import = builtins.__import__
+
+    def _fail_bootstrap(name, *args, **kwargs):
+        if name == "agent.process_bootstrap":
+            raise ImportError(
+                "cannot import name 'build_keepalive_http_client' "
+                "from 'agent.process_bootstrap'"
+            )
+        return real_import(name, *args, **kwargs)
+
+    monkeypatch.setattr(builtins, "__import__", _fail_bootstrap)
+    monkeypatch.setattr(aux, "_WARNED_KEEPALIVE_IMPORT_SKEW", False)
+
+    with caplog.at_level(logging.WARNING, logger="agent.auxiliary_client"):
+        result = aux._openai_http_client_kwargs("https://api.example.com/v1")
+        again = aux._openai_http_client_kwargs("https://api.example.com/v1")
+
+    assert result == {}
+    assert again == {}
+    skew_warnings = [
+        r for r in caplog.records if "mixed/stale install" in r.getMessage()
+    ]
+    assert len(skew_warnings) == 1  # warned once, not per call
+
+
+def test_healthy_bootstrap_still_injects_keepalive_client():
+    """With the helper present, the keepalive http_client is injected."""
+    result = aux._openai_http_client_kwargs("https://api.example.com/v1")
+    assert "http_client" in result
+    assert result["http_client"] is not None

tests/agent/test_bedrock_adapter.py

@@ -1712,3 +1712,85 @@ class TestRequireBoto3VersionCheck:
         with patch.dict("sys.modules", {"boto3": fake_boto3}):
             result = _require_boto3()
             assert result is fake_boto3
+
+class TestImageBase64Decoding:
+    """Image data URLs must be decoded to raw bytes before passing to Converse API.
+
+    boto3 re-encodes at the wire layer, so passing the base64 string directly
+    results in double-encoding. Bedrock rejects with 'Failed to sanitize image'.
+    Ref: #33317.
+    """
+
+    def test_data_url_decoded_to_bytes(self):
+        from agent.bedrock_adapter import _convert_content_to_converse
+        import base64
+
+        # A tiny 1x1 red PNG
+        raw_png = base64.b64decode(
+            "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8/5+hHgAHggJ/PchI7wAAAABJRU5ErkJggg=="
+        )
+        data_url = "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8/5+hHgAHggJ/PchI7wAAAABJRU5ErkJggg=="
+
+        content = [{"type": "image_url", "image_url": {"url": data_url}}]
+        blocks = _convert_content_to_converse(content)
+
+        assert len(blocks) == 1
+        img_block = blocks[0]["image"]
+        assert img_block["format"] == "png"
+        # Must be raw bytes, not a base64 string
+        assert isinstance(img_block["source"]["bytes"], bytes)
+        assert img_block["source"]["bytes"] == raw_png
+
+    def test_invalid_base64_falls_back_to_encode(self):
+        from agent.bedrock_adapter import _convert_content_to_converse
+
+        data_url = "data:image/jpeg;base64,NOT_VALID_BASE64!!!"
+        content = [{"type": "image_url", "image_url": {"url": data_url}}]
+        blocks = _convert_content_to_converse(content)
+
+        # Should not crash — falls back to encoding the string as bytes
+        assert len(blocks) == 1
+        assert isinstance(blocks[0]["image"]["source"]["bytes"], bytes)
+
+
+class TestBearerTokenRoutesToConverse:
+    """Bearer Token users must go through Converse API, not AnthropicBedrock SDK.
+
+    The AnthropicBedrock SDK only supports SigV4 signing — it cannot use
+    AWS_BEARER_TOKEN_BEDROCK. Ref: #28156.
+    """
+
+    def _resolve(self, monkeypatch, *, bearer: bool):
+        import os
+
+        from hermes_cli import runtime_provider as rp
+
+        if bearer:
+            monkeypatch.setenv("AWS_BEARER_TOKEN_BEDROCK", "test-bearer-token-123")
+        else:
+            monkeypatch.delenv("AWS_BEARER_TOKEN_BEDROCK", raising=False)
+        monkeypatch.setenv("AWS_DEFAULT_REGION", "us-east-1")
+        assert "AWS_BEARER_TOKEN_BEDROCK" in os.environ or not bearer
+
+        monkeypatch.setattr(
+            rp,
+            "_get_model_config",
+            lambda: {
+                "default": "us.anthropic.claude-sonnet-4-6",
+                "provider": "bedrock",
+            },
+        )
+        monkeypatch.setattr(rp, "load_config", lambda: {"bedrock": {}})
+        return rp.resolve_runtime_provider(requested="bedrock")
+
+    def test_bearer_token_forces_converse_for_claude(self, monkeypatch):
+        """Claude model + Bearer Token → bedrock_converse, not anthropic_messages."""
+        runtime = self._resolve(monkeypatch, bearer=True)
+        assert runtime["api_mode"] == "bedrock_converse"
+        assert "bedrock_anthropic" not in runtime
+
+    def test_sigv4_claude_still_uses_anthropic_bedrock_sdk(self, monkeypatch):
+        """Without a bearer token, Claude keeps the AnthropicBedrock SDK path."""
+        runtime = self._resolve(monkeypatch, bearer=False)
+        assert runtime["api_mode"] == "anthropic_messages"
+        assert runtime.get("bedrock_anthropic") is True

tests/agent/test_codex_responses_adapter.py

@@ -79,6 +79,21 @@ def test_normalize_codex_response_treats_summary_only_reasoning_as_incomplete():
     assert assistant_message.codex_reasoning_items is None
 
 
+def test_normalize_codex_response_maps_incomplete_content_filter_to_refusal():
+    response = SimpleNamespace(
+        status="incomplete",
+        incomplete_details=SimpleNamespace(reason="content_filter"),
+        output=[],
+        output_text="",
+    )
+
+    assistant_message, finish_reason = _normalize_codex_response(response)
+
+    assert finish_reason == "content_filter"
+    assert assistant_message.content == ""
+    assert response.output
+
+
 # ---------------------------------------------------------------------------
 # Server-side built-in tool calls (xAI native web_search, code interpreter,
 # etc.) come back as discrete ``*_call`` output items that xAI's
@@ -493,14 +508,21 @@ def test_normalize_codex_response_salvage_strips_closing_tag():
 
 
 def test_normalize_codex_response_salvage_is_xai_scoped():
-    """Non-xAI issuers keep the reasoning-only → incomplete classification;
-    the Codex backend replays encrypted reasoning, so its continuation
-    genuinely progresses and must not be short-circuited."""
+    """Non-xAI special-cased issuers (Codex backend) keep the reasoning-only →
+    incomplete classification; the Codex backend replays encrypted reasoning,
+    so its continuation genuinely progresses and must not be short-circuited.
+
+    Pins ``issuer_kind="codex_backend"`` explicitly: with no issuer at all,
+    the unrecognized-backend rule (#64434) trusts ``status="completed"`` and
+    returns ``stop`` — that path is covered by the #64434 regression tests.
+    """
     response = _xai_reasoning_only_response(
         "Thinking.\n<response>The answer.</response>"
     )
 
-    assistant_message, finish_reason = _normalize_codex_response(response)
+    assistant_message, finish_reason = _normalize_codex_response(
+        response, issuer_kind="codex_backend"
+    )
 
     assert finish_reason == "incomplete"
     assert assistant_message.content == ""

tests/agent/test_codex_ttfb_watchdog.py

@@ -462,3 +462,143 @@ def test_large_codex_request_strict_ttfb_env_still_reconnects(tmp_path, monkeypa
         assert "codex_ttfb_kill" in closes
     finally:
         stop["flag"] = True
+
+
+def test_large_codex_request_hard_ceiling_reclaims_silent_stall(tmp_path, monkeypatch):
+    """#64507 regression: a large Codex request (TTFB watchdog disabled by the
+    size gate, stale floor *raised*) that never emits a single byte must still
+    be reclaimed at a finite hard ceiling — not hang for 13+ minutes while the
+    worker stays idle and the session shows as active.
+
+    Uses the real default TTFB threshold (120s) and asserts the request dies at
+    the hard ceiling regardless of the size-based TTFB disable.
+    """
+    from agent import chat_completion_helpers as h
+
+    agent = _make_codex_agent(tmp_path, monkeypatch)
+    # Real default TTFB threshold (no HERMES_CODEX_TTFB_* override) → for a
+    # >10k-token request the no-byte TTFB watchdog is auto-disabled.
+    monkeypatch.setenv("HERMES_CODEX_HARD_TIMEOUT_SECONDS", "3")
+
+    closes: list = []
+    dummy_client = SimpleNamespace()
+    monkeypatch.setattr(agent, "_create_request_openai_client", lambda **k: dummy_client)
+    monkeypatch.setattr(
+        agent, "_abort_request_openai_client",
+        lambda c, reason=None: closes.append(reason),
+    )
+    monkeypatch.setattr(
+        agent, "_close_request_openai_client",
+        lambda c, reason=None: closes.append(reason),
+    )
+
+    stop = {"flag": False}
+
+    def fake_hang(api_kwargs, client=None, on_first_delta=None):
+        # No event marker AND no event ever: the exact issue-64507 stall.
+        deadline = time.time() + 120
+        while time.time() < deadline and not stop["flag"] and not agent._interrupt_requested:
+            time.sleep(0.02)
+        raise RuntimeError("connection closed")
+
+    monkeypatch.setattr(agent, "_run_codex_stream", fake_hang)
+
+    large_input = "x" * 44_000  # ~11k estimated tokens → TTFB disabled, stale raised
+    t0 = time.time()
+    try:
+        with pytest.raises(TimeoutError) as excinfo:
+            h.interruptible_api_call(agent, {"model": "gpt-5.5", "input": large_input})
+        elapsed = time.time() - t0
+        # Must die at the hard ceiling (3s), nowhere near the raised stale floor.
+        assert elapsed < 30, f"hard ceiling took {elapsed:.1f}s — stall not reclaimed"
+        assert "stale_call_kill" in closes, f"stale kill expected, got {closes}"
+        assert "timed out after" in str(excinfo.value)
+        assert "with no response" in str(excinfo.value)
+    finally:
+        stop["flag"] = True
+
+
+def test_large_codex_request_hard_ceiling_disabled_restores_legacy(tmp_path, monkeypatch):
+    """Setting HERMES_CODEX_HARD_TIMEOUT_SECONDS=0 disables the ceiling entirely,
+    restoring the pre-#64507 behavior (request waits out the raised stale floor
+    instead of being capped). Keeps the knob for operators who must.
+    """
+    from agent import chat_completion_helpers as h
+
+    agent = _make_codex_agent(tmp_path, monkeypatch)
+    monkeypatch.setenv("HERMES_CODEX_HARD_TIMEOUT_SECONDS", "0")
+
+    closes: list = []
+    dummy_client = SimpleNamespace()
+    monkeypatch.setattr(agent, "_create_request_openai_client", lambda **k: dummy_client)
+    monkeypatch.setattr(
+        agent, "_abort_request_openai_client",
+        lambda c, reason=None: closes.append(reason),
+    )
+    monkeypatch.setattr(
+        agent, "_close_request_openai_client",
+        lambda c, reason=None: closes.append(reason),
+    )
+
+    sentinel = SimpleNamespace(ok=True)
+
+    def fake_stream(api_kwargs, client=None, on_first_delta=None):
+        # No event, but only briefly — well under the (here 60s) stale timeout.
+        time.sleep(2.0)
+        return sentinel
+
+    monkeypatch.setattr(agent, "_run_codex_stream", fake_stream)
+
+    large_input = "x" * 44_000
+    resp = h.interruptible_api_call(agent, {"model": "gpt-5.5", "input": large_input})
+    assert resp is sentinel
+    assert "codex_ttfb_kill" not in closes
+    assert "stale_call_kill" not in closes
+
+
+def test_large_codex_request_hard_ceiling_caps_raised_stale_floor(tmp_path, monkeypatch):
+    """The hard ceiling must cap the raised stale floor (openai-codex can push
+    the stale timeout to 1200s at >100k tokens). A large silent stall must die
+    at the ceiling, proving the min() wins over the floor.
+    """
+    from agent import chat_completion_helpers as h
+
+    agent = _make_codex_agent(tmp_path, monkeypatch)
+    monkeypatch.setenv("HERMES_CODEX_HARD_TIMEOUT_SECONDS", "4")
+    # Force the >100k-token tier so openai_codex_stale_timeout_floor returns 1200s.
+    monkeypatch.setattr(
+        agent, "_compute_non_stream_stale_timeout", lambda *a, **k: 1200.0
+    )
+
+    closes: list = []
+    dummy_client = SimpleNamespace()
+    monkeypatch.setattr(agent, "_create_request_openai_client", lambda **k: dummy_client)
+    monkeypatch.setattr(
+        agent, "_abort_request_openai_client",
+        lambda c, reason=None: closes.append(reason),
+    )
+    monkeypatch.setattr(
+        agent, "_close_request_openai_client",
+        lambda c, reason=None: closes.append(reason),
+    )
+
+    stop = {"flag": False}
+
+    def fake_hang(api_kwargs, client=None, on_first_delta=None):
+        deadline = time.time() + 200
+        while time.time() < deadline and not stop["flag"] and not agent._interrupt_requested:
+            time.sleep(0.02)
+        raise RuntimeError("connection closed")
+
+    monkeypatch.setattr(agent, "_run_codex_stream", fake_hang)
+
+    huge_input = "x" * 500_000  # ~125k tokens → stale floor 1200s
+    t0 = time.time()
+    try:
+        with pytest.raises(TimeoutError):
+            h.interruptible_api_call(agent, {"model": "gpt-5.5", "input": huge_input})
+        elapsed = time.time() - t0
+        assert elapsed < 40, f"hard ceiling lost to stale floor: {elapsed:.1f}s"
+        assert "stale_call_kill" in closes, f"stale kill expected, got {closes}"
+    finally:
+        stop["flag"] = True

tests/agent/test_compressor_media_stripping.py

@@ -54,3 +54,77 @@ class TestMediaDirectiveStripping:
         result = compressor._serialize_for_summary(turns)
         assert "MEDIA:" not in result
         assert result.count("[media attachment]") == 2
+
+    def test_multimodal_list_content_does_not_crash(self, compressor):
+        """content as a list (multimodal) must be flattened to clean text.
+
+        Without flattening, str() coercion in redact_sensitive_text dumps
+        the raw part-dict repr — including base64 image data — into the
+        summarizer input, burning summary budget on noise.
+        """
+        turns = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What is in this image?"},
+                    {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc123"}},
+                ],
+            },
+        ]
+        result = compressor._serialize_for_summary(turns)
+        assert "What is in this image?" in result
+        assert "[image]" in result
+        assert "base64" not in result
+
+    def test_multimodal_remote_image_keeps_url(self, compressor):
+        """http(s) image parts keep their URL as a referenceable handle."""
+        turns = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "look at this"},
+                    {"type": "image_url", "image_url": {"url": "https://example.com/a.png"}},
+                ],
+            },
+        ]
+        result = compressor._serialize_for_summary(turns)
+        assert "[image: https://example.com/a.png]" in result
+
+    def test_multimodal_unknown_part_type_keeps_marker(self, compressor):
+        """Unknown part types are not silently dropped."""
+        turns = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "see attachment"},
+                    {"type": "document", "title": "spec.pdf"},
+                ],
+            },
+        ]
+        result = compressor._serialize_for_summary(turns)
+        assert "see attachment" in result
+        assert "[document]" in result
+
+    def test_multimodal_list_text_parts_extracted(self, compressor):
+        """Text parts from multimodal list content are preserved in output."""
+        turns = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "first part"},
+                    {"type": "text", "text": "second part"},
+                ],
+            },
+        ]
+        result = compressor._serialize_for_summary(turns)
+        assert "first part" in result
+        assert "second part" in result
+
+    def test_multimodal_list_bare_strings_handled(self, compressor):
+        """Bare strings inside a content list are joined."""
+        turns = [
+            {"role": "user", "content": ["hello", "world"]},
+        ]
+        result = compressor._serialize_for_summary(turns)
+        assert "hello" in result
+        assert "world" in result

tests/agent/test_context_compressor.py

@@ -1,5 +1,6 @@
 """Tests for agent/context_compressor.py — compression logic, thresholds, truncation fallback."""
 
+import json
 import pytest
 import time
 from unittest.mock import patch, MagicMock
@@ -9,6 +10,7 @@ from agent.context_compressor import (
     HISTORICAL_TASK_HEADING,
     SUMMARY_PREFIX,
     COMPRESSED_SUMMARY_METADATA_KEY,
+    _summarize_tool_result,
 )
 from hermes_state import SessionDB
 
@@ -27,6 +29,49 @@ def compressor():
         return c
 
 
+class TestSummarizeToolResultWebExtract:
+    """Pre-compression pruning must survive web_extract calls whose ``urls`` are
+    web_search result dicts ({"url"/"href": ...}), which models routinely forward
+    straight into web_extract.
+    """
+
+    CONTENT = "x" * 500  # >200 chars so the pruning pass actually summarizes
+
+    def test_multiple_dict_urls_do_not_crash(self):
+        # Two dict URLs previously hit ``dict + str`` -> TypeError, aborting
+        # _prune_old_tool_results() (and thus compress()).
+        args = json.dumps({
+            "urls": [
+                {"url": "https://example.com/a", "title": "A"},
+                {"url": "https://example.org/b", "title": "B"},
+            ]
+        })
+        summary = _summarize_tool_result("web_extract", args, self.CONTENT)
+        assert summary == "[web_extract] https://example.com/a (+1 more) (500 chars)"
+
+    def test_single_dict_url_is_unwrapped_not_stringified(self):
+        args = json.dumps({"urls": [{"url": "https://example.com/a", "title": "A"}]})
+        summary = _summarize_tool_result("web_extract", args, self.CONTENT)
+        assert summary == "[web_extract] https://example.com/a (500 chars)"
+        assert "{" not in summary  # no raw dict repr leaked into the summary
+
+    def test_href_key_is_unwrapped(self):
+        args = json.dumps({"urls": [{"href": "https://example.com/h"}]})
+        summary = _summarize_tool_result("web_extract", args, self.CONTENT)
+        assert summary == "[web_extract] https://example.com/h (500 chars)"
+
+    def test_malformed_dict_falls_back_to_placeholder(self):
+        args = json.dumps({"urls": [{"title": "no url here"}, {"title": "still none"}]})
+        summary = _summarize_tool_result("web_extract", args, self.CONTENT)
+        assert summary == "[web_extract] ? (+1 more) (500 chars)"
+
+    def test_plain_string_urls_unchanged(self):
+        # Regression guard: the normal (already-working) string path is intact.
+        args = json.dumps({"urls": ["https://example.com/a", "https://example.org/b"]})
+        summary = _summarize_tool_result("web_extract", args, self.CONTENT)
+        assert summary == "[web_extract] https://example.com/a (+1 more) (500 chars)"
+
+
 class TestShouldCompress:
     def test_below_threshold(self, compressor):
         compressor.last_prompt_tokens = 50000

tests/agent/transports/test_codex_transport.py

@@ -673,6 +673,34 @@ class TestCodexValidateResponse:
         r = SimpleNamespace(output=None, output_text="Some text")
         assert transport.validate_response(r) is False
 
+    def test_empty_output_content_filter_incomplete_is_valid(self, transport):
+        r = SimpleNamespace(
+            status="incomplete",
+            incomplete_details=SimpleNamespace(reason="content_filter"),
+            output=[],
+            output_text="",
+        )
+        assert transport.validate_response(r) is True
+
+    def test_empty_output_content_filter_dict_incomplete_is_valid(self, transport):
+        r = SimpleNamespace(
+            status=" incomplete ",
+            incomplete_details={"reason": " content_filter "},
+            output=[],
+            output_text="",
+        )
+        assert transport.validate_response(r) is True
+
+    @pytest.mark.parametrize("reason", ["max_output_tokens", "length", "", None])
+    def test_empty_output_other_incomplete_reasons_remain_invalid(self, transport, reason):
+        r = SimpleNamespace(
+            status="incomplete",
+            incomplete_details=SimpleNamespace(reason=reason),
+            output=[],
+            output_text="",
+        )
+        assert transport.validate_response(r) is False
+
 
 class TestCodexMapFinishReason:
 

tests/gateway/test_config.py

@@ -1467,3 +1467,93 @@ class TestMultiplexProfilesEnvOverride:
         for noise in ("", "   ", "maybe", "2"):
             monkeypatch.setenv("GATEWAY_MULTIPLEX_PROFILES", noise)
             assert _env_multiplex_profiles_override() is None, repr(noise)
+
+
+class TestMultiplexProfilesConfig:
+    """Tests for parsing multiplex_profiles (top-level and nested forms)."""
+
+    def test_multiplex_profiles_top_level(self, tmp_path, monkeypatch):
+        """Top-level multiplex_profiles is honored."""
+        hermes_home = tmp_path / ".hermes"
+        hermes_home.mkdir()
+        (hermes_home / "config.yaml").write_text(
+            "multiplex_profiles: true\n",
+            encoding="utf-8",
+        )
+        monkeypatch.setenv("HERMES_HOME", str(hermes_home))
+
+        config = load_gateway_config()
+
+        assert config.multiplex_profiles is True
+
+    def test_multiplex_profiles_nested_under_gateway(self, tmp_path, monkeypatch):
+        """gateway.multiplex_profiles (the form written by `hermes config set
+        gateway.multiplex_profiles true`) must be honored. Regression test for
+        the silent-fallback bug where the loader only forwarded the top-level
+        key, so users who wrote it under gateway: got multiplex_profiles=False
+        with no warning."""
+        hermes_home = tmp_path / ".hermes"
+        hermes_home.mkdir()
+        (hermes_home / "config.yaml").write_text(
+            "gateway:\n  multiplex_profiles: true\n",
+            encoding="utf-8",
+        )
+        monkeypatch.setenv("HERMES_HOME", str(hermes_home))
+
+        config = load_gateway_config()
+
+        assert config.multiplex_profiles is True, (
+            "gateway.multiplex_profiles: true was silently ignored — "
+            "loader only forwarded the top-level form"
+        )
+
+    def test_multiplex_profiles_default_false(self, tmp_path, monkeypatch):
+        """Default is False when neither form is present."""
+        hermes_home = tmp_path / ".hermes"
+        hermes_home.mkdir()
+        (hermes_home / "config.yaml").write_text("", encoding="utf-8")
+        monkeypatch.setenv("HERMES_HOME", str(hermes_home))
+
+        config = load_gateway_config()
+
+        assert config.multiplex_profiles is False
+
+    def test_multiplex_profiles_top_level_overrides_nested(self, tmp_path, monkeypatch):
+        """When both forms are present, top-level wins (matches profile_routes
+        and other parity bridges in load_gateway_config)."""
+        hermes_home = tmp_path / ".hermes"
+        hermes_home.mkdir()
+        (hermes_home / "config.yaml").write_text(
+            "multiplex_profiles: true\n"
+            "gateway:\n  multiplex_profiles: false\n",
+            encoding="utf-8",
+        )
+        monkeypatch.setenv("HERMES_HOME", str(hermes_home))
+
+        config = load_gateway_config()
+
+        assert config.multiplex_profiles is True
+
+    def test_multiplex_profiles_explicit_top_level_false_not_consulting_nested(
+        self, tmp_path, monkeypatch
+    ):
+        """Lock in the `is None` vs `is False` distinction: when top-level is
+        explicitly false, the loader must forward False WITHOUT consulting the
+        nested form (so a stale `gateway.multiplex_profiles: true` cannot
+        silently re-enable multiplexing). Guards against a future regression
+        that flips the check to `not _mp`."""
+        hermes_home = tmp_path / ".hermes"
+        hermes_home.mkdir()
+        (hermes_home / "config.yaml").write_text(
+            "multiplex_profiles: false\n"
+            "gateway:\n  multiplex_profiles: true\n",
+            encoding="utf-8",
+        )
+        monkeypatch.setenv("HERMES_HOME", str(hermes_home))
+
+        config = load_gateway_config()
+
+        assert config.multiplex_profiles is False, (
+            "Explicit top-level false was overridden by nested true — "
+            "loader must respect top-level precedence when key is present"
+        )

tests/gateway/test_incomplete_gateway_turns.py

@@ -0,0 +1,190 @@
+"""Regression tests for hidden-reasoning-only incomplete gateway turns."""
+
+import asyncio
+from datetime import datetime
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+import gateway.run as gateway_run
+from gateway.config import GatewayConfig, Platform, PlatformConfig
+from gateway.platforms.base import BasePlatformAdapter, MessageEvent, ProcessingOutcome, SendResult
+from gateway.session import SessionEntry, SessionSource, build_session_key
+
+
+class CaptureSlackAdapter(BasePlatformAdapter):
+    def __init__(self):
+        super().__init__(PlatformConfig(enabled=True, token="fake-token"), Platform.SLACK)
+        self.sent = []
+        self.processing_hooks = []
+
+    async def connect(self) -> bool:
+        return True
+
+    async def disconnect(self) -> None:
+        return None
+
+    async def send(self, chat_id, content, reply_to=None, metadata=None) -> SendResult:
+        self.sent.append(
+            {
+                "chat_id": chat_id,
+                "content": content,
+                "reply_to": reply_to,
+                "metadata": metadata,
+            }
+        )
+        return SendResult(success=True, message_id="slack-1")
+
+    async def send_typing(self, chat_id: str, metadata=None) -> None:
+        return None
+
+    async def get_chat_info(self, chat_id: str):
+        return {"id": chat_id}
+
+    async def on_processing_start(self, event: MessageEvent) -> None:
+        self.processing_hooks.append(("start", event.message_id))
+
+    async def on_processing_complete(self, event: MessageEvent, outcome: ProcessingOutcome) -> None:
+        self.processing_hooks.append(("complete", event.message_id, outcome))
+
+
+def _make_incomplete_result() -> dict:
+    # Mirror the REAL conversation-loop exhaustion shape: the sentinel text is
+    # returned as BOTH final_response and error (agent/conversation_loop.py's
+    # "remained incomplete after 3 continuation attempts" return).
+    _sentinel = "Codex response remained incomplete after 3 continuation attempts"
+    return {
+        "final_response": _sentinel,
+        "messages": [
+            {"role": "user", "content": "hello"},
+            {"role": "assistant", "content": ""},
+        ],
+        "tools": [],
+        "history_offset": 0,
+        "api_calls": 3,
+        "partial": True,
+        "completed": False,
+        "interrupted": False,
+        "error": _sentinel,
+        "last_prompt_tokens": 0,
+    }
+
+
+def _make_runner(adapter: CaptureSlackAdapter) -> gateway_run.GatewayRunner:
+    runner = object.__new__(gateway_run.GatewayRunner)
+    runner.config = GatewayConfig(
+        platforms={Platform.SLACK: PlatformConfig(enabled=True, token="fake-token")}
+    )
+    runner.adapters = {Platform.SLACK: adapter}
+    runner._voice_mode = {}
+    runner.hooks = SimpleNamespace(emit=AsyncMock(), loaded_hooks=False)
+    runner.session_store = MagicMock()
+    runner.session_store.get_or_create_session.return_value = SessionEntry(
+        session_key="agent:main:slack:channel:C123:171717",
+        session_id="sess-1",
+        created_at=datetime.now(),
+        updated_at=datetime.now(),
+        platform=Platform.SLACK,
+        chat_type="channel",
+    )
+    runner.session_store.load_transcript.return_value = []
+    runner.session_store.has_any_sessions.return_value = True
+    runner.session_store.rewrite_transcript = MagicMock()
+    runner.session_store.append_to_transcript = MagicMock()
+    runner.session_store.update_session = MagicMock()
+    # The transient-failure persistence path dedupes on platform message_id
+    # (#47237). A bare MagicMock returns a truthy mock, which would wrongly
+    # mark the user turn as a duplicate and skip persisting it.
+    runner.session_store.has_platform_message_id = MagicMock(return_value=False)
+    runner._running_agents = {}
+    runner._pending_messages = {}
+    runner._pending_approvals = {}
+    runner._session_db = None
+    runner._is_user_authorized = lambda _source: True
+    runner._set_session_env = lambda _context: None
+    runner._run_agent = AsyncMock(return_value=_make_incomplete_result())
+    return runner
+
+
+def _make_event() -> MessageEvent:
+    return MessageEvent(
+        text="hello",
+        source=SessionSource(
+            platform=Platform.SLACK,
+            chat_id="C123",
+            chat_type="channel",
+            thread_id="171717",
+            user_id="U123",
+        ),
+        message_id="m-1",
+    )
+
+
+def test_incomplete_codex_warning_is_not_surfaced_as_chat_text():
+    agent_result = _make_incomplete_result()
+
+    # Mirror the gateway pipeline: the hidden-turn detector blanks the
+    # sentinel final_response BEFORE empty-response normalization runs.
+    response = agent_result.get("final_response") or ""
+    assert gateway_run._is_gateway_hidden_reasoning_incomplete_turn(agent_result)
+    response = ""
+
+    response = gateway_run._normalize_empty_agent_response(
+        agent_result,
+        response,
+        history_len=4,
+    )
+
+    assert response == ""
+
+
+def test_real_answer_alongside_incomplete_error_is_never_suppressed():
+    """A turn whose final_response is genuine model text (not the sentinel
+    echo) must be delivered even when the error field carries the
+    retry-exhaustion sentinel — suppression is only for hidden turns."""
+    agent_result = _make_incomplete_result()
+    agent_result["final_response"] = "Here is the actual answer."
+
+    assert not gateway_run._is_gateway_hidden_reasoning_incomplete_turn(agent_result)
+
+
+def test_interrupted_or_failed_turns_are_not_classified_hidden():
+    for key in ("interrupted", "failed"):
+        agent_result = _make_incomplete_result()
+        agent_result[key] = True
+        assert not gateway_run._is_gateway_hidden_reasoning_incomplete_turn(agent_result)
+
+
+@pytest.mark.asyncio
+async def test_incomplete_codex_turn_stays_out_of_slack_transcript(monkeypatch, tmp_path):
+    adapter = CaptureSlackAdapter()
+    runner = _make_runner(adapter)
+
+    monkeypatch.setattr(gateway_run, "_hermes_home", tmp_path)
+    monkeypatch.setattr(gateway_run, "_resolve_runtime_agent_kwargs", lambda: {"api_key": "fake"})
+    monkeypatch.setattr(
+        "agent.model_metadata.get_model_context_length",
+        lambda *_args, **_kwargs: 100,
+    )
+    monkeypatch.setenv("SLACK_HOME_CHANNEL", "C123")
+
+    adapter.set_message_handler(runner._handle_message)
+    adapter._keep_typing = lambda *_args, **_kwargs: asyncio.Event().wait()
+
+    event = _make_event()
+    await adapter._process_message_background(event, build_session_key(event.source))
+
+    assert adapter.sent == []
+    assert runner.session_store.update_session.called
+
+    transcript_roles = [
+        call.args[1]["role"]
+        for call in runner.session_store.append_to_transcript.call_args_list
+    ]
+    assert transcript_roles == ["session_meta", "user"]
+    assert runner.session_store.append_to_transcript.call_args_list[1].args[1]["content"] == "hello"
+    assert adapter.processing_hooks == [
+        ("start", "m-1"),
+        ("complete", "m-1", ProcessingOutcome.SUCCESS),
+    ]

tests/gateway/test_multiplex_adapter_registry.py

@@ -45,6 +45,58 @@ class TestCredentialFingerprint:
         assert "config-token" not in fp
 
 
+    def test_reads_config_token(self):
+        """Adapters like Discord store token on `config`, not on self.
+
+        Without the config-token fallback, every Discord adapter in a
+        multiplexed gateway returns None here and the same-token conflict
+        check is silently skipped — N adapters start polling the same bot
+        token and race on every inbound message.
+        """
+        class _Config:
+            token = "discord-bot-token"
+        class _ConfigBackedAdapter:
+            config = _Config()
+        fp = GatewayRunner._adapter_credential_fingerprint(_ConfigBackedAdapter())
+        assert fp is not None
+        assert "discord-bot-token" not in fp
+        assert len(fp) == 16
+
+    def test_distinct_config_tokens_distinct_fp(self):
+        class _CfgA:
+            token = "tok-A"
+        class _CfgB:
+            token = "tok-B"
+        class _A:
+            config = _CfgA()
+        class _B:
+            config = _CfgB()
+        a = GatewayRunner._adapter_credential_fingerprint(_A())
+        b = GatewayRunner._adapter_credential_fingerprint(_B())
+        assert a is not None and b is not None
+        assert a != b
+
+    def test_direct_token_takes_precedence_over_config(self):
+        """If both `adapter.token` and `adapter.config.token` exist, direct wins."""
+        class _Cfg:
+            token = "from-config"
+        class _Both:
+            token = "from-direct"
+            config = _Cfg()
+        fp = GatewayRunner._adapter_credential_fingerprint(_Both())
+        import hashlib
+        expected = hashlib.sha256(b"hermes-mux:from-direct").hexdigest()[:16]
+        assert fp == expected
+
+    def test_config_without_token_returns_none(self):
+        """config present but no token attribute → None (no false positive)."""
+        class _Cfg:
+            pass
+        class _Adapter:
+            config = _Cfg()
+        assert GatewayRunner._adapter_credential_fingerprint(_Adapter()) is None
+
+
 class TestProfileMessageHandler:
     @pytest.mark.asyncio
     async def test_stamps_profile_on_unstamped_source(self):

tests/gateway/test_profile_resolution.py

@@ -0,0 +1,471 @@
+"""Tests for GatewayRunner._resolve_profile_home_for_source — profile resolution logic."""
+
+import logging
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from gateway.session import SessionSource, build_session_key
+from gateway.run import GatewayRunner
+from gateway.profile_routing import ProfileRoute
+from gateway.config import Platform
+from gateway.platforms.base import BasePlatformAdapter
+
+
+@pytest.fixture
+def mock_runner():
+    """Create a minimal mock GatewayRunner with the methods we need."""
+    runner = MagicMock(spec=GatewayRunner)
+    runner.config = MagicMock(profile_routes=[])
+    # Bind the actual methods to the mock
+    runner._profile_name_for_source = GatewayRunner._profile_name_for_source.__get__(runner)
+    runner._resolve_profile_home_for_source = GatewayRunner._resolve_profile_home_for_source.__get__(runner)
+    return runner
+
+
+@pytest.fixture
+def discord_source():
+    """Create a basic Discord SessionSource for testing."""
+    return SessionSource(
+        platform=MagicMock(value="discord"),
+        chat_id="123456",
+        guild_id="789",
+        thread_id=None,
+        parent_chat_id=None,
+    )
+
+
+@pytest.fixture
+def telegram_source():
+    """Create a basic Telegram SessionSource for testing.
+
+    Telegram (like Slack/Feishu/etc.) has no ``guild_id`` — only ``chat_id``.
+    Used to prove profile routing is platform-generic, not Discord-only.
+    """
+    return SessionSource(
+        platform=MagicMock(value="telegram"),
+        chat_id="-1001234567890",
+        guild_id=None,
+        thread_id=None,
+        parent_chat_id=None,
+    )
+
+
+class TestResolutionOrder:
+    """Tests that profile resolution follows the correct priority order."""
+    
+    def test_source_profile_wins_over_routing(self, mock_runner, discord_source):
+        """source.profile should be used even if routing would match."""
+        discord_source.profile = "from-source"
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value="active"):
+            with patch("hermes_cli.profiles.get_profile_dir") as mock_get_dir:
+                with patch("hermes_cli.profiles.profile_exists", return_value=True):
+                    mock_get_dir.return_value = Path("/hermes/profiles/from-source")
+                    result = mock_runner._resolve_profile_home_for_source(discord_source)
+                    
+                    assert result == Path("/hermes/profiles/from-source")
+                    mock_get_dir.assert_called_once_with("from-source")
+    
+    def test_routing_wins_over_active_profile(self, mock_runner, discord_source):
+        """When source.profile is empty, routing should win over active profile."""
+        discord_source.profile = None
+        
+        # Mock routing to return a profile
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value="active"):
+            with patch("hermes_cli.profiles.get_profile_dir") as mock_get_dir:
+                with patch("hermes_cli.profiles.profile_exists", return_value=True):
+                    mock_get_dir.return_value = Path("/hermes/profiles/routed")
+                    
+                    # Manually set routing to return a profile
+                    mock_runner._profile_name_for_source = MagicMock(return_value="routed")
+                    
+                    result = mock_runner._resolve_profile_home_for_source(discord_source)
+                    
+                    assert result == Path("/hermes/profiles/routed")
+                    mock_get_dir.assert_called_once_with("routed")
+    
+    def test_active_profile_fallback(self, mock_runner, discord_source):
+        """When source.profile and routing both return None, active profile is used."""
+        discord_source.profile = None
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value="active"):
+            with patch("hermes_cli.profiles.get_profile_dir") as mock_get_dir:
+                mock_get_dir.return_value = Path("/hermes/profiles/active")
+                
+                # No routing match
+                mock_runner._profile_name_for_source = MagicMock(return_value=None)
+                
+                result = mock_runner._resolve_profile_home_for_source(discord_source)
+                
+                assert result == Path("/hermes/profiles/active")
+                mock_get_dir.assert_called_once_with("active")
+    
+    def test_default_fallback_when_no_active(self, mock_runner, discord_source):
+        """When even active profile is None, 'default' is used."""
+        discord_source.profile = None
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value=None):
+            with patch("hermes_cli.profiles.get_profile_dir") as mock_get_dir:
+                mock_get_dir.return_value = Path("/hermes")
+                
+                mock_runner._profile_name_for_source = MagicMock(return_value=None)
+                
+                result = mock_runner._resolve_profile_home_for_source(discord_source)
+                
+                assert result == Path("/hermes")
+                mock_get_dir.assert_called_once_with("default")
+
+
+class TestMissingProfileWarning:
+    """Tests for warning when a profile doesn't exist on disk."""
+    
+    def test_nonexistent_profile_warning(self, mock_runner, discord_source, caplog):
+        """When source.profile points to a nonexistent profile, log a WARNING."""
+        discord_source.profile = "nonexistent"
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value="active"):
+            with patch("hermes_cli.profiles.get_profile_dir") as mock_get_dir:
+                mock_get_dir.return_value = Path("/hermes/profiles/nonexistent")
+                with patch("hermes_cli.profiles.profile_exists", return_value=False):
+                    with patch("hermes_constants.get_hermes_home", return_value=Path("/hermes")):
+                        with caplog.at_level(logging.WARNING):
+                            result = mock_runner._resolve_profile_home_for_source(discord_source)
+                            
+                            # Should fall back to global HERMES_HOME
+                            assert result == Path("/hermes")
+                            
+                            # Should have logged a warning
+                            assert len(caplog.records) == 1
+                            assert caplog.records[0].levelname == "WARNING"
+                            assert "nonexistent" in caplog.records[0].message
+                            assert "does not exist" in caplog.records[0].message
+                            assert "discord" in caplog.records[0].message
+                            assert "123456" in caplog.records[0].message
+    
+    def test_nonexistent_routing_profile_warning(self, mock_runner, discord_source, caplog):
+        """When routing returns a nonexistent profile, log a WARNING."""
+        discord_source.profile = None
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value="active"):
+            with patch("hermes_cli.profiles.get_profile_dir") as mock_get_dir:
+                mock_get_dir.return_value = Path("/hermes/profiles/routed")
+                with patch("hermes_cli.profiles.profile_exists", return_value=False):
+                    with patch("hermes_constants.get_hermes_home", return_value=Path("/hermes")):
+                        # Routing returns a profile that doesn't exist
+                        mock_runner._profile_name_for_source = MagicMock(return_value="routed")
+                        
+                        with caplog.at_level(logging.WARNING):
+                            result = mock_runner._resolve_profile_home_for_source(discord_source)
+                            
+                            # Should fall back to global HERMES_HOME
+                            assert result == Path("/hermes")
+                            
+                            # Should have logged a warning
+                            assert len(caplog.records) == 1
+                            assert "routed" in caplog.records[0].message
+    
+    def test_empty_source_profile_no_warning(self, mock_runner, discord_source, caplog):
+        """When source.profile is empty, silent fallback to active profile (no warning)."""
+        discord_source.profile = None
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value="active"):
+            with patch("hermes_cli.profiles.get_profile_dir") as mock_get_dir:
+                mock_get_dir.return_value = Path("/hermes/profiles/active")
+                with patch("hermes_cli.profiles.profile_exists", return_value=True):
+                    with caplog.at_level(logging.WARNING):
+                        mock_runner._profile_name_for_source = MagicMock(return_value=None)
+                        
+                        result = mock_runner._resolve_profile_home_for_source(discord_source)
+                        
+                        # Should use active profile
+                        assert result == Path("/hermes/profiles/active")
+                        
+                        # No warnings (active profile exists)
+                        assert not any(r.levelname == "WARNING" for r in caplog.records)
+    
+    def test_existing_profile_no_warning(self, mock_runner, discord_source, caplog):
+        """When the profile exists, no warning should be logged."""
+        discord_source.profile = "existing"
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value="active"):
+            with patch("hermes_cli.profiles.get_profile_dir") as mock_get_dir:
+                mock_get_dir.return_value = Path("/hermes/profiles/existing")
+                with patch("hermes_cli.profiles.profile_exists", return_value=True):
+                    with caplog.at_level(logging.WARNING):
+                        result = mock_runner._resolve_profile_home_for_source(discord_source)
+                        
+                        assert result == Path("/hermes/profiles/existing")
+                        
+                        # No warnings
+                        assert not any(r.levelname == "WARNING" for r in caplog.records)
+
+
+class TestExceptionHandling:
+    """Tests for exception handling in profile resolution."""
+    
+    def test_get_profile_dir_exception_logs_warning(self, mock_runner, discord_source, caplog):
+        """When get_profile_dir raises an exception, log a WARNING with context."""
+        discord_source.profile = "bad-profile"
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value="active"):
+            with patch("hermes_cli.profiles.get_profile_dir", side_effect=ValueError("Invalid profile name")):
+                with patch("hermes_constants.get_hermes_home", return_value=Path("/hermes")):
+                    with caplog.at_level(logging.WARNING):
+                        result = mock_runner._resolve_profile_home_for_source(discord_source)
+                        
+                        # Should fall back to global HERMES_HOME
+                        assert result == Path("/hermes")
+                        
+                        # Should have logged a warning with exception info
+                        assert len(caplog.records) == 1
+                        assert caplog.records[0].levelname == "WARNING"
+                        assert "bad-profile" in caplog.records[0].message
+                        assert "Failed to resolve profile directory" in caplog.records[0].message
+    
+    def test_exception_with_no_profile_name(self, mock_runner, discord_source, caplog):
+        """Exception when no profile was set should still log a warning."""
+        discord_source.profile = None
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value=None):
+            with patch("hermes_cli.profiles.get_profile_dir", side_effect=RuntimeError("Filesystem error")):
+                with patch("hermes_constants.get_hermes_home", return_value=Path("/hermes")):
+                    mock_runner._profile_name_for_source = MagicMock(return_value=None)
+                    
+                    with caplog.at_level(logging.WARNING):
+                        result = mock_runner._resolve_profile_home_for_source(discord_source)
+                        
+                        assert result == Path("/hermes")
+                        
+                        # Warning should mention "(no profile)"
+                        assert "(no profile)" in caplog.records[0].message
+
+
+class TestRoutingConsultation:
+    """Tests that _profile_name_for_source is consulted when source.profile is empty."""
+    
+    def test_routing_consulted_when_source_profile_empty(self, mock_runner, discord_source):
+        """_profile_name_for_source should be called when source.profile is empty."""
+        discord_source.profile = None
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value="active"):
+            with patch("hermes_cli.profiles.get_profile_dir") as mock_get_dir:
+                mock_get_dir.return_value = Path("/hermes/profiles/routed")
+                
+                mock_runner._profile_name_for_source = MagicMock(return_value="routed")
+                
+                mock_runner._resolve_profile_home_for_source(discord_source)
+                
+                # Should have called routing
+                mock_runner._profile_name_for_source.assert_called_once_with(discord_source)
+    
+    def test_routing_not_consulted_when_source_profile_set(self, mock_runner, discord_source):
+        """_profile_name_for_source should NOT be called when source.profile is set."""
+        discord_source.profile = "from-source"
+        
+        with patch("hermes_cli.profiles.get_active_profile_name", return_value="active"):
+            with patch("hermes_cli.profiles.get_profile_dir") as mock_get_dir:
+                mock_get_dir.return_value = Path("/hermes/profiles/from-source")
+                
+                mock_runner._profile_name_for_source = MagicMock(return_value="routed")
+                
+                mock_runner._resolve_profile_home_for_source(discord_source)
+                
+                # Should NOT have called routing
+                mock_runner._profile_name_for_source.assert_not_called()
+
+
+class TestNonDiscordProfileRouting:
+    """Profile routing must be platform-generic, not Discord-only.
+
+    Regression coverage for the ``gateway_runner`` injection gap: previously
+    only Discord's adapter pre-declared ``gateway_runner``, so only Discord
+    ever had ``build_source`` call ``_profile_name_for_source``. Telegram /
+    Feishu / Slack / etc. silently fell through to the default profile. These
+    tests pin the resolution half for a non-Discord platform (Telegram).
+    """
+
+    def test_telegram_route_resolves(self, mock_runner, telegram_source):
+        """A configured Telegram route resolves to its profile via the real
+        ``_profile_name_for_source`` (bound onto the mock runner)."""
+        mock_runner.config.profile_routes = [
+            ProfileRoute(name="tg", platform="telegram", profile="tg-profile",
+                         chat_id="-1001234567890"),
+        ]
+        telegram_source.profile = None
+
+        assert mock_runner._profile_name_for_source(telegram_source) == "tg-profile"
+
+    def test_telegram_no_route_returns_none(self, mock_runner, telegram_source):
+        """With no matching Telegram route, resolution returns None (caller
+        falls back to the default/active profile)."""
+        mock_runner.config.profile_routes = [
+            ProfileRoute(name="dc", platform="discord", profile="dc-profile",
+                         chat_id="123456"),
+        ]
+        telegram_source.profile = None
+
+        assert mock_runner._profile_name_for_source(telegram_source) is None
+
+
+class TestGatewayRunnerInjection:
+    """``BasePlatformAdapter`` declares ``gateway_runner`` so the gateway's
+    unconditional injection reaches every platform adapter — the foundation
+    that makes the routing in TestNonDiscordProfileRouting reachable at runtime.
+    """
+
+    def test_base_adapter_declares_gateway_runner(self):
+        from gateway.platforms.base import BasePlatformAdapter
+
+        # Class-level attribute exists and defaults to None.
+        assert hasattr(BasePlatformAdapter, "gateway_runner")
+        assert BasePlatformAdapter.gateway_runner is None
+
+    def test_subclass_inherits_gateway_runner(self):
+        from gateway.platforms.base import BasePlatformAdapter
+
+        class _ToyAdapter(BasePlatformAdapter):
+            pass
+
+        # No manual declaration — yet the attribute is inherited from the base,
+        # so the gateway's ``adapter.gateway_runner = self`` injection reaches
+        # every adapter, not just the ones that pre-declared it (Discord).
+        assert hasattr(_ToyAdapter, "gateway_runner")
+        assert _ToyAdapter.gateway_runner is None
+
+
+# A concrete adapter we can instantiate without the full platform stack.
+# ``build_source`` only reads ``self.platform`` and ``self.gateway_runner``, so a
+# bare instance with those two attrs exercises the real BasePlatformAdapter
+# method end-to-end. Clearing ``__abstractmethods__`` lets ``__new__`` bypass
+# the ABC instantiation guard without stubbing connect/send/get_chat_info/…
+class _StubAdapter(BasePlatformAdapter):
+    pass
+
+
+_StubAdapter.__abstractmethods__ = frozenset()  # type: ignore[attr-defined]
+
+
+def _stub_adapter(platform: Platform, runner) -> "_StubAdapter":
+    a = _StubAdapter.__new__(_StubAdapter)
+    a.platform = platform
+    a.gateway_runner = runner
+    return a
+
+
+class TestAdapterToSessionKeyIntegration:
+    """Adapter -> ``source.profile`` -> session-key integration coverage.
+
+    The review asked for integration coverage for Discord AND a non-Discord
+    platform. These drive a concrete adapter's real ``build_source``
+    (BasePlatformAdapter) with an injected ``gateway_runner``, assert the
+    matched route's profile is stamped on the source, and that the resulting
+    session key is profile-scoped (``agent:<profile>:...`` rather than the
+    shared ``agent:main:...``). The Telegram case is the bug-#2 regression:
+    pre-fix it never received ``gateway_runner`` and fell through to default.
+    """
+
+    @staticmethod
+    def _routes():
+        return [
+            ProfileRoute(name="dc", platform="discord", profile="coder",
+                         guild_id="111", chat_id="222"),
+            ProfileRoute(name="tg", platform="telegram", profile="ops",
+                         chat_id="-1001234567890"),
+        ]
+
+    def test_discord_adapter_stamps_profile_and_scopes_key(self, mock_runner):
+        mock_runner.config.profile_routes = self._routes()
+        adapter = _stub_adapter(Platform.DISCORD, mock_runner)
+
+        source = adapter.build_source(
+            chat_id="222", chat_type="group", guild_id="111", user_id="u1",
+        )
+        assert source.profile == "coder"
+
+        key = build_session_key(source, profile=source.profile)
+        assert key.startswith("agent:coder:"), key
+        # A default-profile key would land in agent:main — must differ.
+        assert key != build_session_key(source, profile=None)
+
+    def test_telegram_adapter_stamps_profile_and_scopes_key(self, mock_runner):
+        """Non-Discord platform (bug #2). The adapter now receives
+        ``gateway_runner``, so ``build_source`` stamps the profile and the
+        session key is isolated under ``agent:ops:`` instead of ``agent:main:``."""
+        mock_runner.config.profile_routes = self._routes()
+        adapter = _stub_adapter(Platform.TELEGRAM, mock_runner)
+
+        source = adapter.build_source(
+            chat_id="-1001234567890", chat_type="group", user_id="u1",
+        )
+        assert source.profile == "ops"
+
+        key = build_session_key(source, profile=source.profile)
+        assert key.startswith("agent:ops:"), key
+        assert key != build_session_key(source, profile=None)
+
+    def test_adapter_without_runner_falls_back_to_default_namespace(self, mock_runner):
+        """Regression anchor: with no ``gateway_runner`` injected (the pre-fix
+        state for non-Discord adapters), ``build_source`` leaves ``profile=None``
+        and the session key is the shared ``agent:main:`` namespace — no
+        per-profile isolation. This is the silent fallback the fix removes for
+        non-Discord platforms."""
+        adapter = _stub_adapter(Platform.TELEGRAM, runner=None)
+
+        source = adapter.build_source(
+            chat_id="-1001234567890", chat_type="group", user_id="u1",
+        )
+        assert source.profile is None
+        key = build_session_key(source, profile=source.profile)
+        assert key.startswith("agent:main:"), key
+
+
+class TestMultiplexGate:
+    """``profile_routes`` only activates under ``gateway.multiplex_profiles``.
+
+    Routing stamps ``source.profile``, which namespaces session/batch keys —
+    but the profile-scoped agent run (``_profile_runtime_scope``) only engages
+    when multiplexing is on. Without the gate, a configured route with
+    multiplexing off would split batch/session keys into ``agent:<profile>``
+    while the agent still served the turn from ``agent:main``'s home.
+    """
+
+    def test_routes_ignored_when_multiplex_off(self, mock_runner, discord_source):
+        mock_runner.config.multiplex_profiles = False
+        mock_runner.config.profile_routes = [
+            ProfileRoute(name="dc", platform="discord", profile="coder",
+                         guild_id="789", chat_id="123456"),
+        ]
+        discord_source.profile = None
+
+        assert mock_runner._profile_name_for_source(discord_source) is None
+
+    def test_routes_active_when_multiplex_on(self, mock_runner, discord_source):
+        mock_runner.config.multiplex_profiles = True
+        mock_runner.config.profile_routes = [
+            ProfileRoute(name="dc", platform="discord", profile="coder",
+                         guild_id="789", chat_id="123456"),
+        ]
+        discord_source.profile = None
+
+        assert mock_runner._profile_name_for_source(discord_source) == "coder"
+
+    def test_build_source_leaves_profile_none_when_multiplex_off(self, mock_runner):
+        """End-to-end through the real adapter ``build_source``: with routes
+        configured but multiplexing off, no profile is stamped and the session
+        key stays in the legacy ``agent:main`` namespace — byte-identical to a
+        gateway with no routes at all."""
+        mock_runner.config.multiplex_profiles = False
+        mock_runner.config.profile_routes = [
+            ProfileRoute(name="dc", platform="discord", profile="coder",
+                         guild_id="111", chat_id="222"),
+        ]
+        adapter = _stub_adapter(Platform.DISCORD, mock_runner)
+
+        source = adapter.build_source(
+            chat_id="222", chat_type="group", guild_id="111", user_id="u1",
+        )
+        assert source.profile is None
+        key = build_session_key(source, profile=source.profile)
+        assert key.startswith("agent:main:"), key

tests/gateway/test_profile_routing.py

@@ -0,0 +1,261 @@
+"""Tests for gateway/profile_routing.py — profile-based routing."""
+
+import pytest
+from gateway.profile_routing import (
+    ProfileRoute,
+    parse_profile_routes,
+    match_profile_route,
+)
+
+
+class TestProfileRoute:
+    def test_specificity_thread(self):
+        r = ProfileRoute(name="t", platform="discord", profile="p",
+                         guild_id="g", chat_id="c", thread_id="t")
+        assert r.specificity == 14  # 2 + 4 + 8
+
+    def test_specificity_channel(self):
+        r = ProfileRoute(name="c", platform="discord", profile="p",
+                         guild_id="g", chat_id="c")
+        assert r.specificity == 6  # 2 + 4
+
+    def test_specificity_guild(self):
+        r = ProfileRoute(name="g", platform="discord", profile="p",
+                         guild_id="g")
+        assert r.specificity == 2
+
+    def test_specificity_minimal(self):
+        r = ProfileRoute(name="m", platform="telegram", profile="p")
+        assert r.specificity == 0
+
+    def test_frozen(self):
+        r = ProfileRoute(name="x", platform="discord", profile="p")
+        with pytest.raises(AttributeError):
+            r.name = "y"
+
+
+class TestProfileRouteMatching:
+    def test_exact_thread_match(self):
+        r = ProfileRoute(name="t", platform="discord", profile="trader",
+                         guild_id="111", chat_id="222", thread_id="333")
+        assert r.matches("discord", guild_id="111", chat_id="222", thread_id="333")
+        assert not r.matches("discord", guild_id="111", chat_id="222", thread_id="444")
+
+    def test_channel_match(self):
+        r = ProfileRoute(name="c", platform="discord", profile="helper",
+                         chat_id="222")
+        assert r.matches("discord", chat_id="222")
+        assert not r.matches("discord", chat_id="333")
+        assert not r.matches("telegram", chat_id="222")
+
+    def test_guild_match(self):
+        r = ProfileRoute(name="g", platform="discord", profile="server",
+                         guild_id="111")
+        assert r.matches("discord", guild_id="111")
+        assert not r.matches("discord", guild_id="222")
+
+    def test_disabled_route_no_match(self):
+        r = ProfileRoute(name="d", platform="discord", profile="off",
+                         guild_id="111", enabled=False)
+        assert not r.matches("discord", guild_id="111")
+
+    def test_guild_route_matches_any_channel_in_guild(self):
+        r = ProfileRoute(name="g", platform="discord", profile="server",
+                         guild_id="111")
+        assert r.matches("discord", guild_id="111", chat_id="222")
+        assert r.matches("discord", guild_id="111", chat_id="222", thread_id="333")
+
+    def test_extra_fields_ignored(self):
+        r = ProfileRoute(name="g", platform="discord", profile="server",
+                         guild_id="111")
+        assert r.matches("discord", guild_id="111", chat_id="any")
+
+    def test_guild_and_chat_are_conjunctive(self):
+        # A route declaring BOTH guild_id and chat_id requires both to match.
+        # Regression guard: previously chat_id was checked first and returned
+        # True before guild_id was ever consulted.
+        r = ProfileRoute(name="gc", platform="discord", profile="scoped",
+                         guild_id="111", chat_id="222")
+        # Both match (direct channel) -> match
+        assert r.matches("discord", guild_id="111", chat_id="222")
+        # Both match via parent (thread inside the channel) -> match
+        assert r.matches("discord", guild_id="111", chat_id="333", parent_chat_id="222")
+        # chat matches but guild differs -> NO match (the bug this guards)
+        assert not r.matches("discord", guild_id="999", chat_id="222")
+        # guild matches but chat differs -> NO match
+        assert not r.matches("discord", guild_id="111", chat_id="333")
+
+
+class TestParseProfileRoutes:
+    def test_empty(self):
+        assert parse_profile_routes(None) == []
+        assert parse_profile_routes([]) == []
+
+    def test_valid_routes_sorted_by_specificity(self):
+        raw = [
+            {"name": "guild", "platform": "discord", "profile": "p", "guild_id": "1"},
+            {"name": "thread", "platform": "discord", "profile": "p",
+             "guild_id": "1", "chat_id": "2", "thread_id": "3"},
+            {"name": "channel", "platform": "discord", "profile": "p", "chat_id": "2"},
+        ]
+        routes = parse_profile_routes(raw)
+        names = [r.name for r in routes]
+        assert names == ["thread", "channel", "guild"]
+
+    def test_skips_invalid(self):
+        raw = [
+            {"platform": "discord"},
+            {"profile": "p"},
+            "not a dict",
+            {"name": "ok", "platform": "telegram", "profile": "p"},
+        ]
+        routes = parse_profile_routes(raw)
+        assert len(routes) == 1
+        assert routes[0].name == "ok"
+
+    def test_enabled_flag(self):
+        raw = [
+            {"name": "off", "platform": "discord", "profile": "p",
+             "guild_id": "1", "enabled": False},
+            {"name": "on", "platform": "discord", "profile": "p", "guild_id": "1"},
+        ]
+        routes = parse_profile_routes(raw)
+        assert not routes[0].enabled
+        assert routes[1].enabled
+
+
+class TestMatchProfileRoute:
+    def test_no_routes(self):
+        assert match_profile_route([], "discord") is None
+
+    def test_returns_first_match(self):
+        routes = [
+            ProfileRoute(name="thread", platform="discord", profile="trader",
+                         guild_id="1", chat_id="2", thread_id="3"),
+            ProfileRoute(name="channel", platform="discord", profile="helper",
+                         chat_id="2"),
+        ]
+        m = match_profile_route(routes, "discord", guild_id="1", chat_id="2", thread_id="3")
+        assert m is not None
+        assert m.profile == "trader"
+
+    def test_falls_through_to_channel(self):
+        routes = [
+            ProfileRoute(name="thread", platform="discord", profile="trader",
+                         guild_id="1", chat_id="2", thread_id="3"),
+            ProfileRoute(name="channel", platform="discord", profile="helper",
+                         chat_id="2"),
+        ]
+        m = match_profile_route(routes, "discord", guild_id="1", chat_id="2")
+        assert m is not None
+        assert m.profile == "helper"
+
+    def test_no_match_returns_none(self):
+        routes = [
+            ProfileRoute(name="r", platform="telegram", profile="p"),
+        ]
+        assert match_profile_route(routes, "discord") is None
+
+
+class TestSessionKeyIntegration:
+    def test_default_profile_key(self):
+        from gateway.session import build_session_key, SessionSource, Platform
+        src = SessionSource(platform=Platform.DISCORD, chat_id="123",
+                            chat_type="channel", user_id="456")
+        key = build_session_key(src)
+        assert key.startswith("agent:main:")
+
+    def test_custom_profile_key(self):
+        from gateway.session import build_session_key, SessionSource, Platform
+        src = SessionSource(platform=Platform.DISCORD, chat_id="123",
+                            chat_type="channel", user_id="456")
+        key = build_session_key(src, profile="trader")
+        assert key.startswith("agent:trader:")
+        assert key == "agent:trader:discord:channel:123:456"
+
+    def test_isolated_sessions(self):
+        from gateway.session import build_session_key, SessionSource, Platform
+        src = SessionSource(platform=Platform.DISCORD, chat_id="123",
+                            chat_type="channel", user_id="456")
+        key_default = build_session_key(src)
+        key_trader = build_session_key(src, profile="trader")
+        assert key_default != key_trader
+
+    def test_dm_profile_scoped(self):
+        from gateway.session import build_session_key, SessionSource, Platform
+        src = SessionSource(platform=Platform.DISCORD, chat_id="999",
+                            chat_type="dm", user_id="111")
+        key = build_session_key(src, profile="bot2")
+        assert key == "agent:bot2:discord:dm:999"
+
+
+
+class TestParentChatIdMatching:
+    """Thread messages carry thread_id as chat_id; parent_chat_id is the channel."""
+
+    def test_channel_route_matches_via_parent_chat_id(self):
+        r = ProfileRoute(name="ch", platform="discord", profile="trader",
+                         chat_id="222")
+        assert r.matches("discord", chat_id="333", parent_chat_id="222")
+
+    def test_channel_route_no_match_wrong_parent(self):
+        r = ProfileRoute(name="ch", platform="discord", profile="trader",
+                         chat_id="222")
+        assert not r.matches("discord", chat_id="333", parent_chat_id="444")
+
+    def test_match_profile_route_with_parent_chat_id(self):
+        routes = [
+            ProfileRoute(name="ch", platform="discord", profile="trader",
+                         chat_id="222"),
+        ]
+        m = match_profile_route(routes, "discord", chat_id="333", parent_chat_id="222")
+        assert m is not None
+        assert m.profile == "trader"
+
+    def test_thread_id_does_not_match_parent_chat_id(self):
+        """thread_id only matches the actual thread_id, never parent_chat_id.
+        Discord snowflakes are globally unique, so thread_id != channel_id."""
+        r = ProfileRoute(name="th", platform="discord", profile="helper",
+                         thread_id="555")
+        assert r.matches("discord", thread_id="555")
+        assert not r.matches("discord", parent_chat_id="555")
+
+    def test_no_parent_chat_id_still_works(self):
+        r = ProfileRoute(name="ch", platform="discord", profile="trader",
+                         chat_id="222")
+        assert r.matches("discord", chat_id="222")
+
+    def test_guild_route_matches_with_parent_chat_id(self):
+        """Guild routes should match regardless of chat_id or parent_chat_id."""
+        r = ProfileRoute(name="g", platform="discord", profile="server",
+                         guild_id="111")
+        assert r.matches("discord", guild_id="111", chat_id="333", parent_chat_id="444")
+
+
+class TestForumPostMatching:
+    """Test that forum posts match via parent_chat_id (direct parent)."""
+
+    def test_forum_channel_route_matches_forum_post(self):
+        """A route on a forum channel should match comments on posts in that forum.
+        
+        In Discord, forum posts (threads) have parent_chat_id = forum channel ID.
+        No cache is needed — the parent relationship is direct.
+        """
+        r = ProfileRoute(name="forum", platform="discord", profile="forum_profile",
+                         chat_id="forum_channel_123")
+        # A comment on a forum post: chat_id=post_thread_id, parent_chat_id=forum_channel_id
+        assert r.matches("discord", chat_id="post_thread_456", parent_chat_id="forum_channel_123")
+
+    def test_forum_post_comment_matches_channel_not_thread_id(self):
+        """Verify that thread_id matching is distinct from parent_chat_id matching."""
+        routes = [
+            ProfileRoute(name="forum", platform="discord", profile="forum_profile",
+                         chat_id="forum_channel_123"),
+            ProfileRoute(name="post", platform="discord", profile="post_profile",
+                         thread_id="post_thread_456"),
+        ]
+        # A comment on the forum post should match the forum channel route, not the thread route
+        m = match_profile_route(routes, "discord", chat_id="post_thread_456", 
+                                 parent_chat_id="forum_channel_123")
+        assert m is not None
+        assert m.profile == "forum_profile"

tests/hermes_cli/test_bedrock_region_scoped_picker.py

@@ -0,0 +1,82 @@
+"""Regression tests for #28156 — Bedrock picker must be region-scoped.
+
+Geo-prefixed cross-region inference profiles (us.*, eu.*, apac.*, ...) only
+route from endpoints in their own geography. Offering us.* profiles to an
+eu-central-2 picker produces configs AWS rejects regardless of credentials.
+"""
+
+from hermes_cli.model_setup_flows import (
+    BEDROCK_GEO_PREFIXES,
+    bedrock_model_routable_from_region,
+    bedrock_region_geo_prefix,
+)
+
+
+class TestRegionGeoPrefix:
+    def test_known_geographies(self):
+        assert bedrock_region_geo_prefix("us-east-1") == "us."
+        assert bedrock_region_geo_prefix("eu-central-2") == "eu."
+        assert bedrock_region_geo_prefix("ap-southeast-1") == "ap."
+        assert bedrock_region_geo_prefix("ca-central-1") == "ca."
+        assert bedrock_region_geo_prefix("sa-east-1") == "sa."
+        assert bedrock_region_geo_prefix("me-south-1") == "me."
+        assert bedrock_region_geo_prefix("af-south-1") == "af."
+
+    def test_unknown_region_is_empty(self):
+        assert bedrock_region_geo_prefix("") == ""
+        assert bedrock_region_geo_prefix("moon-base-1") == ""
+
+
+class TestRoutableFromRegion:
+    def test_us_profile_not_offered_in_eu(self):
+        assert not bedrock_model_routable_from_region(
+            "us.anthropic.claude-sonnet-4-6", "eu-central-2"
+        )
+
+    def test_eu_profile_offered_in_eu(self):
+        assert bedrock_model_routable_from_region(
+            "eu.anthropic.claude-sonnet-4-6", "eu-central-2"
+        )
+
+    def test_global_profile_offered_everywhere(self):
+        for region in ("eu-central-2", "us-east-1", "ap-southeast-1"):
+            assert bedrock_model_routable_from_region(
+                "global.anthropic.claude-sonnet-4-6", region
+            )
+
+    def test_bare_foundation_id_offered_everywhere(self):
+        assert bedrock_model_routable_from_region(
+            "anthropic.claude-3-sonnet-20240229-v1:0", "eu-central-2"
+        )
+
+    def test_apac_spellings_route_in_ap_regions(self):
+        for prefix in ("ap.", "apac.", "jp."):
+            assert bedrock_model_routable_from_region(
+                f"{prefix}anthropic.claude-sonnet-4-6", "ap-northeast-1"
+            )
+
+    def test_eu_profile_not_offered_in_us(self):
+        assert not bedrock_model_routable_from_region(
+            "eu.anthropic.claude-sonnet-4-6", "us-east-1"
+        )
+
+    def test_unknown_region_hides_nothing(self):
+        assert bedrock_model_routable_from_region(
+            "us.anthropic.claude-sonnet-4-6", ""
+        )
+
+
+class TestGeoPrefixContract:
+    def test_every_geo_prefix_maps_to_a_routable_region_or_is_alias(self):
+        """Invariant: each geo prefix is either reachable from some region's
+        geo mapping or an ap-family alias — no dead entries."""
+        mapped = {
+            bedrock_region_geo_prefix(r)
+            for r in (
+                "us-east-1", "eu-central-2", "ap-southeast-1",
+                "ca-central-1", "sa-east-1", "me-south-1", "af-south-1",
+            )
+        }
+        ap_aliases = {"apac.", "jp."}
+        for prefix in BEDROCK_GEO_PREFIXES:
+            assert prefix in mapped or prefix in ap_aliases

tests/hermes_cli/test_config.py

@@ -1645,6 +1645,114 @@ class TestMigrationWriteInvariant:
         assert loaded["display"]["compact"] == DEFAULT_CONFIG["display"]["compact"]
 
 
+class TestSaveConfigPartialWritePreservation:
+    """Regression for #62723: partial migration writes must not drop unrelated sections."""
+
+    def test_merge_existing_preserves_platforms_on_partial_write(self, tmp_path):
+        body = """_config_version: 30
+model:
+  default: deepseek-v4-pro
+  provider: deepseek
+agent:
+  max_turns: 60
+platforms:
+  feishu:
+    enabled: true
+    extra:
+      app_id: cli_xxx
+      app_secret: xxx
+feishu:
+  require_mention: true
+"""
+        (tmp_path / "config.yaml").write_text(body, encoding="utf-8")
+        with patch.dict(os.environ, {"HERMES_HOME": str(tmp_path)}):
+            save_config(
+                {
+                    "_config_version": 30,
+                    "model": {"default": "deepseek-v4-pro", "provider": "deepseek"},
+                    "agent": {"max_turns": 60, "verify_on_stop": False},
+                },
+                merge_existing=True,
+            )
+            raw = yaml.safe_load((tmp_path / "config.yaml").read_text(encoding="utf-8"))
+
+        assert raw["platforms"]["feishu"]["extra"]["app_id"] == "cli_xxx"
+        assert raw["feishu"]["require_mention"] is True
+        assert raw["agent"]["verify_on_stop"] is False
+
+    def test_partial_write_without_merge_drops_omitted_sections(self, tmp_path):
+        """Full-replacement callers (raw YAML editor) rely on merge_existing=False."""
+        body = """_config_version: 30
+model:
+  default: deepseek-v4-pro
+  provider: deepseek
+platforms:
+  feishu:
+    enabled: true
+"""
+        (tmp_path / "config.yaml").write_text(body, encoding="utf-8")
+        with patch.dict(os.environ, {"HERMES_HOME": str(tmp_path)}):
+            save_config({"model": {"default": "other-model", "provider": "openrouter"}})
+            raw = yaml.safe_load((tmp_path / "config.yaml").read_text(encoding="utf-8"))
+
+        assert raw["model"]["default"] == "other-model"
+        assert "platforms" not in raw
+
+    def test_persist_migration_writes_full_read_raw_config(self, tmp_path):
+        from hermes_cli.config import _persist_migration, read_raw_config
+
+        body = """_config_version: 30
+model:
+  default: deepseek-v4-pro
+  provider: deepseek
+agent:
+  max_turns: 60
+platforms:
+  feishu:
+    enabled: true
+    extra:
+      app_id: cli_xxx
+      app_secret: xxx
+"""
+        (tmp_path / "config.yaml").write_text(body, encoding="utf-8")
+        with patch.dict(os.environ, {"HERMES_HOME": str(tmp_path)}):
+            config = read_raw_config()
+            config.setdefault("agent", {})["verify_on_stop"] = False
+            config["_config_version"] = 32
+            _persist_migration(config)
+            raw = yaml.safe_load((tmp_path / "config.yaml").read_text(encoding="utf-8"))
+
+        assert raw["platforms"]["feishu"]["extra"]["app_id"] == "cli_xxx"
+        assert raw["agent"]["verify_on_stop"] is False
+        assert raw["agent"]["max_turns"] == 60
+        assert raw["_config_version"] == 32
+
+    def test_v30_to_latest_migration_keeps_platforms(self, tmp_path):
+        """End-to-end: reporter's v30 feishu profile survives version bump."""
+        body = """_config_version: 30
+model:
+  default: deepseek-v4-pro
+  provider: deepseek
+agent:
+  max_turns: 60
+platforms:
+  feishu:
+    enabled: true
+    extra:
+      app_id: cli_xxx
+      app_secret: xxx
+feishu:
+  require_mention: true
+"""
+        (tmp_path / "config.yaml").write_text(body, encoding="utf-8")
+        with patch.dict(os.environ, {"HERMES_HOME": str(tmp_path)}):
+            migrate_config(interactive=False, quiet=True)
+            raw = yaml.safe_load((tmp_path / "config.yaml").read_text(encoding="utf-8"))
+
+        assert raw["platforms"]["feishu"]["extra"]["app_id"] == "cli_xxx"
+        assert raw["feishu"]["require_mention"] is True
+
+
 class TestVerifyOnStopMigration:
     """v30 → v31: switch verify_on_stop OFF once, preserving explicit choices."""
 

tests/hermes_cli/test_goals.py

@@ -166,8 +166,8 @@ class TestJudgeGoal:
         from hermes_cli import goals
 
         with patch(
-            "agent.auxiliary_client.get_text_auxiliary_client",
-            return_value=(None, None),
+            "agent.auxiliary_client.call_llm",
+            side_effect=RuntimeError("No LLM provider configured"),
         ):
             verdict, _, _, _wd = goals.judge_goal("my goal", "my response")
         assert verdict == "continue"
@@ -176,11 +176,9 @@ class TestJudgeGoal:
         """Judge exception → fail-open continue (don't wedge progress on judge bugs)."""
         from hermes_cli import goals
 
-        fake_client = MagicMock()
-        fake_client.chat.completions.create.side_effect = RuntimeError("boom")
         with patch(
-            "agent.auxiliary_client.get_text_auxiliary_client",
-            return_value=(fake_client, "judge-model"),
+            "agent.auxiliary_client.call_llm",
+            side_effect=RuntimeError("boom"),
         ):
             verdict, reason, _, _wd = goals.judge_goal("goal", "response")
         assert verdict == "continue"
@@ -189,17 +187,11 @@ class TestJudgeGoal:
     def test_judge_says_done(self):
         from hermes_cli import goals
 
-        fake_client = MagicMock()
-        fake_client.chat.completions.create.return_value = MagicMock(
-            choices=[
-                MagicMock(
-                    message=MagicMock(content='{"done": true, "reason": "achieved"}')
-                )
-            ]
-        )
         with patch(
-            "agent.auxiliary_client.get_text_auxiliary_client",
-            return_value=(fake_client, "judge-model"),
+            "agent.auxiliary_client.call_llm",
+            return_value=MagicMock(
+                choices=[MagicMock(message=MagicMock(content='{"done": true, "reason": "achieved"}'))]
+            ),
         ):
             verdict, reason, _, _wd = goals.judge_goal("goal", "agent response")
         assert verdict == "done"
@@ -208,17 +200,11 @@ class TestJudgeGoal:
     def test_judge_says_continue(self):
         from hermes_cli import goals
 
-        fake_client = MagicMock()
-        fake_client.chat.completions.create.return_value = MagicMock(
-            choices=[
-                MagicMock(
-                    message=MagicMock(content='{"done": false, "reason": "not yet"}')
-                )
-            ]
-        )
         with patch(
-            "agent.auxiliary_client.get_text_auxiliary_client",
-            return_value=(fake_client, "judge-model"),
+            "agent.auxiliary_client.call_llm",
+            return_value=MagicMock(
+                choices=[MagicMock(message=MagicMock(content='{"done": false, "reason": "not yet"}'))]
+            ),
         ):
             verdict, reason, _, _wd = goals.judge_goal("goal", "agent response")
         assert verdict == "continue"
@@ -448,11 +434,9 @@ class TestJudgeParseFailureAutoPause:
         """Transient network/API errors must not trip the auto-pause guard."""
         from hermes_cli import goals
 
-        fake_client = MagicMock()
-        fake_client.chat.completions.create.side_effect = RuntimeError("connection reset")
         with patch(
-            "agent.auxiliary_client.get_text_auxiliary_client",
-            return_value=(fake_client, "judge-model"),
+            "agent.auxiliary_client.call_llm",
+            side_effect=RuntimeError("connection reset"),
         ):
             verdict, _, parse_failed, _wd = goals.judge_goal("goal", "response")
         assert verdict == "continue"
@@ -462,13 +446,9 @@ class TestJudgeParseFailureAutoPause:
         """End-to-end: judge returns empty content → parse_failed=True."""
         from hermes_cli import goals
 
-        fake_client = MagicMock()
-        fake_client.chat.completions.create.return_value = MagicMock(
-            choices=[MagicMock(message=MagicMock(content=""))]
-        )
         with patch(
-            "agent.auxiliary_client.get_text_auxiliary_client",
-            return_value=(fake_client, "judge-model"),
+            "agent.auxiliary_client.call_llm",
+            return_value=MagicMock(choices=[MagicMock(message=MagicMock(content=""))]),
         ):
             verdict, _, parse_failed, _wd = goals.judge_goal("goal", "response")
         assert verdict == "continue"
@@ -747,22 +727,11 @@ class TestJudgeGoalWithSubgoals:
             message = _FakeMsg()
         class _FakeResp:
             choices = [_FakeChoice()]
-        class _FakeClient:
-            class chat:
-                class completions:
-                    @staticmethod
-                    def create(**kwargs):
-                        captured.update(kwargs)
-                        return _FakeResp()
+        def _fake_call_llm(**kwargs):
+            captured.update(kwargs)
+            return _FakeResp()
 
-        with patch.object(goals, "get_text_auxiliary_client",
-                          return_value=(_FakeClient, "fake-model"), create=True), \
-             patch.object(goals, "get_auxiliary_extra_body",
-                          return_value=None, create=True), \
-             patch("agent.auxiliary_client.get_text_auxiliary_client",
-                   return_value=(_FakeClient, "fake-model")), \
-             patch("agent.auxiliary_client.get_auxiliary_extra_body",
-                   return_value=None):
+        with patch("agent.auxiliary_client.call_llm", side_effect=_fake_call_llm):
             verdict, reason, parse_failed, _wd = goals.judge_goal(
                 "ship the feature",
                 "ok shipped",
@@ -790,18 +759,11 @@ class TestJudgeGoalWithSubgoals:
             message = _FakeMsg()
         class _FakeResp:
             choices = [_FakeChoice()]
-        class _FakeClient:
-            class chat:
-                class completions:
-                    @staticmethod
-                    def create(**kwargs):
-                        captured.update(kwargs)
-                        return _FakeResp()
+        def _fake_call_llm(**kwargs):
+            captured.update(kwargs)
+            return _FakeResp()
 
-        with patch("agent.auxiliary_client.get_text_auxiliary_client",
-                   return_value=(_FakeClient, "fake-model")), \
-             patch("agent.auxiliary_client.get_auxiliary_extra_body",
-                   return_value=None):
+        with patch("agent.auxiliary_client.call_llm", side_effect=_fake_call_llm):
             goals.judge_goal("ship it", "done", subgoals=None)
 
         sent_messages = captured.get("messages") or []
@@ -1363,7 +1325,8 @@ class TestGoalManagerContract:
 
 
 class TestJudgeWithContract:
-    def _fake_client(self, captured, content='{"done": false, "reason": "more"}'):
+    def _fake_call_llm(self, captured, content='{"done": false, "reason": "more"}'):
+        """judge_goal routes through call_llm (#35566) — capture its kwargs."""
         class _FakeMsg:
             pass
         _FakeMsg.content = content
@@ -1371,14 +1334,11 @@ class TestJudgeWithContract:
             message = _FakeMsg()
         class _FakeResp:
             choices = [_FakeChoice()]
-        class _FakeClient:
-            class chat:
-                class completions:
-                    @staticmethod
-                    def create(**kwargs):
-                        captured.update(kwargs)
-                        return _FakeResp()
-        return _FakeClient
+
+        def _fake(**kwargs):
+            captured.update(kwargs)
+            return _FakeResp()
+        return _fake
 
     def test_judge_uses_contract_template(self, hermes_home):
         from unittest.mock import patch
@@ -1386,10 +1346,8 @@ class TestJudgeWithContract:
         from hermes_cli.goals import GoalContract
 
         captured = {}
-        client = self._fake_client(captured)
-        with patch("agent.auxiliary_client.get_text_auxiliary_client",
-                   return_value=(client, "fake-model")), \
-             patch("agent.auxiliary_client.get_auxiliary_extra_body", return_value=None):
+        with patch("agent.auxiliary_client.call_llm",
+                   side_effect=self._fake_call_llm(captured)):
             goals.judge_goal(
                 "ship it", "I think it's done",
                 contract=GoalContract(verification="pytest -q passes"),
@@ -1407,10 +1365,8 @@ class TestJudgeWithContract:
         from hermes_cli.goals import GoalContract
 
         captured = {}
-        client = self._fake_client(captured)
-        with patch("agent.auxiliary_client.get_text_auxiliary_client",
-                   return_value=(client, "fake-model")), \
-             patch("agent.auxiliary_client.get_auxiliary_extra_body", return_value=None):
+        with patch("agent.auxiliary_client.call_llm",
+                   side_effect=self._fake_call_llm(captured)):
             goals.judge_goal(
                 "ship it", "done",
                 subgoals=["write changelog"],
@@ -1438,16 +1394,8 @@ class TestDraftContract:
             message = _FakeMsg()
         class _FakeResp:
             choices = [_FakeChoice()]
-        class _FakeClient:
-            class chat:
-                class completions:
-                    @staticmethod
-                    def create(**kwargs):
-                        return _FakeResp()
-
-        with patch("agent.auxiliary_client.get_text_auxiliary_client",
-                   return_value=(_FakeClient, "fake-model")), \
-             patch("agent.auxiliary_client.get_auxiliary_extra_body", return_value=None):
+        with patch("agent.auxiliary_client.call_llm",
+                   return_value=_FakeResp()):
             contract = goals.draft_contract("Migrate auth to JWT")
         assert contract is not None
         assert contract.outcome == "auth on JWT"
@@ -1464,24 +1412,16 @@ class TestDraftContract:
             message = _FakeMsg()
         class _FakeResp:
             choices = [_FakeChoice()]
-        class _FakeClient:
-            class chat:
-                class completions:
-                    @staticmethod
-                    def create(**kwargs):
-                        return _FakeResp()
-
-        with patch("agent.auxiliary_client.get_text_auxiliary_client",
-                   return_value=(_FakeClient, "fake-model")), \
-             patch("agent.auxiliary_client.get_auxiliary_extra_body", return_value=None):
+        with patch("agent.auxiliary_client.call_llm",
+                   return_value=_FakeResp()):
             assert goals.draft_contract("anything") is None
 
     def test_draft_returns_none_when_no_client(self, hermes_home):
         from unittest.mock import patch
         from hermes_cli import goals
 
-        with patch("agent.auxiliary_client.get_text_auxiliary_client",
-                   return_value=(None, None)):
+        with patch("agent.auxiliary_client.call_llm",
+                   side_effect=RuntimeError("No LLM provider configured")):
             assert goals.draft_contract("anything") is None
 
 
@@ -1495,7 +1435,8 @@ class TestContractAndBackgroundCompose:
     the contract block and the background-process list to the judge, so it
     can return either done (evidence met) or wait (parked on the poller)."""
 
-    def _capture_client(self, captured, content='{"verdict": "wait", "wait_on_pid": 4242, "reason": "CI still running"}'):
+    def _capture_call_llm(self, captured, content='{"verdict": "wait", "wait_on_pid": 4242, "reason": "CI still running"}'):
+        """judge_goal routes through call_llm (#35566) — capture its kwargs."""
         class _FakeMsg:
             pass
         _FakeMsg.content = content
@@ -1503,14 +1444,11 @@ class TestContractAndBackgroundCompose:
             message = _FakeMsg()
         class _FakeResp:
             choices = [_FakeChoice()]
-        class _FakeClient:
-            class chat:
-                class completions:
-                    @staticmethod
-                    def create(**kwargs):
-                        captured.update(kwargs)
-                        return _FakeResp()
-        return _FakeClient
+
+        def _fake(**kwargs):
+            captured.update(kwargs)
+            return _FakeResp()
+        return _fake
 
     def test_judge_prompt_carries_contract_and_background(self, hermes_home):
         from unittest.mock import patch
@@ -1518,14 +1456,12 @@ class TestContractAndBackgroundCompose:
         from hermes_cli.goals import GoalContract
 
         captured = {}
-        client = self._capture_client(captured)
         bg = [{
             "session_id": "ci-watch", "pid": 4242, "status": "running",
             "command": "wait_for_pr_green.sh 50501", "trigger": "exit",
         }]
-        with patch("agent.auxiliary_client.get_text_auxiliary_client",
-                   return_value=(client, "fake-model")), \
-             patch("agent.auxiliary_client.get_auxiliary_extra_body", return_value=None):
+        with patch("agent.auxiliary_client.call_llm",
+                   side_effect=self._capture_call_llm(captured)):
             verdict, reason, parse_failed, wait_directive = goals.judge_goal(
                 "ship the PR",
                 "I pushed and started the CI watcher; waiting on it now.",
@@ -1550,14 +1486,12 @@ class TestContractAndBackgroundCompose:
         from hermes_cli.goals import GoalContract
 
         captured = {}
-        client = self._capture_client(
-            captured,
-            content='{"verdict": "done", "reason": "CI is green, evidence shown"}',
-        )
         bg = [{"session_id": "ci", "pid": 4242, "status": "running", "command": "ci", "trigger": "exit"}]
-        with patch("agent.auxiliary_client.get_text_auxiliary_client",
-                   return_value=(client, "fake-model")), \
-             patch("agent.auxiliary_client.get_auxiliary_extra_body", return_value=None):
+        with patch("agent.auxiliary_client.call_llm",
+                   side_effect=self._capture_call_llm(
+                       captured,
+                       content='{"verdict": "done", "reason": "CI is green, evidence shown"}',
+                   )):
             verdict, reason, parse_failed, wait_directive = goals.judge_goal(
                 "ship the PR",
                 "CI finished: 30 passed, 0 failed. Done.",

tests/hermes_cli/test_input_sanitize.py

@@ -0,0 +1,51 @@
+"""Tests for shared user prompt input sanitization."""
+
+from hermes_cli.input_sanitize import (
+    collapse_repeated_input_artifacts,
+    sanitize_user_prompt_text,
+    strip_leaked_bracketed_paste_wrappers,
+)
+
+
+class TestStripLeakedBracketedPasteWrappers:
+    def test_plain_text_unchanged(self):
+        assert strip_leaked_bracketed_paste_wrappers("hello world") == "hello world"
+
+    def test_strips_canonical_escape_wrappers(self):
+        assert strip_leaked_bracketed_paste_wrappers("\x1b[200~hello\x1b[201~") == "hello"
+
+    def test_strips_visible_caret_escape_wrappers(self):
+        assert strip_leaked_bracketed_paste_wrappers("^[[200~hello^[[201~") == "hello"
+
+    def test_does_not_strip_non_wrapper_bracket_forms_in_normal_text(self):
+        text = "literal[200~tag and literal[201~tag should stay"
+        assert strip_leaked_bracketed_paste_wrappers(text) == text
+
+
+class TestCollapseRepeatedInputArtifacts:
+    def test_issue_62557_corruption_tail(self):
+        prefix = "需要时随时叫我。"
+        tail = "[e~[[e" + "~[[e" * 20
+        assert collapse_repeated_input_artifacts(prefix + tail) == prefix
+
+    def test_plain_text_unchanged(self):
+        text = "build00~tag should stay"
+        assert collapse_repeated_input_artifacts(text) == text
+
+    def test_mid_string_marker_followed_by_suffix_preserved(self):
+        text = "notes ~[[e more text here"
+        assert collapse_repeated_input_artifacts(text) == text
+
+    def test_trailing_punctuation_preserved(self):
+        assert collapse_repeated_input_artifacts("wait....") == "wait...."
+
+    def test_insufficient_tail_repeats_preserved(self):
+        text = "hello~[[e~[[e"
+        assert collapse_repeated_input_artifacts(text) == text
+
+
+class TestSanitizeUserPromptText:
+    def test_combines_wrapper_strip_and_tail_collapse(self):
+        prefix = "hello["
+        corrupted = prefix + "~[[e" * 8
+        assert sanitize_user_prompt_text(corrupted) == "hello"

tests/hermes_cli/test_kanban_cli.py

@@ -465,11 +465,10 @@ def test_run_slash_specify_end_to_end(kanban_home, monkeypatch):
     resp.choices[0].message.content = (
         '{"title": "Spec: rough idea", "body": "**Goal**\\nShip it."}'
     )
-    fake_client = MagicMock()
-    fake_client.chat.completions.create = MagicMock(return_value=resp)
+    # specify_task routes through call_llm now (#35566) — mock it directly.
     monkeypatch.setattr(
-        "agent.auxiliary_client.get_text_auxiliary_client",
-        lambda *a, **kw: (fake_client, "test-model"),
+        "agent.auxiliary_client.call_llm",
+        MagicMock(return_value=resp),
     )
 
     # Specify via slash.

tests/hermes_cli/test_kanban_decompose.py

@@ -41,18 +41,19 @@ def _mock_client_returning(content: str):
 
 
 def _patch_aux_client(content: str, *, model: str = "test-model"):
-    client = _mock_client_returning(content)
+    # decompose_task now routes through call_llm (see #35566) — mock it at
+    # the source module so task config, extra_body, and retries stay out of
+    # unit-test scope.
     return patch(
-        "agent.auxiliary_client.get_text_auxiliary_client",
-        return_value=(client, model),
+        "agent.auxiliary_client.call_llm",
+        return_value=_fake_aux_response(content),
     )
 
 
 def _patch_extra_body():
-    return patch(
-        "agent.auxiliary_client.get_auxiliary_extra_body",
-        return_value={},
-    )
+    # No-op shim retained for call-site compatibility: extra_body plumbing
+    # now lives inside call_llm, which _patch_aux_client already mocks.
+    return patch("agent.auxiliary_client.get_auxiliary_extra_body", return_value={})
 
 
 def _patch_list_profiles(names: list[str]):
@@ -334,9 +335,11 @@ def test_decompose_no_aux_client_configured(kanban_home):
     for p in patches:
         p.start()
     try:
+        # call_llm raises RuntimeError when no provider is configured; the
+        # decomposer must convert that into a failed outcome, not a crash.
         with patch(
-            "agent.auxiliary_client.get_text_auxiliary_client",
-            return_value=(None, ""),
+            "agent.auxiliary_client.call_llm",
+            side_effect=RuntimeError("No LLM provider configured"),
         ):
             outcome = decomp.decompose_task(tid, author="me")
     finally:
@@ -344,4 +347,5 @@ def test_decompose_no_aux_client_configured(kanban_home):
             p.stop()
 
     assert outcome.ok is False
-    assert "no auxiliary client" in outcome.reason
+    # call_llm's no-provider RuntimeError surfaces via the LLM-error branch.
+    assert "LLM error" in outcome.reason

tests/hermes_cli/test_kanban_specify.py

@@ -48,15 +48,12 @@ def _mock_client_returning(content: str):
 
 
 def _patch_aux_client(content: str, *, model: str = "test-model"):
-    """Patch get_text_auxiliary_client at its source + at the module that
-    imported it lazily inside specify_task. Both patches are needed
-    because kanban_specify imports the function inside the function body.
+    """Patch call_llm at its source module — specify_task now routes through
+    it (#35566) instead of building a raw client. Returns (patcher, mock) so
+    callers can still assert on the call.
     """
-    client = _mock_client_returning(content)
-    return patch(
-        "agent.auxiliary_client.get_text_auxiliary_client",
-        return_value=(client, model),
-    ), client
+    mock_fn = MagicMock(return_value=_fake_aux_response(content))
+    return patch("agent.auxiliary_client.call_llm", mock_fn), mock_fn
 
 
 # ---------------------------------------------------------------------------
@@ -159,13 +156,14 @@ def test_specify_task_no_aux_client_configured(kanban_home):
         tid = kb.create_task(conn, title="rough", triage=True)
 
     with patch(
-        "agent.auxiliary_client.get_text_auxiliary_client",
-        return_value=(None, ""),
+        "agent.auxiliary_client.call_llm",
+        side_effect=RuntimeError("No LLM provider configured"),
     ):
         outcome = spec.specify_task(tid)
 
     assert outcome.ok is False
-    assert "auxiliary client" in outcome.reason
+    # call_llm's no-provider RuntimeError surfaces via the LLM-error branch.
+    assert "LLM error" in outcome.reason
     # Task must stay in triage — we never touched it.
     with kb.connect() as conn:
         assert kb.get_task(conn, tid).status == "triage"
@@ -176,10 +174,9 @@ def test_specify_task_llm_api_error_keeps_task_in_triage(kanban_home):
         tid = kb.create_task(conn, title="rough", triage=True)
 
     client = MagicMock()
-    client.chat.completions.create = MagicMock(side_effect=RuntimeError("429 rate limited"))
     with patch(
-        "agent.auxiliary_client.get_text_auxiliary_client",
-        return_value=(client, "test-model"),
+        "agent.auxiliary_client.call_llm",
+        side_effect=RuntimeError("429 rate limited"),
     ):
         outcome = spec.specify_task(tid)
 
@@ -288,8 +285,8 @@ def test_cli_specify_all_returns_1_when_every_task_fails(kanban_home, capsys):
         kb.create_task(conn, title="b", triage=True)
 
     with patch(
-        "agent.auxiliary_client.get_text_auxiliary_client",
-        return_value=(None, ""),  # no aux client → every task fails
+        "agent.auxiliary_client.call_llm",
+        side_effect=RuntimeError("No LLM provider configured"),  # every task fails
     ):
         rc = _run_cli("specify", "--all")
 

tests/hermes_cli/test_mcp_catalog.py

@@ -7,6 +7,7 @@ launch an MCP is mocked.
 
 from __future__ import annotations
 
+import re
 from pathlib import Path
 from unittest.mock import patch
 
@@ -197,6 +198,32 @@ class TestManifestParsing:
         assert get_entry("official/demo") is not None
         assert get_entry("missing") is None
 
+    def test_transport_env_parsed_and_written_to_server_config(self, catalog_dir):
+        body = _basic_manifest()
+        body["transport"]["env"] = {"DISABLE_TELEMETRY": "true"}
+        _write_manifest(catalog_dir, "demo", body)
+        from hermes_cli.mcp_catalog import _build_server_config
+
+        e = _entry("demo")
+        assert e.transport.env == {"DISABLE_TELEMETRY": "true"}
+        cfg = _build_server_config(e, None)
+        assert cfg["env"] == {"DISABLE_TELEMETRY": "true"}
+
+    def test_transport_env_absent_leaves_config_without_env_key(self, catalog_dir):
+        _write_manifest(catalog_dir, "demo", _basic_manifest())
+        from hermes_cli.mcp_catalog import _build_server_config
+
+        cfg = _build_server_config(_entry("demo"), None)
+        assert "env" not in cfg
+
+    def test_transport_env_bad_shape_rejected(self, catalog_dir):
+        body = _basic_manifest()
+        body["transport"]["env"] = ["DISABLE_TELEMETRY=true"]  # list, not mapping
+        _write_manifest(catalog_dir, "demo", body)
+        from hermes_cli.mcp_catalog import list_catalog
+
+        assert list_catalog() == []
+
 
 # ---------------------------------------------------------------------------
 # Install flow
@@ -812,3 +839,59 @@ class TestShippedCatalog:
             assert entry.name
             assert entry.description
             assert entry.transport.type in ("stdio", "http")
+
+    def test_all_shipped_manifests_are_version_locked(self, monkeypatch):
+        """Contract: catalog entries follow the same supply-chain rules as
+        pyproject dependencies — everything Hermes fetches/launches is pinned
+        to an exact version.
+
+        - git installs must pin a full 40-char commit SHA (branches and tags
+          can be moved by the upstream owner; SHAs cannot).
+        - package-launcher stdio transports (uvx/npx and their pkg-manager
+          equivalents) must carry an exact version specifier on the package
+          arg (``pkg==X`` for Python, ``pkg@X`` for npm).
+
+        http transports and ${INSTALL_DIR}-anchored commands have nothing to
+        pin at the transport layer (the server runs elsewhere / comes from the
+        SHA-pinned clone), so they're exempt.
+        """
+        monkeypatch.delenv("HERMES_OPTIONAL_MCPS", raising=False)
+        from hermes_cli.mcp_catalog import _catalog_root, _parse_manifest
+
+        root = _catalog_root()
+        if not root.exists():
+            pytest.skip("optional-mcps/ not present in this checkout")
+
+        launcher_commands = {"uvx", "npx", "pipx", "bunx", "pnpx"}
+        problems = []
+        for m in root.glob("*/manifest.yaml"):
+            entry = _parse_manifest(m)
+
+            if entry.install is not None:
+                if not re.fullmatch(r"[0-9a-f]{40}", entry.install.ref):
+                    problems.append(
+                        f"{entry.name}: install.ref {entry.install.ref!r} is not "
+                        "a full 40-char commit SHA"
+                    )
+
+            t = entry.transport
+            if t.type == "stdio" and (t.command or "") in launcher_commands:
+                pkg_args = [a for a in t.args if not a.startswith("-")]
+                if not pkg_args:
+                    problems.append(f"{entry.name}: launcher {t.command} has no package arg")
+                    continue
+                pkg = pkg_args[0]
+                # Exact-pin shapes: pkg==1.2.3 (uvx/pipx) or pkg@1.2.3 /
+                # @scope/pkg@1.2.3 (npx/bunx/pnpx). The version must start
+                # with a digit — a bare name, a range operator, or an npm
+                # dist-tag (@latest, @next) floats and is rejected.
+                exact = re.fullmatch(r"[^=@\s]+==\d[\w.\-+]*", pkg) or re.fullmatch(
+                    r"(@[\w.\-]+/)?[\w.\-]+@\d[\w.\-+]*", pkg
+                )
+                if not exact:
+                    problems.append(
+                        f"{entry.name}: package arg {pkg!r} is not pinned to an "
+                        "exact version (expected pkg==X or pkg@X)"
+                    )
+
+        assert not problems, "unpinned catalog entries:\n" + "\n".join(problems)

tests/hermes_cli/test_moa_config.py

@@ -304,3 +304,115 @@ def test_reference_max_tokens_in_flattened_view():
     active preset's reference_max_tokens."""
     cfg = normalize_moa_config(_preset(reference_max_tokens=750))
     assert cfg["reference_max_tokens"] == 750
+
+
+# ── validate_moa_payload (write-boundary validation, #64156) ─────────────────
+#
+# normalize_moa_config is deliberately tolerant at READ time (hand-edited
+# configs degrade to defaults). validate_moa_payload is the strict WRITE-time
+# counterpart: it must flag exactly the payloads normalize would silently
+# repair, so API save paths reject them instead of corrupting user config.
+
+
+def _valid_preset_payload():
+    return {
+        "reference_models": [{"provider": "openrouter", "model": "deepseek/deepseek-v4-pro"}],
+        "aggregator": {"provider": "openrouter", "model": "anthropic/claude-opus-4.8"},
+    }
+
+
+def test_validate_moa_payload_accepts_complete_presets():
+    from hermes_cli.moa_config import validate_moa_payload
+
+    assert validate_moa_payload({"presets": {"default": _valid_preset_payload()}}) == []
+
+
+def test_validate_moa_payload_accepts_legacy_flat_payload():
+    from hermes_cli.moa_config import validate_moa_payload
+
+    assert validate_moa_payload(_valid_preset_payload()) == []
+
+
+def test_validate_moa_payload_flags_half_filled_reference_slot():
+    """The #64156 shape: provider picked, model still empty (mid-edit autosave)."""
+    from hermes_cli.moa_config import validate_moa_payload
+
+    preset = _valid_preset_payload()
+    preset["reference_models"].append({"provider": "kilo", "model": ""})
+    problems = validate_moa_payload({"presets": {"default": preset}})
+
+    assert problems
+    assert any("reference 2" in p and "model is required" in p for p in problems)
+
+
+def test_validate_moa_payload_flags_half_filled_aggregator():
+    from hermes_cli.moa_config import validate_moa_payload
+
+    preset = _valid_preset_payload()
+    preset["aggregator"] = {"provider": "openrouter", "model": ""}
+    problems = validate_moa_payload({"presets": {"default": preset}})
+
+    assert any("aggregator" in p and "model is required" in p for p in problems)
+
+
+def test_validate_moa_payload_flags_empty_references():
+    from hermes_cli.moa_config import validate_moa_payload
+
+    preset = _valid_preset_payload()
+    preset["reference_models"] = []
+    problems = validate_moa_payload({"presets": {"default": preset}})
+
+    assert any("at least one complete reference model" in p for p in problems)
+
+
+def test_validate_moa_payload_flags_recursive_moa_slot():
+    from hermes_cli.moa_config import validate_moa_payload
+
+    preset = _valid_preset_payload()
+    preset["aggregator"] = {"provider": "MoA", "model": "default"}
+    problems = validate_moa_payload({"presets": {"default": preset}})
+
+    assert any("recursive MoA" in p for p in problems)
+
+
+def test_validate_moa_payload_names_the_broken_preset():
+    """Multi-preset payloads must say WHICH preset is broken."""
+    from hermes_cli.moa_config import validate_moa_payload
+
+    problems = validate_moa_payload(
+        {
+            "presets": {
+                "good": _valid_preset_payload(),
+                "broken": {
+                    "reference_models": [{"provider": "", "model": ""}],
+                    "aggregator": {"provider": "a", "model": "b"},
+                },
+            }
+        }
+    )
+
+    assert problems
+    assert all("'broken'" in p for p in problems)
+    assert not any("'good'" in p for p in problems)
+
+
+def test_validate_moa_payload_agrees_with_clean_slot():
+    """Contract: a payload validate accepts must survive normalize UNCHANGED in
+    its slots — validate and _clean_slot can never disagree (else a payload
+    could pass validation and still be swapped for defaults)."""
+    from hermes_cli.moa_config import validate_moa_payload
+
+    payload = {"presets": {"p": _valid_preset_payload()}}
+    assert validate_moa_payload(payload) == []
+
+    cfg = normalize_moa_config(payload)
+    assert cfg["presets"]["p"]["reference_models"] == payload["presets"]["p"]["reference_models"]
+    assert cfg["presets"]["p"]["aggregator"] == payload["presets"]["p"]["aggregator"]
+
+
+def test_validate_moa_payload_rejects_non_dict():
+    from hermes_cli.moa_config import validate_moa_payload
+
+    assert validate_moa_payload(None)
+    assert validate_moa_payload([1, 2])
+    assert validate_moa_payload({"presets": {"p": "not-a-dict"}})

tests/hermes_cli/test_profile_describer.py

@@ -79,11 +79,11 @@ def _fake_aux_response(content: str):
 
 
 def _patch_aux_client(content: str):
-    client = MagicMock()
-    client.chat.completions.create = MagicMock(return_value=_fake_aux_response(content))
+    # describe_profile now routes through call_llm (#35566) — mock it at the
+    # source module.
     return patch(
-        "agent.auxiliary_client.get_text_auxiliary_client",
-        return_value=(client, "test-model"),
+        "agent.auxiliary_client.call_llm",
+        return_value=_fake_aux_response(content),
     )
 
 

tests/hermes_cli/test_web_server.py

@@ -861,6 +861,72 @@ class TestWebServerEndpoints:
         assert cfg["moa"]["reference_models"] == payload["reference_models"]
         assert cfg["moa"]["aggregator"] == payload["aggregator"]
 
+    def test_put_moa_models_rejects_half_filled_slot_with_422(self):
+        """#64156: a mid-edit autosave (provider picked, model empty) used to be
+        silently normalized into the hardcoded default preset — the user's
+        config was replaced without any error. The write path must reject it."""
+        from hermes_cli.config import load_config
+
+        original = load_config().get("moa")
+
+        payload = {
+            "presets": {
+                "default": {
+                    "reference_models": [{"provider": "kilo", "model": ""}],
+                    "aggregator": {"provider": "openrouter", "model": "anthropic/claude-opus-4.8"},
+                }
+            }
+        }
+
+        resp = self.client.put("/api/model/moa", json=payload)
+        assert resp.status_code == 422
+        assert "model is required" in resp.json()["detail"]
+        # Config untouched — not swapped for defaults.
+        assert load_config().get("moa") == original
+
+    def test_put_moa_models_rejects_half_filled_aggregator_with_422(self):
+        payload = {
+            "presets": {
+                "default": {
+                    "reference_models": [{"provider": "openrouter", "model": "deepseek/deepseek-v4-pro"}],
+                    "aggregator": {"provider": "openrouter", "model": ""},
+                }
+            }
+        }
+
+        resp = self.client.put("/api/model/moa", json=payload)
+        assert resp.status_code == 422
+        assert "aggregator" in resp.json()["detail"]
+
+    def test_put_moa_models_round_trips_fanout_and_reference_max_tokens(self):
+        """GET → PUT round-trip must not erase newer per-preset knobs. The old
+        Pydantic payload didn't declare fanout / reference_max_tokens, so any
+        client save silently wiped hand-set values back to defaults."""
+        from hermes_cli.config import load_config
+
+        payload = {
+            "presets": {
+                "default": {
+                    "reference_models": [{"provider": "openrouter", "model": "deepseek/deepseek-v4-pro"}],
+                    "aggregator": {"provider": "openrouter", "model": "anthropic/claude-opus-4.8"},
+                    "fanout": "user_turn",
+                    "reference_max_tokens": 600,
+                }
+            }
+        }
+
+        resp = self.client.put("/api/model/moa", json=payload)
+        assert resp.status_code == 200
+
+        saved = load_config()["moa"]["presets"]["default"]
+        assert saved["fanout"] == "user_turn"
+        assert saved["reference_max_tokens"] == 600
+
+        # And the GET view carries them back to the client.
+        fetched = self.client.get("/api/model/moa").json()
+        assert fetched["presets"]["default"]["fanout"] == "user_turn"
+        assert fetched["presets"]["default"]["reference_max_tokens"] == 600
+
     # ── GET /api/media (remote image display) ───────────────────────────
 
     def test_get_media_serves_image_in_root(self):

tests/plugins/model_providers/test_custom_profile.py

@@ -49,20 +49,26 @@ class TestCustomReasoningWireShape:
         assert tl == {}
 
     def test_disabled_sends_think_false(self, custom_profile):
-        """enabled=False → extra_body.think = False (Ollama thinking-off flag)."""
+        """enabled=False → reasoning_effort='none' top-level + think=False.
+
+        Both fields are required: Ollama's /v1/chat/completions silently
+        ignores extra_body.think (only /api/chat honours it — ollama#14820)
+        but respects top-level reasoning_effort (#25758). think=False stays
+        for proxies and the native /api/chat path.
+        """
         eb, tl = custom_profile.build_api_kwargs_extras(
             reasoning_config={"enabled": False}, model="glm-5.2"
         )
         assert eb == {"think": False}
-        assert tl == {}
+        assert tl == {"reasoning_effort": "none"}
 
     def test_effort_none_sends_think_false(self, custom_profile):
-        """effort='none' is the disable alias → think=False, no effort."""
+        """effort='none' is the disable alias → same dual emission."""
         eb, tl = custom_profile.build_api_kwargs_extras(
             reasoning_config={"enabled": True, "effort": "none"}, model="glm-5.2"
         )
         assert eb == {"think": False}
-        assert tl == {}
+        assert tl == {"reasoning_effort": "none"}
 
     @pytest.mark.parametrize(
         "effort", ["minimal", "low", "medium", "high", "xhigh", "max"]

tests/plugins/test_kanban_dashboard_plugin.py

@@ -2182,13 +2182,10 @@ def _patch_specifier_response(monkeypatch, *, content, model="test-model"):
     resp = MagicMock()
     resp.choices = [MagicMock()]
     resp.choices[0].message.content = content
-    fake_client = MagicMock()
-    fake_client.chat.completions.create = MagicMock(return_value=resp)
-    monkeypatch.setattr(
-        "agent.auxiliary_client.get_text_auxiliary_client",
-        lambda *a, **kw: (fake_client, model),
-    )
-    return fake_client
+    # specify_task routes through call_llm now (#35566) — mock it directly.
+    fake_call = MagicMock(return_value=resp)
+    monkeypatch.setattr("agent.auxiliary_client.call_llm", fake_call)
+    return fake_call
 
 
 def test_specify_happy_path(client, monkeypatch):
@@ -2250,11 +2247,11 @@ def test_specify_no_aux_client_surfaces_reason(client, monkeypatch):
         json={"title": "rough", "triage": True},
     ).json()["task"]
 
-    # Simulate "no auxiliary client configured".
-    monkeypatch.setattr(
-        "agent.auxiliary_client.get_text_auxiliary_client",
-        lambda *a, **kw: (None, ""),
-    )
+    # Simulate "no auxiliary client configured" — call_llm raises when
+    # no provider resolves (#35566 routing).
+    def _no_provider(**kwargs):
+        raise RuntimeError("No LLM provider configured")
+    monkeypatch.setattr("agent.auxiliary_client.call_llm", _no_provider)
 
     r = client.post(
         f"/api/plugins/kanban/tasks/{t['id']}/specify",
@@ -2263,7 +2260,8 @@ def test_specify_no_aux_client_surfaces_reason(client, monkeypatch):
     assert r.status_code == 200
     body = r.json()
     assert body["ok"] is False
-    assert "auxiliary client" in body["reason"]
+    # call_llm's no-provider RuntimeError surfaces via the LLM-error branch.
+    assert "LLM error" in body["reason"]
 
     # Task must stay in triage — nothing was touched.
     detail = client.get(f"/api/plugins/kanban/tasks/{t['id']}").json()["task"]

tests/run_agent/test_run_agent.py

@@ -4149,6 +4149,66 @@ class TestRunConversation:
         assert result["final_response"] == "Final answer"
         assert result["completed"] is True
 
+    def test_codex_content_filter_incomplete_routes_to_policy_fallback(self, agent):
+        self._setup_agent(agent)
+        agent.api_mode = "codex_responses"
+        agent.provider = "openai-codex"
+        agent.base_url = "https://chatgpt.com/backend-api/codex"
+        agent._base_url_lower = agent.base_url.lower()
+        agent._base_url_hostname = "chatgpt.com"
+        agent.model = "gpt-5.5"
+        agent._fallback_chain = [
+            {"provider": "openrouter", "model": "anthropic/claude-sonnet-4.7"},
+        ]
+        agent._fallback_index = 0
+
+        content_filter_response = SimpleNamespace(
+            status="incomplete",
+            incomplete_details=SimpleNamespace(reason="content_filter"),
+            output=[],
+            output_text="",
+            model="gpt-5.5",
+            usage=None,
+        )
+        fallback_response = SimpleNamespace(
+            status="completed",
+            incomplete_details=None,
+            output=[
+                SimpleNamespace(
+                    type="message",
+                    status="completed",
+                    content=[SimpleNamespace(type="output_text", text="Recovered on fallback")],
+                )
+            ],
+            model="fallback/model",
+            usage=None,
+        )
+        hook_events = []
+
+        def _fake_activate(reason=None):
+            agent._fallback_index = len(agent._fallback_chain)
+            return True
+
+        with (
+            patch.object(agent, "_create_request_openai_client", return_value=MagicMock()),
+            patch.object(agent, "_close_request_openai_client"),
+            patch.object(agent, "_run_codex_stream", side_effect=[content_filter_response, fallback_response]) as mock_run_codex_stream,
+            patch.object(agent, "_try_activate_fallback", side_effect=_fake_activate) as mock_try_activate_fallback,
+            patch.object(agent, "_invoke_api_request_error_hook", side_effect=lambda **kw: hook_events.append(kw)),
+            patch.object(agent, "_persist_session"),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+        ):
+            result = agent.run_conversation("summarize this large Slack thread")
+
+        assert result["final_response"] == "Recovered on fallback"
+        assert result["completed"] is True
+        mock_try_activate_fallback.assert_called_once_with()
+        assert mock_run_codex_stream.call_count == 2
+        assert hook_events[0]["error_type"] == "ContentPolicyBlocked"
+        assert hook_events[0]["retryable"] is False
+        assert hook_events[0]["reason"] == FailoverReason.content_policy_blocked.value
+
     def test_ollama_small_runtime_context_fails_before_api_call(self, agent, caplog):
         self._setup_agent(agent)
         agent.model = "qwen3.5:9b"

tests/run_agent/test_run_agent_codex_responses.py

@@ -1704,6 +1704,90 @@ def test_mid_turn_compaction_does_not_double_persist_in_place_rows(monkeypatch,
     )
 
 
+def _codex_incomplete_with_reasoning(text: str, reasoning_id: str = "rs_default"):
+    """Incomplete response with a reasoning item whose id/encrypted_content
+    can vary independently of the visible message text."""
+    return SimpleNamespace(
+        output=[
+            SimpleNamespace(
+                type="reasoning",
+                id=reasoning_id,
+                encrypted_content=f"opaque_{reasoning_id}",
+                summary=[SimpleNamespace(text="thinking...")],
+            ),
+            SimpleNamespace(
+                type="message",
+                status="in_progress",
+                content=[SimpleNamespace(type="output_text", text=text)],
+            ),
+        ],
+        usage=SimpleNamespace(input_tokens=4, output_tokens=2, total_tokens=6),
+        status="in_progress",
+        model="gpt-5-codex",
+    )
+
+
+def test_codex_incomplete_visible_dedup_suppresses_duplicate_interims(monkeypatch):
+    """Two consecutive incomplete responses with identical visible content
+    but different opaque reasoning items should be collapsed — only the first
+    interim is emitted to the user (#52711)."""
+    agent = _build_agent(monkeypatch)
+    # 2 incompletes with same text but different reasoning ids, then a final.
+    # (Only 2 to avoid hitting the cap of 3.)
+    responses = [
+        _codex_incomplete_with_reasoning("Working on it...", "rs_1"),
+        _codex_incomplete_with_reasoning("Working on it...", "rs_2"),
+        _codex_message_response("Done."),
+    ]
+    monkeypatch.setattr(agent, "_interruptible_api_call", lambda api_kwargs: responses.pop(0))
+
+    emitted: list = []
+    original_emit = agent._emit_interim_assistant_message
+    def _capture_emit(msg):
+        emitted.append(msg.get("content"))
+        original_emit(msg)
+    monkeypatch.setattr(agent, "_emit_interim_assistant_message", _capture_emit)
+
+    result = agent.run_conversation("test dedup")
+
+    assert result["completed"] is True
+    # Only ONE interim should have been emitted (the first), not two.
+    assert len(emitted) == 1
+    assert emitted[0] == "Working on it..."
+
+
+def test_codex_incomplete_opaque_state_updated_in_place(monkeypatch):
+    """When visible content is a duplicate, the last message's opaque state
+    (codex_reasoning_items) should be updated in-place without emitting a new
+    interim (#52711)."""
+    agent = _build_agent(monkeypatch)
+    responses = [
+        _codex_incomplete_with_reasoning("Partial output...", "rs_1"),
+        _codex_incomplete_with_reasoning("Partial output...", "rs_2"),
+        _codex_message_response("Final."),
+    ]
+    monkeypatch.setattr(agent, "_interruptible_api_call", lambda api_kwargs: responses.pop(0))
+
+    result = agent.run_conversation("test opaque update")
+
+    assert result["completed"] is True
+    # Find the incomplete interim message in the result.
+    incompletes = [
+        m for m in result["messages"]
+        if m.get("role") == "assistant" and m.get("finish_reason") == "incomplete"
+    ]
+    # Only one incomplete message should exist (the second was deduped).
+    assert len(incompletes) == 1
+    # The opaque state should reflect the LATEST reasoning item (rs_2),
+    # updated in-place on the single message.
+    items = incompletes[0].get("codex_reasoning_items")
+    if items:
+        assert any(
+            (i.get("id") if isinstance(i, dict) else getattr(i, "id", None)) == "rs_2"
+            for i in items
+        )
+
+
 def test_normalize_codex_response_marks_commentary_only_message_as_incomplete(monkeypatch):
     agent = _build_agent(monkeypatch)
     from agent.codex_responses_adapter import _normalize_codex_response
@@ -2608,7 +2692,8 @@ def test_codex_message_item_status_survives_conversion_and_preflight(monkeypatch
 
 def test_duplicate_detection_distinguishes_different_codex_reasoning(monkeypatch):
     """Two consecutive reasoning-only responses with different encrypted content
-    must NOT be treated as duplicates."""
+    are deduped on visible content — only one interim is kept, but opaque state
+    is updated in-place (#52711)."""
     agent = _build_agent(monkeypatch)
     responses = [
         # First reasoning-only response
@@ -2641,23 +2726,23 @@ def test_duplicate_detection_distinguishes_different_codex_reasoning(monkeypatch
 
     assert result["completed"] is True
     assert result["final_response"] == "Final answer after thinking."
-    # Both reasoning-only interim messages should be in history (not collapsed)
+    # Only one reasoning-only interim should be in history (deduped on
+    # visible content — both have empty visible output).
     interim_msgs = [
         msg for msg in result["messages"]
         if msg.get("role") == "assistant"
         and msg.get("finish_reason") == "incomplete"
     ]
-    assert len(interim_msgs) == 2
-    encrypted_contents = [
-        msg["codex_reasoning_items"][0]["encrypted_content"]
-        for msg in interim_msgs
-    ]
-    assert "enc_first" in encrypted_contents
-    assert "enc_second" in encrypted_contents
+    assert len(interim_msgs) == 1
+    # But the opaque state should reflect the LATEST reasoning item.
+    items = interim_msgs[0].get("codex_reasoning_items")
+    if items:
+        assert items[0].get("encrypted_content") == "enc_second"
 
 
 def test_duplicate_detection_distinguishes_different_codex_message_items(monkeypatch):
-    """Incomplete turns with new message ids/phases/statuses must not be collapsed."""
+    """Incomplete turns with same visible content but different message ids
+    are deduped — only one interim kept, opaque state updated in-place (#52711)."""
     agent = _build_agent(monkeypatch)
     responses = [
         SimpleNamespace(
@@ -2700,12 +2785,12 @@ def test_duplicate_detection_distinguishes_different_codex_message_items(monkeyp
         if msg.get("role") == "assistant"
         and msg.get("finish_reason") == "incomplete"
     ]
-    assert len(interim_msgs) == 2
-    assert [msg["codex_message_items"][0]["id"] for msg in interim_msgs] == [
-        "msg_first",
-        "msg_second",
-    ]
-    assert all(msg["codex_message_items"][0]["status"] == "in_progress" for msg in interim_msgs)
+    # Only one interim — deduped on visible content ("Still working..." == "Still working...").
+    assert len(interim_msgs) == 1
+    # Opaque state should reflect the latest message item.
+    items = interim_msgs[0].get("codex_message_items")
+    if items:
+        assert items[0].get("id") == "msg_second"
 
 
 def test_chat_messages_to_responses_input_deduplicates_reasoning_ids(monkeypatch):

tests/test_tui_gateway_server.py

@@ -5093,6 +5093,51 @@ def test_prompt_submit_history_version_mismatch_surfaces_warning(monkeypatch):
         server._sessions.pop("sid", None)
 
 
+def test_prompt_submit_sanitizes_bracketed_paste_before_agent(monkeypatch):
+    """prompt.submit must sanitize corrupted user text before run_conversation."""
+    captured: dict[str, str] = {}
+
+    class _Agent:
+        def run_conversation(
+            self, prompt, conversation_history=None, stream_callback=None
+        ):
+            captured["prompt"] = prompt
+            return {
+                "final_response": "ok",
+                "messages": [{"role": "assistant", "content": "ok"}],
+            }
+
+    class _ImmediateThread:
+        def __init__(self, target=None, daemon=None, **kw):
+            self._target = target
+
+        def start(self):
+            self._target()
+
+    corrupted = "hello[" + "~[[e" * 8
+    server._sessions["sid"] = _session(agent=_Agent())
+    try:
+        monkeypatch.setattr(server.threading, "Thread", _ImmediateThread)
+        monkeypatch.setattr(server, "_get_usage", lambda _a: {})
+        monkeypatch.setattr(server, "render_message", lambda _t, _c: "")
+        monkeypatch.setattr(server, "_emit", lambda *a: None)
+        monkeypatch.setattr(server, "_start_agent_build", lambda *a, **k: None)
+        monkeypatch.setattr(server, "_ensure_session_db_row", lambda *a, **k: None)
+        monkeypatch.setattr(server, "_persist_branch_seed", lambda *a, **k: None)
+
+        resp = server.handle_request(
+            {
+                "id": "1",
+                "method": "prompt.submit",
+                "params": {"session_id": "sid", "text": corrupted},
+            }
+        )
+        assert resp.get("result"), f"got error: {resp.get('error')}"
+        assert captured["prompt"] == "hello"
+    finally:
+        server._sessions.pop("sid", None)
+
+
 def test_prompt_submit_history_version_match_persists_normally(monkeypatch):
     """Regression guard: the backstop does not affect the happy path."""
 

tests/test_wal_checkpoint_strategy.py

@@ -0,0 +1,138 @@
+"""Tests for SessionDB WAL checkpoint strategy (issue #45383).
+
+Verifies that periodic checkpoints use PASSIVE mode (safe for large DBs)
+while close() and pre-VACUUM paths still use TRUNCATE.
+"""
+
+import sqlite3
+import logging
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from hermes_state import SessionDB
+
+
+@pytest.fixture()
+def db(tmp_path):
+    """Create a SessionDB with a temp database file."""
+    db_path = tmp_path / "test_state.db"
+    session_db = SessionDB(db_path=db_path)
+    yield session_db
+    try:
+        session_db.close()
+    except Exception:
+        pass
+
+
+class TestTryWalCheckpointPassive:
+    """_try_wal_checkpoint() should use PASSIVE mode for periodic use."""
+
+    def test_checkpoint_uses_passive_mode(self, db):
+        """PASSIVE checkpoint does not require exclusive lock — safe for large DBs."""
+        # Capture the real connection's execute before mocking
+        real_conn = db._conn
+        execute_calls = []
+
+        def tracking_execute(sql, *args, **kwargs):
+            execute_calls.append(sql)
+            return real_conn.execute(sql, *args, **kwargs)
+
+        # sqlite3.Connection.execute is read-only (C extension) — replace _conn
+        mock_conn = MagicMock()
+        mock_conn.execute.side_effect = tracking_execute
+        mock_conn.fetchone.return_value = None
+        db._conn = mock_conn
+
+        db._try_wal_checkpoint()
+
+        passive_calls = [c for c in execute_calls if "wal_checkpoint(PASSIVE)" in c]
+        truncate_calls = [c for c in execute_calls if "wal_checkpoint(TRUNCATE)" in c]
+        assert len(passive_calls) == 1, (
+            f"Expected 1 PASSIVE checkpoint call, got {len(passive_calls)}"
+        )
+        assert len(truncate_calls) == 0, (
+            "Periodic checkpoint should NOT use TRUNCATE"
+        )
+
+    def test_checkpoint_logs_warning_on_failure(self, db, caplog):
+        """Failed PASSIVE checkpoint logs a warning instead of silent pass."""
+        mock_conn = MagicMock()
+        mock_conn.execute.side_effect = sqlite3.OperationalError("disk I/O error")
+        db._conn = mock_conn
+
+        with caplog.at_level(logging.WARNING):
+            db._try_wal_checkpoint()
+
+        assert any("WAL checkpoint (PASSIVE) failed" in r.message for r in caplog.records), (
+            f"Expected warning log about PASSIVE checkpoint failure, got: {caplog.text}"
+        )
+
+    def test_checkpoint_returns_result_on_success(self, db):
+        """Successful PASSIVE checkpoint does not raise."""
+        db._try_wal_checkpoint()
+
+
+class TestCloseUsesTruncate:
+    """close() should still use TRUNCATE to shrink WAL on shutdown."""
+
+    def test_close_uses_truncate_mode(self, db):
+        """TRUNCATE at close is safe — no concurrent writers during shutdown."""
+        real_conn = db._conn
+        execute_calls = []
+
+        def tracking_execute(sql, *args, **kwargs):
+            execute_calls.append(sql)
+            return real_conn.execute(sql, *args, **kwargs)
+
+        mock_conn = MagicMock()
+        mock_conn.execute.side_effect = tracking_execute
+        db._conn = mock_conn
+
+        db.close()
+
+        truncate_calls = [c for c in execute_calls if "wal_checkpoint(TRUNCATE)" in c]
+        assert len(truncate_calls) == 1, (
+            f"Expected 1 TRUNCATE checkpoint at close, got {len(truncate_calls)}"
+        )
+
+    def test_close_logs_debug_on_failure(self, db, caplog):
+        """Failed TRUNCATE at close logs debug (not warning — close is best-effort)."""
+        mock_conn = MagicMock()
+        mock_conn.execute.side_effect = sqlite3.OperationalError("database is locked")
+        db._conn = mock_conn
+
+        with caplog.at_level(logging.DEBUG):
+            db.close()
+
+        assert any("WAL checkpoint (TRUNCATE) at close failed" in r.message for r in caplog.records), (
+            f"Expected debug log about TRUNCATE failure at close, got: {caplog.text}"
+        )
+
+
+class TestCheckpointFrequency:
+    """Checkpoint triggers every N writes."""
+
+    def test_checkpoint_triggers_at_interval(self, db):
+        """_try_wal_checkpoint is called every _CHECKPOINT_EVERY_N_WRITES writes."""
+        call_count = [0]
+        original = db._try_wal_checkpoint
+
+        def counting_checkpoint():
+            call_count[0] += 1
+            original()
+
+        db._try_wal_checkpoint = counting_checkpoint
+
+        # Write exactly _CHECKPOINT_EVERY_N_WRITES sessions to trigger one checkpoint
+        n = db._CHECKPOINT_EVERY_N_WRITES
+        import time as _time
+        for i in range(n):
+            db._execute_write(lambda conn, _i=i: conn.execute(
+                "INSERT INTO sessions (id, source, started_at) VALUES (?, ?, ?)",
+                (f"sess_{_i}", "test", _time.time()),
+            ))
+
+        assert call_count[0] == 1, (
+            f"Expected 1 checkpoint after {n} writes, got {call_count[0]}"
+        )

tui_gateway/server.py

@@ -8463,7 +8463,11 @@ def _(rid, params: dict) -> dict:
 
 @method("prompt.submit")
 def _(rid, params: dict) -> dict:
-    sid, text = params.get("session_id", ""), params.get("text", "")
+    from hermes_cli.input_sanitize import sanitize_user_prompt_text
+
+    sid = params.get("session_id", "")
+    raw_text = params.get("text", "")
+    text = sanitize_user_prompt_text(raw_text) if isinstance(raw_text, str) else raw_text
     truncate_user_ordinal = params.get("truncate_before_user_ordinal")
     session, err = _sess_nowait(params, rid)
     if err:

web/src/lib/api.ts

@@ -2301,6 +2301,8 @@ export interface AuxiliaryModelsResponse {
 export interface MoaModelSlot {
   provider: string;
   model: string;
+  /** Optional per-slot reasoning effort — round-tripped, not edited here. */
+  reasoning_effort?: string;
 }
 
 export interface MoaConfigResponse {
@@ -2312,6 +2314,10 @@ export interface MoaConfigResponse {
     reference_temperature: number;
     aggregator_temperature: number;
     max_tokens: number;
+    /** Optional advisor output cap — round-tripped, not edited here. */
+    reference_max_tokens?: number | null;
+    /** Fan-out cadence (per_iteration | user_turn) — round-tripped. */
+    fanout?: string;
     enabled: boolean;
   }>;
   reference_models: MoaModelSlot[];

website/docs/reference/optional-skills-catalog.md

@@ -57,7 +57,7 @@ hermes skills uninstall <skill-name>
 |-------|-------------|
 | [**baoyu-article-illustrator**](/docs/user-guide/skills/optional/creative/creative-baoyu-article-illustrator) | Article illustrations: type × style × palette consistency. |
 | [**baoyu-comic**](/docs/user-guide/skills/optional/creative/creative-baoyu-comic) | Knowledge comics (知识漫画): educational, biography, tutorial. |
-| [**blender-mcp**](/docs/user-guide/skills/optional/creative/creative-blender-mcp) | Control Blender directly from Hermes via socket connection to the blender-mcp addon. Create 3D objects, materials, animations, and run arbitrary Blender Python (bpy) code. Use when user wants to create or modify anything in Blender. |
+| [**blender-mcp**](/docs/user-guide/skills/optional/creative/creative-blender-mcp) | Drive Blender via the catalog blender MCP, with bpy recipes. |
 | [**concept-diagrams**](/docs/user-guide/skills/optional/creative/creative-concept-diagrams) | Generate flat, minimal light/dark-aware SVG diagrams as standalone HTML files, using a unified educational visual language with 9 semantic color ramps, sentence-case typography, and automatic dark mode. Best suited for educational and no... |
 | [**creative-ideation**](/docs/user-guide/skills/optional/creative/creative-creative-ideation) | Generate ideas via named methods from creative practice. |
 | [**hyperframes**](/docs/user-guide/skills/optional/creative/creative-hyperframes) | Create HTML-based video compositions, animated title cards, social overlays, captioned talking-head videos, audio-reactive visuals, and shader transitions using HyperFrames. HTML is the source of truth for video. Use when the user wants... |

website/docs/user-guide/messaging/slack.md

@@ -217,6 +217,12 @@ hermes gateway install      # Install as a user service
 sudo hermes gateway install --system   # Linux only: boot-time system service
 ```
 
+:::tip Codex reasoning-effort safety
+For Codex-backed Slack peer-agent channels, prefer `agent.reasoning_effort: high` or lower. `xhigh`
+can spend the entire turn in hidden reasoning and never produce visible assistant text; Hermes now
+suppresses those incomplete-turn warnings from the thread and keeps the diagnostics in gateway logs.
+:::
+
 ---
 
 ## Step 9: Invite the Bot to Channels

website/docs/user-guide/multi-profile-gateways.md

@@ -189,6 +189,50 @@ Kanban workers only ever see their own profile's secrets). Kanban,
 profile-scoped skills/memory/SOUL, and model routing all behave per-profile
 exactly as they do with separate gateways.
 
+### Routing shared-bot chats to profiles (`profile_routes`)
+
+Multiplexing selects a profile per **credential** (each profile's own bot
+token) or per **URL prefix** (`/p/<profile>/` for HTTP platforms). When several
+communities share **one** bot token — for example one Discord bot serving many
+guilds — you can additionally route specific guilds/channels/threads to
+different profiles with `gateway.profile_routes`:
+
+```yaml
+gateway:
+  multiplex_profiles: true
+  profile_routes:
+    # An entire Discord server → one profile
+    - name: acme-server
+      platform: discord
+      guild_id: "1234567890"
+      profile: acme
+
+    # One channel in that server → a different profile
+    - name: acme-support
+      platform: discord
+      guild_id: "1234567890"
+      chat_id: "9876543210"
+      profile: acme-support
+
+    # A Telegram group (no guild concept — chat_id only)
+    - name: tg-group
+      platform: telegram
+      chat_id: "-1001234567890"
+      profile: tg-profile
+```
+
+Routes are matched most-specific-first (`thread_id` > `chat_id` > `guild_id`),
+all declared fields must hold (AND), and a route keyed on a channel also
+matches threads/forum posts whose parent is that channel. Messages that match
+no route stay on the default/active profile. The routed profile gets the full
+per-profile isolation described above (config, skills, memory, credentials,
+session namespace). Routing works on every platform adapter, not just Discord.
+
+`profile_routes` requires `gateway.multiplex_profiles: true`; with
+multiplexing off the routes are ignored. If a route names a profile that does
+not exist on disk, the gateway logs a warning naming the profile and source and
+falls back to the default home.
+
 ## Start, stop, or restart all gateways at once
 
 The CLI ships with single-profile lifecycle commands. To act across every

website/docs/user-guide/skills/optional/creative/creative-blender-mcp.md

@@ -1,14 +1,14 @@
 ---
-title: "Blender Mcp — Control Blender directly from Hermes via socket connection to the blender-mcp addon"
+title: "Blender Mcp — Drive Blender via the catalog blender MCP, with bpy recipes"
 sidebar_label: "Blender Mcp"
-description: "Control Blender directly from Hermes via socket connection to the blender-mcp addon"
+description: "Drive Blender via the catalog blender MCP, with bpy recipes"
 ---
 
 {/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
 
 # Blender Mcp
 
-Control Blender directly from Hermes via socket connection to the blender-mcp addon. Create 3D objects, materials, animations, and run arbitrary Blender Python (bpy) code. Use when user wants to create or modify anything in Blender.
+Drive Blender via the catalog blender MCP, with bpy recipes.
 
 ## Skill metadata
 
@@ -16,8 +16,8 @@ Control Blender directly from Hermes via socket connection to the blender-mcp ad
 |---|---|
 | Source | Optional — install with `hermes skills install official/creative/blender-mcp` |
 | Path | `optional-skills/creative/blender-mcp` |
-| Version | `1.0.0` |
-| Author | alireza78a |
+| Version | `2.0.0` |
+| Author | alireza78a + Hermes Agent |
 | Platforms | linux, macos, windows |
 
 ## Reference: full SKILL.md
@@ -26,87 +26,81 @@ Control Blender directly from Hermes via socket connection to the blender-mcp ad
 The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
 :::
 
-# Blender MCP
+# Blender MCP Skill
 
-Control a running Blender instance from Hermes via socket on TCP port 9876.
+Companion skill for the `blender` entry in the Hermes MCP catalog. The MCP
+server provides the connection to Blender; this skill teaches the bpy idioms
+and pitfalls for driving it well. It does not cover Blender UI workflows —
+everything here goes through the MCP tools against a live Blender session.
 
-## Setup (one-time)
+## When to Use
 
-### 1. Install the Blender addon
+Use when the user wants to create or modify anything in a running Blender
+instance: meshes, materials, animations, lighting, renders. Requires the
+blender MCP server installed and a Blender desktop session with the addon
+connected.
 
-    curl -sL https://raw.githubusercontent.com/ahujasid/blender-mcp/main/addon.py -o ~/Desktop/blender_mcp_addon.py
+## Prerequisites
 
-In Blender:
-    Edit > Preferences > Add-ons > Install > select blender_mcp_addon.py
-    Enable "Interface: Blender MCP"
+1. Install the MCP server from the Nous catalog (one-time):
 
-### 2. Start the socket server in Blender
+       hermes mcp install blender
 
-Press N in Blender viewport to open sidebar.
-Find "BlenderMCP" tab and click "Start Server".
+   This configures the pinned `blender-mcp` stdio server with the curated
+   tool set: `get_scene_info`, `get_object_info`, `get_viewport_screenshot`,
+   `execute_blender_code`.
 
-### 3. Verify connection
+2. Install the addon inside Blender (one-time — the catalog entry's
+   post-install notes cover this too):
+   - Download https://raw.githubusercontent.com/ahujasid/blender-mcp/main/addon.py
+   - Blender > Edit > Preferences > Add-ons > Install... > select addon.py,
+     enable "Interface: Blender MCP".
 
-    nc -z -w2 localhost 9876 && echo "OPEN" || echo "CLOSED"
+3. Every session: start Blender FIRST, press N in the viewport, open the
+   "BlenderMCP" tab, click "Connect to Claude" (starts the local bridge
+   socket). Then start your Hermes session so the MCP tools are loaded.
 
-## Protocol
+   The addon refuses to start under `blender -b` (background mode). On a
+   machine without a display, run Blender under a virtual one:
+   `xvfb-run blender`. GPU rendering works fine under Xvfb.
 
-Plain UTF-8 JSON over TCP -- no length prefix.
+## Quick Reference
 
-Send:     {"type": "<command>", "params": {<kwargs>}}
-Receive:  {"status": "success", "result": <value>}
-          {"status": "error",   "message": "<reason>"}
+| MCP tool                  | Use for                                    |
+|---------------------------|--------------------------------------------|
+| `get_scene_info`          | List objects before touching the scene     |
+| `get_object_info`         | Inspect one object (transform, materials)  |
+| `get_viewport_screenshot` | Visual check of what you built             |
+| `execute_blender_code`    | Everything else — arbitrary bpy Python     |
 
-## Available Commands
+Optional asset-service tools (PolyHaven, Sketchfab, Hyper3D, Hunyuan3D) are
+disabled by default. If the user has enabled a service in the addon panel,
+opt into its tools with `hermes mcp configure blender`.
 
-| type                    | params            | description                     |
-|-------------------------|-------------------|---------------------------------|
-| execute_code            | code (str)        | Run arbitrary bpy Python code   |
-| get_scene_info          | (none)            | List all objects in scene       |
-| get_object_info         | object_name (str) | Details on a specific object    |
-| get_viewport_screenshot | (none)            | Screenshot of current viewport  |
+## Procedure
 
-## Python Helper
+1. Call `get_scene_info` first — never assume the scene is empty.
+2. Build with `execute_blender_code`, in small focused calls (one logical
+   step per call: add objects, then materials, then animation). Large
+   monolithic scripts hit the bridge timeout.
+3. Verify visually with `get_viewport_screenshot` between major steps.
+4. Render to an absolute path and tell the user where the file is.
 
-Use this inside execute_code tool calls:
+### Common bpy Patterns
 
-    import socket, json
+Clear scene:
 
-    def blender_exec(code: str, host="localhost", port=9876, timeout=15):
-        s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
-        s.connect((host, port))
-        s.settimeout(timeout)
-        payload = json.dumps({"type": "execute_code", "params": {"code": code}})
-        s.sendall(payload.encode("utf-8"))
-        buf = b""
-        while True:
-            try:
-                chunk = s.recv(4096)
-                if not chunk:
-                    break
-                buf += chunk
-                try:
-                    json.loads(buf.decode("utf-8"))
-                    break
-                except json.JSONDecodeError:
-                    continue
-            except socket.timeout:
-                break
-        s.close()
-        return json.loads(buf.decode("utf-8"))
-
-## Common bpy Patterns
-
-### Clear scene
     bpy.ops.object.select_all(action='SELECT')
     bpy.ops.object.delete()
 
-### Add mesh objects
+Add mesh objects:
+
     bpy.ops.mesh.primitive_uv_sphere_add(radius=1, location=(0, 0, 0))
     bpy.ops.mesh.primitive_cube_add(size=2, location=(3, 0, 0))
     bpy.ops.mesh.primitive_cylinder_add(radius=0.5, depth=2, location=(-3, 0, 0))
 
-### Create and assign material
+Create and assign material:
+
     mat = bpy.data.materials.new(name="MyMat")
     mat.use_nodes = True
     bsdf = mat.node_tree.nodes.get("Principled BSDF")
@@ -115,21 +109,43 @@ Use this inside execute_code tool calls:
     bsdf.inputs["Metallic"].default_value = 0.0
     obj.data.materials.append(mat)
 
-### Keyframe animation
+Keyframe animation:
+
     obj.location = (0, 0, 0)
     obj.keyframe_insert(data_path="location", frame=1)
     obj.location = (0, 0, 3)
     obj.keyframe_insert(data_path="location", frame=60)
 
-### Render to file
+Render to file:
+
     bpy.context.scene.render.filepath = "/tmp/render.png"
     bpy.context.scene.render.engine = 'CYCLES'
     bpy.ops.render.render(write_still=True)
 
 ## Pitfalls
 
-- Must check socket is open before running (nc -z localhost 9876)
-- Addon server must be started inside Blender each session (N-panel > BlenderMCP > Connect)
-- Break complex scenes into multiple smaller execute_code calls to avoid timeouts
-- Render output path must be absolute (/tmp/...) not relative
-- shade_smooth() requires object to be selected and in object mode
+- The blender MCP tools only exist if the server is installed and the session
+  started after install. If they're missing, run `hermes mcp install blender`
+  and start a new session.
+- The addon bridge must be (re)connected inside Blender each Blender session
+  (N-panel > BlenderMCP > Connect). "Connection refused" from the tools means
+  Blender isn't running or the addon isn't connected — fix that, don't retry.
+- Break complex scenes into multiple smaller `execute_blender_code` calls to
+  avoid bridge timeouts.
+- Render output paths must be absolute (`/tmp/render.png`), not relative —
+  they resolve on the BLENDER host's filesystem, which matters if Hermes and
+  Blender run on different machines.
+- `shade_smooth()` requires the object to be selected and in object mode.
+- `execute_blender_code` runs arbitrary Python inside Blender with no sandbox
+  — same trust level as the `terminal` tool. Don't paste untrusted code into
+  it.
+- Do NOT hand-roll raw TCP JSON to port 9876 from `execute_code` — that was
+  this skill's pre-MCP workaround. It bypasses the catalog's version pinning
+  and tool curation. The MCP tools are the supported path.
+
+## Verification
+
+- `get_scene_info` returns the expected object list after each build step.
+- `get_viewport_screenshot` shows the scene you intended.
+- After a render, confirm the output file exists and report its absolute
+  path to the user.

website/docusaurus.config.ts

@@ -62,6 +62,14 @@ const config: Config = {
           /^user-guide\/skills\/bundled\//,
           /^user-guide\/skills\/optional\//,
         ],
+        // Exact-or-prefix matching only (default is edit distance 1).
+        // With fuzzy distance 1, "keet" matched "meetings"/"keep" (one
+        // edit away after stemming), and multi-word typo queries against
+        // our ~14 MB index could stall the single-threaded search worker
+        // for 25s+, backing up every subsequent keystroke's search until
+        // the bar appeared dead. Distance 0 keeps "word or its extension"
+        // semantics (keet -> keet*) and removes the pathological scans.
+        fuzzyMatchingDistance: 0,
       }),
     ],
   ],