Two related hardening fixes for auxiliary calls (which include MoA reference
advisors — a pinned-model path where provider fallback is not a meaningful
recovery):
1. Transient-transport retries: the same-provider retry on a connection reset /
timeout / 5xx / 408 was a single attempt, then fallback. For a pinned aux
call a second blip silently loses the call (root of the run2 double-advisor
'Connection error' collapse — a genuine upstream blip). Now retries N times
with exponential backoff, N = auxiliary.transient_retries (default 2 -> 3
total attempts, clamped [0,6]). Compression-on-timeout fast-fail carve-out
preserved.
2. Per-model client-cache isolation: _client_cache_key excluded the model, so
two concurrent auxiliary calls to the same provider/base_url/key but
different models (e.g. an opus + gpt-5.5 MoA fan-out) shared one cache entry
and could race each other's client lifecycle. Model now participates in the
key -> distinct clients, no cross-call races. Same-model reuse unchanged.
- agent/auxiliary_client.py: _transient_retry_count() + backoff loop; model in
_client_cache_key and both call sites.
- hermes_cli/config.py: auxiliary.transient_retries default (2).
- tests: new retry/isolation tests; updated 2 stale-expectation tests to the
corrected behavior (per-model resolve; N-retry escalation).
Backoff base is overridable (_TRANSIENT_RETRY_BACKOFF_BASE) so tests don't sleep.
Attribution audit gate: the salvaged contributor commits carry
kiljadn@gmail.com (Nick Mason / @designnotdrum). Add the mapping so
contributor_audit.py resolves the author on this PR.
Follow-up on the #56480 salvage: the include_registry=False docstring said
None is returned only for registry/MCP-only toolsets; it also applies to
registry-derived aliases, which have no static TOOLSETS counterpart.
_get_platform_tools reverse-maps a platform composite to configurable
toolsets with an all-tools subset test. Because get_toolset() merges
registry-registered tools into a toolset, a tool added to a toolset
(delegate_cli -> delegation; desktop-only read_terminal -> terminal) that the
static composite never listed made the subset test fail, silently dropping the
entire toolset on api_server and other inference-based platforms. Compare the
toolset's static membership at all three reverse-map sites.
Fixes#49622.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds include_registry=True kwarg to resolve_toolset/get_toolset. When False,
returns only the static TOOLSETS view with no registry-merged tools — the
composite-authored membership platform reverse-mapping must compare against.
Default True preserves all existing behavior; this is the enabling half of
the api_server toolset-drop fix (#49622).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Extends the browser private-network eval guard to the Camofox backend.
On main, _browser_eval() returned early in Camofox mode before running the
shared private-URL literal pre-scan and before re-checking the page URL
after eval, leaving Camofox as a sibling backend that could execute
browser_console(expression=...) against private/internal targets.
- move the eval private-URL literal pre-scan before the Camofox early return
- add a Camofox current-page private-URL probe via the evaluate endpoint
- withhold Camofox eval results when the page is now private/internal
Follow-up to browser private-network hardening in #56173, #56526, #56664.
Salvage of #56764 by @rayjun (rayoo), cherry-picked to preserve authorship.
The fake _SessionStore tracked peer_records but no test read it, leaving
#55300's peer-record behavior unasserted. Add a positive assertion on the
persist path and negative (== []) assertions on the two stale/moved-binding
skip paths, so the peer-record side effect is bound.
Mutation-verified: removing the production _record_gateway_session_peer call
makes the positive assertion fail.
Co-authored-by: João Vitor Cunha <jvsantos.cunha@gmail.com>
The #55300 peer-recording call now fires on the failed-turn compression
split path; the fake _SessionStore in test_compression_failure_session_sync
(carried in with #55721's test changes) lacked that method. Add a
call-tracking no-op so the combined salvage's tests pass.
Co-authored-by: João Vitor Cunha <jvsantos.cunha@gmail.com>
Manual /compress built a temporary AIAgent without the originating
platform / stable gateway session key, so an external context engine
ingested the retained transcript tail as source=cli during /compress
and again as the real platform on resume (duplicate cli,telegram rows).
Pass platform=_platform_config_key(source.platform) + the in-scope
gateway_session_key, mirroring the normal gateway turn. Assigned into
runtime_kwargs (single-valued, authoritative) so they neither collide
into a duplicate-kwarg TypeError nor lose to a stale resolver value.
Fixes#50422.
Map the two contributor emails whose commits are cherry-picked into the
compression-routing-integrity salvage so scripts/contributor_audit.py
attributes them at release time:
- jvsantos.cunha@gmail.com -> plcunha (PR #55300)
- jakepresent1@gmail.com -> jakepresent (PR #55721)
r266-tech (PR #50517) is already mapped.
MoA per-turn latency is dominated by advisor GENERATION: turn wall time
correlates ~0.88 with output tokens and ~-0.03 with input tokens (measured over
52 turns). Each turn waits for the slowest advisor to finish writing, and
advisors were uncapped — writing multi-thousand-token essays the aggregator
only needs the gist of.
Add an opt-in per-preset reference_max_tokens knob (mirrors reference_temperature)
that caps ADVISOR output only; the acting aggregator is never capped. Default
None = uncapped, so existing presets are byte-for-byte unchanged (no regression).
Wired through both MoA execution paths (MoAChatCompletions.create and
aggregate_moa_context).
E2E: same task, closed preset uncapped vs reference_max_tokens=600 -> 59s to 33s
(~44% faster), final answer identical/correct.
- hermes_cli/moa_config.py: _coerce_int_or_none helper + reference_max_tokens
in _normalize_preset/_default_preset/flattened view
- agent/moa_loop.py: read preset.reference_max_tokens, pass to reference fan-out
- agent/conversation_loop.py: pass reference_max_tokens on the per-turn path
- tests + docs
Ben caught that the initial approach (widening _NOUS_PORTAL_ALLOWED_HOSTS to
include the staging host) was the wrong fix -- env vars are supposed to
override the allowlist, mirroring how NOUS_INFERENCE_BASE_URL already
bypasses _ALLOWED_NOUS_INFERENCE_HOSTS via _nous_inference_env_override().
The actual bug: both resolve_nous_access_token and
resolve_nous_runtime_credentials read
`_optional_base_url(state.get("portal_base_url")) or os.getenv(...) or ...`
-- a plain `or` chain where the STORED state value wins first (short-circuits
before the env vars are even read), and then whichever value won gets run
through the same _NOUS_PORTAL_ALLOWED_HOSTS gate regardless of its source.
So a hosted agent stamped with HERMES_PORTAL_BASE_URL=<staging> in its env
AND a staging portal_base_url already persisted to auth.json would still
get silently rewritten to prod on every refresh, because the env var never
even got a chance to be consulted.
Revert the previous _NOUS_PORTAL_ALLOWED_HOSTS widening entirely --
staying prod-only preserves the allowlist's actual job (rejecting an
untrusted network-provided portal_base_url persisted to auth.json by a
compromised Portal response).
Add _nous_portal_env_override() (mirrors _nous_inference_env_override())
and restructure both call sites so the env override is checked FIRST and,
when set, wins outright and skips the allowlist gate entirely -- the
allowlist only ever runs against the fallback (stored-state-or-default)
path now.
Rewrote tests/hermes_cli/test_nous_portal_staging_allowlist.py to test the
actual fix: the helper function, and an end-to-end
resolve_nous_access_token proof that the env override wins even when state
ALSO has the staging host stored (the exact incident shape), that it wins
over a stored PROD host too, and that the allowlist's heal-to-prod
behaviour for an untrusted stored value is preserved when no override is
set.
* fix(streaming): handle completed responses with empty/None choices
The streaming fallback guard added in #55932 recognized a completed
response object only when its `choices` was a non-empty list. But an
adapter can return a completed response whose `choices` is `None` or an
empty list (an error / content-filter / terminal frame) — still a whole,
non-iterable response, not a token stream. Those shapes fell through to
`for chunk in stream` and crashed with
'types.SimpleNamespace' object is not iterable
which is exactly issue #55933 (MoA `openai-codex` aggregator on
TUI/Desktop, where a stream consumer forces the streaming path).
Broaden the guard to discriminate on the PRESENCE of a `choices`
attribute (a genuine provider Stream object exposes none), disable
streaming for the session, and return the completed object so the outer
loop's normal invalid-response validation handles empty/None choices via
its retry path instead of iterating.
Based on the diagnosis in #56525 by @spiky02plateau (that PR normalized
the MoA aggregator return with a one-shot chunk iterator; the common
text/tool-call crash was already fixed at this seam by #55932, so this
extends the existing guard to cover only the remaining empty/None-choices
gap).
Fixes#55933
* refactor(streaming): simplify empty-choices guard body and parametrize tests
Post-review cleanup (no behavior change):
- Inline the single-use `response_choices` local and drop the redundant
`if first_choice is not None else None` guard (getattr(None, ...) already
returns the default safely).
- Collapse the two near-identical empty/None-choices regression tests into
one `@pytest.mark.parametrize` case.
Mutation-verified: reverting the guard to the old non-empty-list condition
still makes both parametrized cases fail with the historical
'types.SimpleNamespace' object is not iterable.
---------
Co-authored-by: spiky02plateau <155588579+spiky02plateau@users.noreply.github.com>
Follow-up to the salvaged fix: the regression test asserted a frozen
max_tokens == 128_000 literal, coupling it to the Opus-4-8 model table.
Assert against _get_anthropic_max_output("claude-opus-4-8") plus > 2000
instead, so the test survives model-table churn while still catching a
regression to the old `or 2000` fallback.
agent/vertex_adapter.py resolved VERTEX_CREDENTIALS_PATH,
GOOGLE_APPLICATION_CREDENTIALS, VERTEX_PROJECT_ID, and VERTEX_REGION via raw
os.environ.get() instead of the profile-scoped get_secret() every other
credential lookup in hermes_cli/runtime_provider.py uses. In a multiplex
gateway serving several profiles from one process, os.environ still holds
whichever profile's .env python-dotenv loaded at boot — so a raw read here
let one profile's turn silently mint a Vertex OAuth2 token from, and get
billed against, a different profile's GCP service account. No error, no
fail-closed guard: the multiplex UnscopedSecretError protection was bypassed
entirely because these reads never went through get_secret().
- _resolve_credentials_path/_resolve_project_override/_resolve_region now
call agent.secret_scope.get_secret(), matching the _getenv() pattern
already used for every other provider's credentials.
- get_vertex_credentials()'s ADC fallback (google.auth.default()) reads
GOOGLE_APPLICATION_CREDENTIALS from os.environ internally, bypassing
get_secret() entirely — closed with a narrow guard: when multiplexing is
active and this profile's scope has no Vertex credentials of its own, but
os.environ still carries a value (left by a different profile's boot-time
dotenv load), refuse ADC rather than silently authenticate as a stranger.
- Zero behavior change for single-profile installs: get_secret() falls
through to os.environ transparently whenever multiplexing is off.
Same bug class as the already-fixed _HERMES_OAUTH_FILE/_AUTH_JSON_PATH/
HOOKS_DIR cross-profile leaks, now closed for Vertex's OAuth2 credential
path.
The salvaged fix wired per-provider ssl_ca_cert / ssl_verify (and
HERMES_CA_BUNDLE) into the MAIN OpenAI client. This follow-up:
- Auxiliary client parity: process_bootstrap.build_keepalive_http_client
accepts and forwards verify; auxiliary_client._resolve_aux_verify mirrors
the main-client TLS resolution (via load_config_readonly, the read-only
fast path) so compression/vision/web_extract/title-gen/session_search
honor the same per-provider CA. Without this, chat worked against a
private-CA endpoint but every auxiliary call still failed APIConnectionError.
- switch_model now reads custom_providers from live config (load_config_readonly)
instead of the init-time agent._custom_providers snapshot, so ssl_ca_cert /
ssl_verify edits are honored on mid-session model switch — matching the
context-length reload (#15779).
- Drop the dead client-level verify= where a custom httpx transport is used
(httpx ignores it there); verify lives on the transport. Fix docstrings.
Applies to both run_agent._build_keepalive_http_client and process_bootstrap.
- resolve_httpx_verify: add CURL_CA_BUNDLE to the env chain (consistency with
agent/ssl_guard._CA_BUNDLE_ENV_VARS) and emit a loud logger.warning naming
the endpoint whenever ssl_verify:false disables verification.
- get_custom_provider_tls_settings: case-insensitive base_url match (config
dedup already lowercases; scheme/host are case-insensitive) so a mixed-case
entry doesn't silently drop its CA. Exact match preserved — no prefix bypass.
- Demote best-effort except Exception: pass in agent_init/switch_model to
logger.debug(exc_info=True).
- Tests for aux verify forwarding, _resolve_aux_verify, case-insensitive
match, and prefix-bypass rejection.
Wire ssl_ca_cert and ssl_verify through custom_providers config and env
vars into the keepalive httpx client, fixing APIConnectionError against
mkcert/self-signed Ollama proxies behind HTTPS.
Desktop chat bubbles render plain text, but a worker-routed command that
builds its own Rich Console (e.g. /journey) picks up truecolor from the
gateway's inherited COLORTERM and leaks raw escapes into the bubble. Strip
ANSI at the single worker-return choke point so every command renders cleanly.
The TUI opens /journey as an overlay, so it never travels this path.
In the interactive CLI, /journey dispatched straight to `args.func(args)`,
letting Rich write ANSI to stdout — which patch_stdout's StdoutProxy passes
through as literal `?[38;2;…m` garbage. Route the read-only views (default +
`list`) through a captured, force-color Console and re-emit via `_cprint`
(prompt_toolkit's ANSI parser), matching the `ChatConsole` idiom.
`delete`/`edit` stay on real stdio since they prompt / open `$EDITOR`.
parse_frontmatter's malformed-YAML fallback stores every value as a string,
so a skill's `metadata` can be a str. `_category`/`_related` chained
`.get("metadata", {}).get("hermes", {})` and blew up with `'str' object has
no attribute 'get'`, taking down `build_learning_graph()` (and thus /journey
and `hermes journey`) whenever any installed skill had bad frontmatter.
Extract a `_hermes_meta()` helper that returns the nested dict only when it
really is one. Fixes the whole class, not just the two call sites.
- claude-fable-5 placed above claude-opus-4.8 in both curated lists
- claude-sonnet-5 replaces claude-sonnet-4.6
- sakana/fugu-ultra added near the bottom (before routers/free tier)
- regenerated website/static/api/model-catalog.json via scripts/build_model_catalog.py (live-pulled by CLI, published on merge — no release needed)
Two Hermes bots sharing a channel could volley replies at each other
indefinitely. Root cause: Discord reply-pings (allowed_mentions
replied_user=true) add the replied-to bot to message.mentions without a
literal <@bot> token in the body, so the existing bot-admission gate
treated a reply chip as an explicit @mention and re-triggered the peer.
Adds opt-in discord.bots_require_inline_mention (default false; env
DISCORD_BOTS_REQUIRE_INLINE_MENTION). When enabled, bot-authored
messages must carry a raw inline <@id>/<@!id> mention in the content;
reply-ping-only mentions no longer admit the message. Human messages and
all existing defaults are unchanged.
The new _self_is_raw_mentioned helper deliberately ignores the resolved
message.mentions list (which reply-ping populates) and checks only the
raw content token via the shared _raw_mentioned_user_ids primitive.
Vertex AI authenticates via OAuth2 (service-account JSON path / ADC), not
PROVIDER_REGISTRY, and VERTEX_CREDENTIALS_PATH is declared with
password=False (it's a path, not a bare key) under category="provider" —
a category the registry-derived blocklist loop never checks. Both it and
GOOGLE_APPLICATION_CREDENTIALS (the ADC fallback the adapter also reads)
fell through every existing blocklist source and leaked the on-disk
location of a GCP service-account key into every spawned subprocess
(terminal, codex/copilot app-server, browser workers) — the same leak
class already closed for every other provider's credentials in #53503.
Follow-up review fixes on the salvage of #54872 (原作者 张满良/@zmlgit):
1. [HIGH] Adapter selection now goes through the shared
_authorization_adapter chokepoint (gateway/authz_mixin.py) instead of a
local inline lookup that fell back to the DEFAULT profile's same-platform
adapter when the owning profile had a registry entry but no adapter for
that platform. That fallback re-introduced the exact cross-profile
mis-delivery ([230002] Bot can NOT be out of the chat) this change exists
to fix. Adds a mutation-verified guard test
(test_notifier_owning_profile_adapter_no_default_fallback).
2. [HIGH→documented] The creator-wake SessionSource cannot faithfully
reconstruct a DM/thread creator's session key because chat_type is neither
persisted on the subscription nor carried on the session-context bridge.
Documented the limitation inline; behavior degrades to a fresh group
session (never an exception). The end-to-end fix (stamp + persist
chat_type) is a scoped follow-up, not bundled into this salvage.
3. [MED] Documented that archived/unblocked are intentionally claimed (cursor
hygiene) but silent, and excluded from wake kinds.
4. [MED] Wake-injection failure now logs at WARNING with exc_info=True (the
cursor has already advanced, so a broken wake must not be a silent no-op).
Addresses @tonydwb's review on PR #54872 (12:05 UTC, 2026-06-29):
> the hardcoded Chinese text in the wake messages (lines 118-128 of
> the diff) should be replaced with English or internationalized.
> The rest of the codebase uses English for user-facing messages,
> and hardcoded Chinese will confuse non-Chinese users. Consider
> using a constants dict or the existing i18n infrastructure.
Used the existing i18n infrastructure (agent/i18n.py::t()) — the same
surface gateway/run.py and slash_commands.py already use for static
user-facing strings.
## Changes
- gateway/kanban_watchers.py: import `t` from agent.i18n; replace the
hardcoded Chinese strings in the synthetic wake-up message with
t("gateway.kanban.wake.*") lookups. Behavior unchanged for zh users
(zh catalog preserves the original Chinese phrasing).
- locales/en.yaml: new `gateway.kanban.wake.*` baseline keys (English):
completed / gave_up / crashed / timed_out / blocked / status_default
/ status_joiner / message (with {task_id} {status} {title}
{assignee} {board} placeholders).
- locales/zh.yaml: Chinese translation of the new keys, preserving the
exact wording the original code used (so existing zh users see no
visible change).
- locales/{zh-hant,ja,de,es,fr,tr,uk,af,ko,it,ga,pt,ru,hu}.yaml: added
the same key set with English fallback values. The i18n invariant
test (tests/agent/test_i18n.py::test_catalog_keys_match_english)
requires every catalog to carry the same key set as en.yaml; native
translations can land incrementally without breaking users (the
loader falls back to en.yaml per-key when a translation is missing,
but the key must still exist).
## Verification
- scripts/run_tests.sh tests/agent/test_i18n.py
tests/gateway/test_kanban_watchers_mixin.py
tests/gateway/test_kanban_notifier.py
tests/gateway/test_kanban_notifier_watcher_dispatch_gate.py
→ 60 passed, 0 failed (i18n catalog parity + placeholders parity +
existing kanban notifier behavior).
- Manual: with HERMES_LANGUAGE=en, t("gateway.kanban.wake.completed")
returns "completed"; with HERMES_LANGUAGE=zh, returns "已完成";
with HERMES_LANGUAGE=ja (translation pending), falls back to
"completed" per-key.
Three connected changes that fix kanban notifications in multiplex_profile
gateways and enable event-driven agent collaboration:
1. Session profile propagation
- Add HERMES_SESSION_PROFILE ContextVar (session_context.py)
- Gateway stamps source.profile at dispatch time (run.py)
- _maybe_auto_subscribe reads profile from ContextVar instead of
os.environ which is unset in the gateway main process (kanban_tools.py)
2. Notifier profile-aware routing (kanban_watchers.py)
- Adapter selection: prefer _profile_adapters[sub.notifier_profile]
so each profile's bot delivers its own task notifications
- Relax profile skip-filter: process cross-profile subscriptions when
the gateway has an adapter for the owning profile
- Extend TERMINAL_KINDS with status/archived/unblocked
3. Creator agent wakeup on terminal events (kanban_watchers.py)
- After delivering completed/blocked/gave_up/crashed/timed_out
notifications, inject a synthetic MessageEvent into the creator's
session via adapter.handle_message to trigger their agent loop
- SessionSource built from subscription metadata — no session_store
lookup needed
Self-review follow-up (hermes-pr-review Phase 2, non-blocking clarity findings).
- Collapse the reapplying_enable predicate to a single chained comparison
(new_value == current == "codex_app_server") instead of a two-clause AND
that re-tested new_value == current.
- Dedent the msg_lines list literals (drop trailing single-element commas).
No behavior change: reapply still falls through to the idempotent migrate()
while skipping set_runtime/persist (prompt cache preserved), and the auto-disable
early-return is unchanged. 31/31 tests green.
The /codex-runtime slash command short-circuits with "openai_runtime
already set" when invoked with the same value as the current config,
and crucially skips the entire migration block below. The check
conflates two things: (a) "the config value is correct" and (b) "the
world state (managed block in ~/.codex/config.toml, hermes-tools MCP
callback, plugin discovery) is converged".
Common footgun this exposes: a user who pre-sets
`model.openai_runtime: codex_app_server` directly in config.yaml
(reasonable thing to do) and then runs /codex-runtime codex_app_server
to trigger migration sees "already set" and silently gets no migration.
~/.codex/config.toml never receives the managed block, the hermes-tools
MCP callback never registers, and codex falls through to its default
runtime instead of the app-server one — visibly successful but
functionally partial setup.
The migration is idempotent by design (it replaces its own managed
block in place between MIGRATION_MARKER and MIGRATION_END_MARKER), so
re-running it is safe and cheap. Fix the short-circuit to fall through
to migration when re-applying codex_app_server while skipping the
config persist (no value-level change needed). The disable case
(re-applying "auto") still short-circuits because disabling doesn't
touch ~/.codex/config.toml at all.
The user-visible message changes to "openai_runtime already set to
codex_app_server — re-applying migration" so re-runs surface what
happened.
Regression test (test_reapply_codex_app_server_runs_migration) asserts:
- migrate() was called when re-applying
- persist_callback was NOT called (no config write on no-op transitions)
- migration output (MCP servers, sandbox default) surfaces in the
user-visible message
- requires_new_session is True so callers know to /reset
Verified RED→GREEN: the test fails on origin/main with
"migration must run on reapply, not just first enable" and passes with
this fix. Full test_codex_runtime_switch.py suite: 31 passed.
Self-review follow-up on the salvaged approval-routing fix.
The initial adaptation re-read os.getenv("HERMES_YOLO_MODE") at session-build
time. That diverges from the repo's security invariant: HERMES_YOLO_MODE is
frozen into tools.approval._YOLO_MODE_FROZEN at import time precisely so a skill
running mid-process cannot set the env var and instantly flip the approval
bypass (a prompt-injection escalation path). A live re-read re-opened that hole
for the codex routing path.
- Add tools.approval.is_approval_bypass_active() — the canonical three-source
bypass check (frozen --yolo/HERMES_YOLO_MODE + session /yolo + approvals.mode
off) in one place. This is the 4th inline copy of that OR-chain (the three
sites in approval.py and tui_gateway/server.py:3121 all use the same idiom);
the helper is the shared chokepoint they can collapse onto.
- codex_runtime.py now calls is_approval_bypass_active() instead of the
hand-rolled mode-or-session check plus a runtime env re-read.
- Update the env-yolo test to patch _YOLO_MODE_FROZEN (the canonical test
pattern, e.g. tests/tools/test_yolo_mode.py) rather than setenv, which is
dead-on-arrival against the frozen constant.
Fail-closed default preserved on every branch; 28 integration + 77 session/yolo
tests pass; E2E confirms the real exec decision flips decline->accept only when
bypass is active.
On gateway/cron/non-CLI contexts the codex app-server runtime has no UI to
surface codex's exec/apply_patch approval requests, so they fail closed
(silently decline) — the bot appears responsive but cannot write files, with
no approval prompt anywhere ("patch rejected by user").
When the user has explicitly opted out of Hermes approvals (approvals.mode: off,
the /yolo session toggle, or HERMES_YOLO_MODE=1), collapse to codex's own
sandbox permission profile (~/.codex/config.toml) as the policy gate by passing
_ServerRequestRouting(auto_approve_exec=True, auto_approve_apply_patch=True) to
the session. Defaults (manual/smart/unset) preserve the current fail-closed
behavior — a no-op for users who have not opted out.
Reads the mode via the canonical tools.approval._get_approval_mode() (which
already normalizes the YAML-1.1 bare-'off'->False case) at session-build time,
so a mid-session /yolo toggle is honored too.
5 integration tests: each opt-out mechanism (config off, YAML False, env var,
session yolo) plus the default fail-closed regression guard.
Closes#26530
Co-authored-by: snav <jake@nousresearch.com>
Every other content-returning browser tool entry point
(browser_snapshot/vision/console/eval, and click/type/press via
_blocked_private_page_action) re-checks window.location.href against the
private/internal/cloud-metadata floor after the page could have changed --
because a redirect chain or client-side navigation can land on an address
the initial browser_navigate preflight never saw. browser_back was the one
navigation-triggering entry point missing this: it called
_run_browser_command(..., "back", []) and returned the resulting URL
straight to the model with no re-check.
On a cloud/CDP (non-local) backend, if browser history contains a
private/internal address (e.g. a prior redirect touched an internal host),
browser_back would navigate the live browser there and hand the URL back
to the model with no guard -- the exact class of gap the private-page
guard exists to close, just on the one entry point it hadn't reached yet.
Re-check happens after the navigation succeeds (not before, unlike
click/type/press) since it's the resulting page -- not the one being left
-- whose safety matters. A failed back navigation (no history) skips the
check entirely since nothing changed. Verified live: the new regression
test fails (returns the private URL instead of a blocked payload) on the
pre-fix code and passes after.
The MoA preset section in the composer model dropdown presented presets like
persistent model selections, but selecting one dispatched the one-shot `/moa`
command (command.dispatch name=moa) — it ran a single turn through MoA and then
silently reverted to the prior model. The user saw MoA context for one message,
then it vanished with no indication.
Route MoA preset selection through the same persistent path real provider
selections use: onSelectModel({ model: preset, provider: 'moa' }) →
config.set model="<preset> --provider moa" → the gateway's switch_model. The
check mark now reflects the real current selection (currentProvider === 'moa'
&& currentModel === preset) instead of transient local state, and the
now-unused activeMoaPreset state is removed.
Tests: new model-menu-panel.test.tsx (2) — selecting a preset calls
onSelectModel with provider 'moa' (persistent), and the check renders on the
active preset. tsc -b clean.
With the default busy_input_mode=interrupt, a burst of rapid gateway
messages arriving while context compression is in flight could interrupt
the current turn and start a fresh turn against the pre-rotation parent
session. Because compression is interrupt-immune (#23975), the still-
running compression later rotates the id out from under that new turn,
and if the new turn also grew past the compression threshold it started
its own uncancellable compression on the same stale parent — forking
multiple orphaned one-shot sibling continuations (#56391).
While a state.db compression lock is held for the session, demote
'interrupt' busy-input mode to 'queue' semantics (mirroring the subagent
protection in #30170), so the follow-up message waits for the in-flight
compression + its id rotation to land instead of racing a new turn
against the stale parent. Ack copy explains the compression demotion.
Fixes#56391.
prompt.submit is fire-and-forget — turn completion is signaled by stream /
message.complete events, not the RPC return — but it inherited the generic 30s
default RPC timeout. A turn that legitimately takes >30s to ACK (MoA presets
running references + aggregator in series, deep reasoning, large tool chains)
popped a false 'request timed out: prompt.submit' toast at 30s while the turn
was still running and streamed its real answer in 60-120s later (#55024).
Add PROMPT_SUBMIT_REQUEST_TIMEOUT_MS (1_800_000 = the backend's
agent.gateway_timeout ceiling) and pass it on all four prompt.submit call sites
(submit, resume-recovery retry, regenerate, rewind), mirroring the existing
SESSION_LIST_REQUEST_TIMEOUT_MS opt-out precedent. Widen the GatewayRequest
type (+ the inline requestGateway prop type) to carry the optional timeoutMs the
runtime impl already accepts.
Tests: use-prompt-actions/index.test.tsx 34/34 pass; tsc -b clean.
Two independent MoA auxiliary-call fixes:
#53866 — auxiliary.moa_reference.timeout and auxiliary.moa_aggregator.timeout
were 600s while moa_agent was 120s. Raise both to 900s so a genuinely long
reference/aggregator turn (mixed providers, deep reasoning, long tool chains)
has headroom instead of being cut mid-generation.
#53735 — _CodexCompletionsAdapter (the Codex/Responses auxiliary path used by
the MoA acting-aggregator, compression, web_extract, session_search, etc.)
never set prompt_cache_key, so it stayed cache-cold while the MAIN Responses
transport (agent/transports/codex.py) was warm. Derive the same
content-addressed key via the shared _content_cache_key(instructions, tools)
helper and set it on the aux Responses request, with the same host guards the
main transport uses (xAI carries the key in extra_body; GitHub/Copilot opts out
of cache-key routing).
Tests: 5 new prompt_cache_key cases (set+prefixed, stable across identical
prefix, differs on different instructions, skipped for xai/github hosts).
tests/agent/test_auxiliary_client.py 279 pass; tests/hermes_cli/test_config.py
130 pass.
On the MoA path agent.model/provider are the virtual preset name (e.g.
"closed") and "moa", which have no pricing entry. estimate_usage_cost()
returned None for the aggregator turn, so the `if amount_usd is not None`
guard skipped it and the session's estimated_cost_usd reflected only the
advisor fan-out — a ~50% undercount when the aggregator does the full acting
loop (verified: $0.91 advisor-only vs $1.96 true, aggregator = 54%).
MoAChatCompletions.create() now stashes the resolved aggregator slot as
last_aggregator_slot (exposed via MoAClient); conversation_loop reads it to
price the aggregator turn at its real model/provider. cost_source flips from
'none' to 'provider_models_api'.
Follow-up to the END-MARKER reorder: moving the summary prefix after the
[PRIOR CONTEXT] wrapper meant _is_context_summary_content (prefix-at-start)
no longer recognized a merged-tail summary. That silently broke three
consumers — the last-real-user anchor (would pick the merged summary as a
real user turn, causing active-task loss), the carry-forward summary find,
and the auto-focus skip. _strip_summary_prefix would also carry the wrapper
+ stale tail content forward as the next summary body.
Extract the two delimiter strings into _MERGED_PRIOR_CONTEXT_HEADER /
_MERGED_SUMMARY_DELIMITER constants (writer + detector stay in sync), teach
_is_context_summary_content and _strip_summary_prefix to look past the
delimiter, and add a regression test. Standalone summaries unchanged.
When the compression summary is merged into the first tail message
(the alternation corner case where a standalone summary role would
collide with both head and tail), the old format was
SUMMARY + END_MARKER + OLD_TAIL_CONTENT — so the preserved tail content
appeared AFTER the end marker and the model could read it as a fresh
message to respond to.
Reorder so the END MARKER is always last: old tail content is wrapped in
[PRIOR CONTEXT ...][END OF PRIOR CONTEXT — COMPACTION SUMMARY BELOW]
delimiters, then the summary, then the END MARKER. _append_text_to_content
handles both string and multimodal-list content.
Salvaged from #56372 by @Gromykoss. Only the END-MARKER reorder half is
carried over. The PR's second change (a post-compaction pass that strips
user-role messages before the first summary marker on compression_count>=2)
was dropped: on 2nd+ compactions the protected head decays to system-only
(_effective_protect_first_n -> 0, #11996) so the targeted 'ghost head user'
does not occur, and where the strip does fire it deletes legitimate recent
tail user turns (data loss) and can leave consecutive assistant messages
(role-alternation violation).
