Replace the vertical option-row lists with a compact SegmentedControl
(grouped pill buttons on a single track), dropping the per-option
descriptions since the section subtitle already covers the context.
Add a base-layer rule giving every interactive control (button, select,
menu item, switch, tab, summary) cursor:pointer, and strip the now-redundant
hardcoded cursor-pointer from those elements (plain clickable divs/labels
keep theirs). Remove the divide-y separators from settings list sections so
they breathe.
Color Mode and Tool Call Display become flat radio-style rows (no tile
border/fill, no inner icon box, no filled check badge — just a subtle active
bg and a check). Theme drops its outer card wrapper so only the preview
swatch shows, with a primary ring marking the active palette.
Remove the outer card chrome (border/bg/shadow/rounded) wrapping each
appearance section so they're flat headings + option grids instead of
boxes nested inside boxes, matching the other settings pages.
Mirror the buttonVariants exercise for non-composer form controls: add a
single controlVariants source of truth (2.5px radius, 12px text,
padding-driven sizing, chrome via desktop-input-chrome) and consume it from
Input, Textarea, and SelectTrigger. Drop per-call radius/height/font
overrides that fought the shared look.
Bump default/sm vertical padding a step (the 12px pass read too small) and give
non-icon buttons a subtle 2.5px radius instead of square corners. Icon buttons
keep their 4px.
New reusable `text` variant renders a button as inline label text (no
bg/border, muted -> foreground, underline-on-hover affordance). Emphasize the
actionable word by adding `font-semibold`/`underline` at the call site. Applied
to the auxiliary-model "Set to main" (plain), "Change" and "Reset all to main"
(bold + underlined) actions, replacing the boxed ghost/outline buttons.
- Button base font down to 12px (text-xs) for the dense desktop scale.
- Remove the decorative Sparkles glyph from the model "Apply" button (keep the
spinner while applying).
- Drop the page-level section titles that just restate the left nav ("Main
model", "Appearance", "MCP servers") — the sidebar already labels the pane.
Sub-section headings (Auxiliary models, LLM providers, etc.) stay.
- Buttons: smaller default font (14px -> 13px) and tighter padding-driven sizes
across every variant; the chunky shadcn scale read as oversized in a dense
desktop UI.
- Overlay split layout (settings / command center): the shared OverlayView top
padding left the card surface showing as a gap above the sidebar. Move the
titlebar clearance into each column so the sidebar background runs flush to
the card's top edge.
- Consolidate buttons that hardcoded size/radius/font onto the proper size
variants (tooltip-icon-button, overlay close, cron IconAction, SidebarTrigger,
gateway system button, session-row actions radius, title chip radius, release
notes link) so styling flows from variant props, not per-call overrides.
Composer and the inline approval strip are intentionally left as-is.
Default button sizing was vanilla-shadcn chunky (fixed h-9, 16px padding) and
inconsistent with the icon-button radius pass. Size text variants by
padding + line-height instead of fixed heights so they stay snug and scale
with content, and drop the radius on non-icon buttons (icon buttons keep the
shared 4px). Move the update-overlay CTAs off a hardcoded h-10 onto the
padding-based lg variant. Composer and the inline approval strip are untouched.
Replace the background-clarify toast (expired on alt-tab, easy to miss) with a
persistent, glowing amber "needs input" dot on the session's sidebar row,
driven off a new ClientSessionState.needsInput flag mirrored into a
$attentionSessionIds store. The flag is set on clarify.request and cleared the
moment the turn resumes (tool.complete) or ends.
Also: redesign the clarify tool UI (borderless choices, pseudo-radio dots,
right-aligned checkmark, arc border, tighter padding), make Button the single
source of icon-button styling (4px radius, new icon-titlebar variant, titlebar
buttons rendered polymorphically via asChild, Codicons throughout), put the
file-tree refresh action first, and .trim() pasted composer text.
* Port from google-gemini/gemini-cli#21541: back up corrupted config.yaml
When config.yaml fails to parse, load_config() silently falls back to
DEFAULT_CONFIG and leaves the broken file on disk. If the user then re-runs
the setup wizard or hermes config set (both rewrite config.yaml), their
broken-but-recoverable overrides are lost for good.
Adapts the policy-file recovery from gemini-cli#21541: on the first parse
warning for a given broken file, snapshot it to config.yaml.corrupt.<ts>.bak
(best-effort, symlink-guarded, size-deduped) and tell the user where it
landed. Unlike Gemini's version we deliberately do NOT reset config.yaml to a
clean state — hermes never silently mutates user config, and leaving it means
a hand-fixed file is re-read on the next load.
Tests: 3 new cases (backup created + content preserved + original untouched;
same-size backup dedup; symlink not copied). E2E verified with isolated
HERMES_HOME and a real tab-indented broken config.
* feat(dashboard): add Debug Share to the System page
Surface `hermes debug share` in the dashboard. The System > Operations
section gets a dedicated card that uploads a redacted report + full logs
and returns the paste URLs as real, copyable links instead of a log tail.
- debug.py: factor a pure build_debug_share() returning structured
{urls, failures, redacted, auto_delete_seconds}; run_debug_share now
calls it (CLI output unchanged).
- web_server.py: POST /api/ops/debug-share runs the share core in a
worker thread and returns the structured payload synchronously (the
URLs are the whole point — not a backgrounded action).
- api.ts: runDebugShare() + DebugShareResponse.
- SystemPage.tsx: share card with a redaction toggle (on by default),
per-link + copy-all buttons, and the 6h auto-delete countdown.
- tests: build_debug_share core + endpoint (redact toggle, failure 502,
token gate).
clarify.request is a one-shot blocking event: the gateway turn blocks on
clarify.respond. The desktop handler dropped it for any non-focused session
(`if (!isActiveEvent) return`) and stored at most one request in a single
global atom, so a background session that asked a clarifying question hung
forever and re-focusing it could never recover (the event was already gone).
- store/clarify.ts: key pending requests by runtime session id; expose the
active session's request via a focus-scoped computed view (ClarifyTool is
unchanged). clearClarifyRequest takes an optional session id for targeted
clears, with a request-id fallback.
- use-message-stream.ts: park every session's clarify (drop the isActiveEvent
early return); toast when one lands for a background session since the row
otherwise just keeps spinning like normal work.
- clarify-tool.tsx: clear by session id so answering one chat can't wipe
another's pending request.
- store/clarify.test.ts: concurrent independence, focus-scoped view,
targeted/stale/fallback clears.
* fix(desktop): render approval/sudo/secret prompts so tools stop silently timing out
The desktop app's gateway event handler (use-message-stream.ts) handled
clarify.request but had no case for approval.request, sudo.request, or
secret.request. When a tool needed approval, the gateway emitted
approval.request and blocked the agent thread in _await_gateway_decision()
for up to 5 min (approvals.gateway_timeout); the desktop dropped the unknown
event, never showed a dialog, then the agent returned BLOCKED. No prompt,
just a stall then a block.
The Ink TUI already handles all three (createGatewayEventHandler.ts); this
brings the Electron app to parity.
- store/prompts.ts: approval/sudo/secret atoms (+ request-id-guarded clears)
- components/prompt-overlays.tsx: Radix dialogs; close/Esc maps to refusal so
silence is never mistaken for consent (parity with TUI Esc->deny)
- use-message-stream.ts: wire the three *.request cases; clearAllPrompts on
message.complete so an overlay can't outlive its turn
- chat-messages.ts: GatewayEventPayload gains command/description/env_var/prompt
- mount PromptOverlays in the chat shell
* feat(desktop): inline tool-call approval bar (Cursor-style "Run")
Render dangerous-command / execute_code approval inline on the pending
tool row instead of as a modal. Binding is positional: the desktop
tool.start payload carries no structured args, but approval.request only
fires from the terminal/execute_code guards and the agent blocks on one
approval at a time, so the single pending row of those tools is the one
that raised it. Command/description text comes from $approvalRequest.
Drops ApprovalDialog from PromptOverlays (sudo/secret stay modal).
* style(desktop): make inline approval bar match Cursor's command card
Drop the amber alert styling for a neutral elevated card: command on a
terminal-prefixed row up top, a divided footer with the muted description
on the left and right-aligned controls — a ghost "Reject" (Esc) plus a
split primary "Run" (⌘⏎) whose chevron opens "Allow this session" /
"Always allow" / "Reject". Wire ⌘/Ctrl+Enter → Run and Esc → Reject to
match Cursor's accept/skip bindings, guarded against double-send via the
$approvalRequest atom.
* style(desktop): shrink inline approval to a tiny Cursor-style button strip
The running tool row already shows the command, so drop the whole card +
command echo + description band. What's left is a compact strip under the
row: a small split "Run ⌘⏎" button (chevron → Allow this session / Always
allow / Reject) and a ghost "Reject Esc", indented to sit under the row's
title text.
* style(desktop): drop the loud blue Run button for a quiet outlined control
Swap the primary (blue) Run for a subtle outlined split control — neutral
border, transparent fill, hover-accent — so the approval strip reads as
quiet inline affordance rather than a big CTA. Reject stays ghost.
* style(desktop): make Run a soft primary badge
Tint the Run split control with the primary color as a badge (bg-primary/10,
primary text, primary/25 border, rounded-md, hover primary/15) instead of a
solid CTA or a neutral outline.
* style(desktop): slim the approval chevron and space out Reject
The chevron button had ballooned because dropping the size prop fell back
to the big default size (h-9 + has-svg px-3). Pin size=xs everywhere and
give the chevron a tight w-5/px-0. Bump the gap between the Run badge and
Reject (gap-2.5) and loosen Reject's internal spacing.
* feat(desktop): confirm before "Always allow" persists an approval
"Always allow" writes the matched pattern to ~/.hermes/config.yaml and
suppresses the prompt in every future session — too consequential to fire
straight from a menu click. Route it through a confirm dialog that names
the pattern + command and the file it touches. The dialog owns the
keyboard while open so Esc closes it instead of denying the approval.
* fix(gateway): make sudo + secret prompts actually fire in the desktop
Tek's PR added the sudo/secret overlays and callback wiring, but neither
reached the live path:
- Sudo: the sudo password callback is thread-local (terminal_tool
_callback_tls), and _wire_callbacks runs on the agent-build thread, not
the turn thread that executes tools. At command time the callback was
missing, so terminal sudo fell through to /dev/tty and hung the headless
gateway. Re-wire callbacks at the top of the prompt-submit turn thread.
- Secret: skills_tool short-circuited to the "secret entry unsupported"
hint for any gateway surface, before invoking the callback. Interactive
surfaces (desktop/TUI) register a secret-capture callback that routes to
the secret.request overlay; only short-circuit when no callback exists,
so messaging still gets the hint but the desktop prompts.
* docs(desktop): drop Cursor references from approval comments
* docs(desktop): drop Cursor reference from prompt-overlays comment
* fix(skills): gate in-band secret capture on HERMES_INTERACTIVE, not callback presence
The desktop/sudo PR switched the gateway secret-capture short-circuit from
"any gateway surface" to "gateway surface with no callback registered". That
made a messaging gateway (telegram/discord/...) attempt interactive in-band
secret capture whenever any callback happened to be registered, instead of
returning the safe "setup unsupported" hint — and broke
test_gateway_still_loads_skill_but_returns_setup_guidance.
Discriminate on HERMES_INTERACTIVE instead: the desktop app / TUI set it in
_enable_gateway_prompts (alongside registering the secret.request callback),
while messaging platforms never do. This is the same flag tools/approval.py
uses to tell an interactive surface from a messaging one, so messaging keeps
the hint and desktop/TUI still prompt.
---------
Co-authored-by: Brooklyn Nicholson <brooklyn.bb.nicholson@gmail.com>
Salvage of #35508 (@dchenk), rebased onto current main. Resolved the
tests/tools/test_stage2_hook_puid_pgid.py conflict (kept both the
envdir-creation regression test on main and the new config-migration
tests).
Docker image upgrades replace code under $INSTALL_DIR but preserve
$HERMES_HOME on the mounted volume, so the persisted config.yaml never
received the schema migrations that non-Docker `hermes update` runs
(#35406). This adds scripts/docker_config_migrate.py, invoked from
stage2-hook after first-boot seeding and before gateway services start:
it backs up config.yaml + .env, runs migrate_config(interactive=False),
and honors HERMES_SKIP_CONFIG_MIGRATION=1 for manual control.
Also fixes a latent bug in check_config_version(): it called load_config()
which deep-merges DEFAULT_CONFIG, so a legacy config with no raw
_config_version falsely reported as already-current. It now reads the raw
on-disk file so legacy configs are correctly detected for migration.
Differs from #35508 as submitted (Option B cleanup): dropped the
`_config_version` line added to cli-config.yaml.example and removed the
accompanying test_cli_config_example_declares_latest_version change-detector
test. The example is a copy-template and has no business asserting a schema
version; check_config_version() reads the user's real config.yaml, not the
example. This removes a second sync point that drifts on every version bump.
Closes#35508. Fixes#35406.
Co-authored-by: Dmitriy Cherchenko <17372886+dchenk@users.noreply.github.com>
`docker run --user $(id -u):$(id -g)` was a tini-era trick to make
container-written files match the host user. Under s6-overlay it no longer
works: the bootstrap (UID remap, volume + build-tree chown, config seeding)
needs root, and the baked image dirs (/opt/data, /opt/hermes/.venv, ui-tui,
node_modules) are owned by the hermes build UID (10000). A pinned arbitrary
UID can't write them, so the runtime fails with EACCES on a bind mount or
hard-crashes on a named volume (Docker inits the volume from the image as
10000; the non-root start can't even `cd /opt/data`, and the profile
reconciler dies with PermissionError on gateway_state.json).
Detect that start early in both the cont-init hook (stage2-hook.sh) and the
CMD wrapper (main-wrapper.sh) and fail fast with actionable guidance pointing
at the supported path: root start + HERMES_UID/HERMES_GID (or the PUID/PGID
aliases), which remaps the hermes user and chowns the volume — the same
host-UID-matching outcome --user was used for, without breaking s6.
The guard fires only when the current UID is neither root NOR the hermes UID.
This preserves the supported non-root start from #34648/#34837 (running with
`--user 10000:10000`, i.e. pinned to the hermes UID itself), which is
unaffected — only the arbitrary-UID variant that #34837 never actually made
writable is rejected.
Verified live across five scenarios (built image, bind + named volume):
arbitrary --user on bind -> rejected with guidance, hermes does not run;
arbitrary --user on named volume -> guidance shown, no raw 'can't cd' crash;
--user 10000:10000 -> boots; root + HERMES_UID=4242 remap -> boots, guard not
tripped; default root start -> boots. Pre-fix control reproduces the raw
PermissionError + 'can't cd' crash with no guidance.
`hermes mcp add --auth header` built `Authorization: Bearer ${MCP_X_API_KEY}`
and passed it straight to the discovery probe without interpolation, so the
probe sent the literal placeholder and auth-requiring servers (e.g. n8n)
returned 401. Runtime tool loading worked because `_load_mcp_config()`
interpolates, but the four CLI probe call sites (add/test/login/configure)
all used unresolved config.
Resolve `${ENV}` inside `_probe_single_server` via a new
`_resolve_mcp_server_config()` (load_hermes_dotenv + _interpolate_env_vars),
mirroring runtime loading. This covers all four call sites, not just add.
Also strip a leading `Bearer ` from pasted tokens before saving to
`MCP_*_API_KEY`, so a token pasted with the prefix doesn't produce
`Bearer Bearer <jwt>` (also a 401).
Reported with a precise root-cause analysis in #37792.
Co-authored-by: ThyFriendlyFox <116314616+ThyFriendlyFox@users.noreply.github.com>
Onboarding's "Local / custom endpoint" only wrote the OPENAI_BASE_URL env
var, which runtime resolution ignores — so a self-hosted endpoint was never
wired in and setup failed with "No usable credentials found for custom" even
though local servers need no key.
Route the local option through saveOnboardingLocalEndpoint: probe the
endpoint, auto-discover a model from /v1/models, persist provider=custom +
base_url + model via /api/model/set, then verify the runtime directly
(not via completeWithModelConfirm, which would re-assign the model without
base_url and wipe it). No onboarding form/UI changes — the existing single
URL field is enough.
The runtime resolver reads model.base_url from config and ignores the
OPENAI_BASE_URL env var, so a self-hosted endpoint could not be configured
from the GUI. Two changes enable it:
- POST /api/model/set accepts an optional base_url and persists it as
model.base_url when provider=custom (still clearing stale base_url for
hosted providers).
- POST /api/providers/validate now returns the model ids a custom endpoint
advertises at /v1/models, so the GUI can auto-pick a default without
asking the user to type a model name.
Refs desktop onboarding "Local / custom endpoint" bug.
The setup-flow provider list showed two Anthropic/Claude entries with
ambiguous labels ('Anthropic (Claude API)' and 'Claude Code (subscription)')
in no deliberate order. Relabel and reorder so the distinction and the
subscription caveat are explicit:
- 'Anthropic API Key' (PKCE, API path)
- 'Anthropic OAuth: Required Extra Usage Credits to Use Subscription' (external)
- Both Anthropic entries moved to the bottom of the list.
- 'OpenAI Codex (ChatGPT)' -> 'OpenAI OAuth (ChatGPT)', now first after Nous.
Applied consistently to the backend OAuth catalog (web_server.py) and the
desktop onboarding overlay's PROVIDER_DISPLAY title/order map; test
assertions updated to the new titles.
Hammer createTokenizer with the worst stalls a terminal can produce —
split + flush at every interior byte, and a 200-report byte-by-byte feed
that flushes after every single byte — and assert the two invariants that
make the SGR-leak class structurally impossible: nothing ever leaks as a
text token, and every complete report reassembles whole. A mixed
mouse+keystroke variant proves real input survives the same storm.
hermes -c / --resume now reopen a session in its original working
directory. The sessions table already had a cwd column; the classic CLI
just never wrote or read it.
- run_agent._ensure_db_session stamps cwd for local CLI sessions only
(new _launch_cwd_for_session gates out gateway/cron and non-local
terminal backends, where a host cwd is meaningless to restore).
- cli._restore_session_cwd chdir's the process AND retargets TERMINAL_CWD
so the terminal tool, code-exec tool, and relative-path resolution all
land in the restored dir. Called from both resume paths (interactive
run() and the -q single-query path).
- Robust degradation: no-op when no cwd recorded, when already there, or
when the dir is gone (single dim warning, stays put — no crash).
With the tokenizer reassembling split CSI sequences across a flush (prior
commit), no SGR mouse fragment can reach a text token anymore — terminals
write a mouse report as one atomic sequence, and any read/flush split now
re-joins in the tokenizer buffer instead of leaking. That makes the whole
downstream recovery layer dead code:
- SGR_MOUSE_FRAGMENT_RE, MOUSE_BURST_NOISE_RE, MOUSE_BURST_RESIDUE_RE
- parseTextWithSgrMouseFragments / parseSgrMouseFragment /
normalizeSgrMouseFragment
- the whole-text mouse-burst noise fast path in parseMultipleKeypresses
Remove all of it (~185 lines) and the tests that only exercised it. The
narrow legacy X10 wheel-tail resynth stays (distinct mechanism, kept with
its own test). This retires the #17701 → #18113 → #26781 → #28463 → #35512
regex hardening chain in favor of the one correct parser fix.
The guard it covered was removed in the previous commit (fragments no
longer reach input-event — they reassemble at the tokenizer). Reassembly
is now covered by termio/tokenize.test.ts and the flush-boundary cases in
parse-keypress.test.ts.
Root-cause fix for the SGR mouse fragment leak (`46M35;40M...` typed into
the prompt). The leak was never really about the fragments — it was the
flush emitting them. When App's 50ms watchdog fires mid-CSI during a render
stall, the tokenizer was force-emitting the buffered partial as a token and
resetting to ground, so both the prefix and the ESC-less remainder surfaced
as unparseable input.
Make the flush state-aware (xterm.js discipline): a bare ESC still flushes
to the Escape key (the legitimate ESCDELAY case), but a buffer still inside
a multi-byte control sequence (csi/osc/dcs/apc/ss3/intermediate) is NOT
emitted — it's kept so the continuation reassembles on the next feed. A
one-tick truncation valve in createTokenizer.flush() drops a partial that
survives a second flush with no progress, so a genuinely truncated write
can't fuse into the next keypress.
With partials never entering the input stream, the downstream scrubber is
dead code: remove the SGR fragment guard from input-event.ts (both the
original `/^\[<\d+;\d+;\d+[Mm]/` and the consolidated form added earlier in
this PR). The parse-keypress burst-recovery regexes (MOUSE_BURST_*) are now
also redundant but left in place as a safety net for one release; they can
be removed in a follow-up once this soaks.
Tests: tokenize.test.ts proves a mid-CSI flush keeps/reassembles and that a
stale partial is dropped after a second flush and a bare ESC still emits;
parse-keypress.test.ts adds the end-to-end split-then-reassemble case
yielding a single clean mouse event with no leaked key.
Supersedes #29337.
The stage2 hook gates the recursive chown of the build trees under
$INSTALL_DIR (.venv, ui-tui, node_modules) so a HERMES_UID/PUID remap
leaves them writable by the new runtime UID — needed for lazy_deps
'uv pip install' of platform extras (#15012, #21100) and the TUI esbuild
rebuild into ui-tui/dist (#28851).
#35027 folded that chown under the $HERMES_HOME ownership check
('stat $HERMES_HOME != hermes_uid'). But 'usermod -u <new> hermes'
re-chowns the hermes home dir ($HERMES_HOME == /opt/data) to the new UID
as a side effect, so after any remap that stat is already satisfied and
needs_chown is false — silently skipping the build-tree chown on the
common PUID/NAS path. The venv stays owned by the build-time UID (10000),
so lazy installs and TUI rebuilds fail with EACCES.
Probe the build trees directly instead: chown only when /opt/hermes/.venv
is not already owned by the runtime hermes UID. Independent of
$HERMES_HOME ownership, idempotent across restarts.
Verified live: built the image, booted with HERMES_UID/HERMES_GID on a
fresh named volume, confirmed .venv/ui-tui/node_modules end up owned by
the remapped UID and 'uv pip install' into the venv succeeds; confirmed
the recursive chown fires once and is skipped on restart.
When App's 50ms flush watchdog fires mid-CSI during a render stall, an
SGR mouse report (ESC[<btn;col;row M/m) is split across stdin chunks: the
tokenizer force-emits the buffered prefix and resets to ground, so both
the prefix and the ESC-less remainder reach InputEvent as nameless tokens.
The previous guard only matched a full `[<\d+;\d+;\d+[Mm]` fragment, so
the flushed prefixes (`ESC[<0;35;`) and the 1-/2-field and leading-`;`
tails (`46M`, `35;46M`, `;46M`) still leaked into the composer as
`46M35;40M...` during long sessions.
Replace the three would-be narrow regexes with one consolidated rule that
covers every split position. A `(?=...\d)` lookahead keeps typed `<`, `[`,
`;`, and `M` safe (no coordinate digit), and the embedded M/m terminator
in the param class leaves stuck-together fragments / prose intact. The
existing `!keypress.name` gate continues to protect real keystrokes, which
arrive one char per chunk with a name set.
Supersedes #29337 (covers the prefix-leak and leading-`;`/1-/2-field tail
cases that PR's two added guards missed).
The HERMES_AGENT_HELP_GUIDANCE block (added #16535) only fired when the
user explicitly asked about configuring/setting up Hermes. Broaden it so
the agent treats the docs as a standing source of self-knowledge for any
Hermes-related help and for understanding its own features/tools, points
to the hermes-agent skill for additional guidance, and treats the docs as
the authoritative/latest source of truth when the two differ.
Static constant in the cache-safe stable tier — no prompt-cache impact.
Dashboard plugins (kanban, hermes-achievements) read
window.__HERMES_SESSION_TOKEN__ directly and hand-assembled WebSocket
URLs with ?token=. That works in loopback/--insecure mode but is
rejected on OAuth-gated deployments, where the session token is absent
and _ws_auth_ok only accepts single-use ?ticket= auth. The result was
401s on plugin REST calls and 1008/403 on the kanban live-events WS
whenever the dashboard ran behind OAuth (e.g. hosted Fly agents).
Make the plugin SDK the single sanctioned auth surface:
- web/src/lib/api.ts: add authedFetch() (raw Response for FormData
uploads / blob downloads, token-or-cookie auth, no throw / no 401
redirect) and buildWsUrl() (assembles a ws(s):// URL with the correct
auth param for the active mode — fresh single-use ticket in gated
mode, token in loopback).
- web/src/plugins/registry.ts: expose authedFetch, buildWsUrl,
buildWsAuthParam, and sdkVersion on window.__HERMES_PLUGIN_SDK__;
add SDK_CONTRACT_VERSION.
- web/src/plugins/sdk.d.ts: hand-authored typed contract for the
plugin SDK + registry globals (single source of truth for the
Window declarations).
- plugins/kanban + hermes-achievements dist bundles: stop reading the
session token directly; route uploads/downloads through
SDK.authedFetch and the live-events WS through SDK.buildWsUrl.
- plugins/kanban plugin_api.py: _ws_upgrade_authorized() delegates the
/events WS upgrade to the canonical web_server._ws_auth_ok gate, so
it transparently accepts loopback token / gated ticket / internal
credential and can never drift from core auth again.
- tests: guard test asserting no plugin dist reads
__HERMES_SESSION_TOKEN__ directly; kanban gated-ticket WS test.
Verified live on a gated staging Fly agent: kanban /events upgrades
101 with a minted ticket (ticket_len=43, ws_auth_ok=True) where the
old code got 403.
uv selects the project Python from requires-python and from the UV_PYTHON
env var, both of which override an already-created venv on the next
'uv sync'. With no upper bound on requires-python, an inherited
UV_PYTHON=3.14 (or a fresh distro whose newest interpreter uv auto-picks)
silently recreated the installer's 3.11 venv at 3.14, where Rust-backed
transitives (pydantic-core) have no cp314 wheel and fall back to a maturin
source build that fails. This bit a Windows/WSL user with UV_PYTHON set in
their shell and a fresh WSL-arch box where uv auto-picked 3.14.
Two layers:
- pyproject: requires-python '>=3.11' -> '>=3.11,<3.14' (+ uv lock regen).
uv now refuses a 3.14 interpreter with a clear error instead of attempting
the maturin build. Backstop independent of the installer.
- install.sh / install.ps1: pin UV_PYTHON to the venv interpreter after
creating it (in both the venv step and the deps step, since bootstrap runs
those stages as separate processes). An inherited UV_PYTHON can no longer
hijack the sync/pip tiers, so the install just works regardless of shell env.
Verified E2E: hostile UV_PYTHON=3.14 + uv venv --python 3.11 + uv sync now
installs into 3.11 with pydantic-core's 3.11 wheel; without the re-pin the
capped requires-python produces a legible incompatibility error rather than a
cryptic build failure.
Fireworks/Mistral reject HTTP 400 'Extra inputs are not permitted, field:
messages[N].tool_calls[M].extra_content' on any session whose history
contains prior Gemini tool calls. Gemini 3 thinking models attach
extra_content (thought_signature) to tool_calls; it survived to the wire
because the sanitize paths only stripped call_id/response_item_id.
Strip extra_content from the outgoing wire copy in both sanitize paths
(ChatCompletionsTransport.convert_messages + _sanitize_tool_calls_for_strict_api),
but gate it on the target model: keep extra_content for Gemini-family
targets (the thought_signature MUST be replayed or Gemini 400s), strip it
for everyone else — including non-Gemini models that inherit a stale Gemini
signature earlier in a mixed-provider session. Native Gemini is unaffected
(GeminiNativeClient bypasses these paths).
Original stored history is never mutated (only the per-call copy).
Fixes#17986.
The Desktop Chat section described chat-only and gave no signpost that
remote-hosted Hermes connection is documented. Adds a pointer to the
in-page remote-backend section and to the deeper Web Dashboard doc.
The TUI hardcoded --max-old-space-size=8192. V8 is not cgroup-aware, so in a
Docker/k8s container capped below ~9-10GB the heap grows past the container
limit and the cgroup OOM-killer SIGKILLs the Node parent BEFORE V8's own heap
monitor fires. SIGKILL runs no JS handler, writes no [tui-parent] breadcrumb,
and closes the gateway child's stdin — the user sees only a bare gateway
'stdin EOF'. Complements #38224 (trail-text cap), which reduced pressure but
left the 8GB-vs-container mismatch in place.
- _read_cgroup_memory_limit(): read cgroup v2 (memory.max) then v1
(memory.limit_in_bytes); handle 'max', the v1 unlimited sentinel, blank/zero,
and >=1PB as unconstrained.
- _resolve_tui_heap_mb(): unconstrained -> 8192; constrained -> 75% of the
cgroup limit (headroom for non-heap RSS + the Python child sharing the
cgroup), floored at 1536MB, never above 8192.
- NODE_OPTIONS block uses the sized value; still respects a user-supplied
--max-old-space-size.
Net: V8 now GCs/exits gracefully (onCritical breadcrumb fires) instead of being
reaped silently. Display/transport only — no agent context or behavior change.
Tests: tests/hermes_cli/test_tui_heap_sizing.py (20 tests).
The stash/restore cycle in the update path was observed to clobber
freshly-pulled source files (apps/desktop/ deletion -> Vite
'[UNRESOLVED_ENTRY] Cannot resolve entry module index.html'). On a
managed clone the user never edits the source tree, so any 'dirty' state
is pure git artifact (CRLF renormalization, npm lockfile churn, files
left behind when a directory was deleted upstream such as
apps/bootstrap-installer/). Stashing that and re-applying it after a pull
is fragile and unnecessary.
- hermes update (hermes_cli/main.py): on a non-fork (managed) clone,
discard working-tree dirt via reset --hard HEAD + clean -fd instead of
stash/apply. Forks keep the stash machinery so intentional edits
survive. Also pin core.autocrlf=false on Windows so the dirt is never
created (mirrors install.ps1 #38239).
- install.sh: replace the update-path stash/restore dance with a hard
reset to origin/<branch>; the installer is a managed-only entry point.
- install.sh + install.ps1 desktop stage: prefer 'npm ci' (wipes and
reinstalls node_modules from the lockfile) over bare 'npm install',
which can report 'up to date' against a stale marker while node_modules
is empty -- leaving tsc unresolved so 'npm run pack' fails.
Tests: managed clone cleans instead of stashing; fork still stashes;
existing stash tests force the stash path explicitly.
The launch provider setup screen rejected too many legitimate users:
a live credential probe ("key rejected"), a post-save runtime check
("still cannot reach X"), and an 8-char minimum all gated progression.
Corporate proxies, regional blocks, rate-limited/flaky probes, and
self-hosted endpoints all tripped these. Now we just require a
non-empty value and save it; a genuinely bad key surfaces later at
chat time instead of blocking onboarding.
Desktop's readiness probe only checks GET /api/status (public), but the
live chat rides /api/ws, which is gated by --tui (4403), a matching
session token (4401), and a non-loopback bind. The web-dashboard doc
covered --tui and the OAuth gate but never the Desktop remote-connection
flow, so the three independent failure modes weren't documented together.
Adds a 'Connecting Hermes Desktop to a remote backend' section: pin
HERMES_DASHBOARD_SESSION_TOKEN, run with --host 0.0.0.0 --insecure --tui,
the curl token-verification one-liner, and WS close-code triage.
The desktop chat app's slash curation (desktop-slash-commands.ts) only
suggested the ~19 curated built-ins. isDesktopSlashSuggestion required
membership in DESKTOP_COMMANDS, so every skill-derived command and user
quick_command was silently dropped from both completion paths
(commands.catalog empty-query + complete.slash typed-query) and from
filterDesktopCommandsCatalog — even though isDesktopSlashCommand let them
EXECUTE when typed in full. The tui_gateway backend already includes skills
in both RPCs; the gap was purely renderer-side.
Add isDesktopSlashExtensionCommand() (= not-a-known-Hermes-built-in, the
same predicate that already gates execution) and let extensions through the
suggestion path. The catalog filter routes through isDesktopSlashSuggestion,
so skill/quick-command categories and pairs are kept automatically.
The native Hindsight memory provider lazy-installs hindsight-client into
/opt/hermes/.venv at first use (tools/lazy_deps.py: memory.hindsight).
That venv lives inside the immutable image layer, not the mounted
/opt/data volume, so the dependency is wiped on every container recreate
/ image update. After an update, profile config still points at Hindsight
and the Hindsight server is healthy, but recall/retain fails with:
ModuleNotFoundError: No module named 'hindsight_client'
The manual workaround (uv pip install hindsight-client inside the running
container) doesn't survive the next recreate, and pip-install-into-.venv
is not an officially supported durable Docker workflow.
Fix: add --extra hindsight to the image's uv sync line, same pattern as
the --extra anthropic/bedrock/azure-identity providers (#30504) and
--extra messaging (#24698) — bake the optional dependency into the build
layer so it survives container recreate. The pyproject [hindsight] pin
(hindsight-client==0.6.1) already matches tools/lazy_deps.py and uv.lock,
so this is a pure additive --extra with no lockfile churn.
Verified: 'uv sync --frozen --no-install-project --extra hindsight'
against the committed uv.lock installs hindsight-client 0.6.1 and the
module imports cleanly.
Adds a regression test (mirrors test_dockerfile_preinstalls_gateway_
messaging_dependencies) so a future Dockerfile cleanup can't silently
drop the extra.
* fix(packaging): modernize project.license to PEP 639 SPDX string
Drops the SetuptoolsDeprecationWarning ('project.license as a TOML table
is deprecated') emitted on every editable build under setuptools>=77 by
switching license = { text = "MIT" } to the SPDX string form plus an
explicit license-files entry. Bumps build-system requires to
setuptools>=77 so an older build backend can't reject the string form.
The warning was non-fatal (builds succeed with it) but surfaces
prominently in install.ps1 build-failure output, where it gets mistaken
for the cause of unrelated Windows build_editable crashes.
* fix(packaging): bound setuptools build requirement per supply-chain policy
Add the <83 upper bound to setuptools>=77.0 so the dep-bounds supply-chain
gate (>=floor,<next_major) passes.
Self-review of #38465 surfaced three real items:
1. SystemExit escape (defense): `_login_nous` raises SystemExit(130)/(1) on
cancel/failure. The logged-out login path inside `_model_flow_nous` catches
it, but the expired-session re-login path (main.py) only catches Exception,
so a Ctrl-C during re-auth could propagate past `_run_portal_one_shot` and
kill the CLI. Add SystemExit to the portal handler so all cancel/abort cases
end with the graceful 'Setup cancelled / retry later' message.
2. Doc sweep: the model-pick step was only added to the bare-`hermes portal`
prose. Propagate it to the surfaces describing `hermes setup --portal`
behavior that still omitted model selection:
- `--portal` argparse help (main.py)
- nous-portal.md intro + the numbered 'what it does' step list (EN + zh-Hans)
- run-hermes-with-nous-portal.md 'default model after setup --portal' line,
which was now contradictory (there's a picker, not a forced default) (EN + zh)
3. Test coverage: add parametrized regression test asserting the portal handler
swallows KeyboardInterrupt / EOFError / SystemExit (returns None, no escape).
Note on 'Skip (keep current)': delegating to _model_flow_nous means picking
Skip preserves the prior provider instead of force-switching to nous — this is
intentional and matches quick setup exactly; docs now say 'sets Nous as your
provider (when you pick a model)' rather than unconditionally.
`hermes portal` / `hermes setup --portal` previously logged in and set
provider=nous but left the model UNSELECTED (blank -> runtime default) and
never showed a picker — unlike the first-time quick setup, which runs the
model picker.
Route `_run_portal_one_shot` through `_model_flow_nous` — the exact same
routine quick setup (`_run_first_time_quick_setup`) and `hermes model` -> Nous
use. It handles both the logged-out path (device-code OAuth, which picks a
model internally) and the logged-in path (curated Nous model picker), then
offers the Tool Gateway opt-in and sets provider=nous. Net effect: `hermes
portal` now offers a model picker every time and is a true single-command
collapse of quick setup's Nous step.
Removes the hand-rolled auth_add_command + manual provider write + separate
Tool Gateway prompt (now a single source of truth). Re-syncs the in-memory
config from disk afterward so a caller's later save_config can't clobber the
model/provider written by the login flow.
Docs (CLI help, portal_cli docstrings, nous-portal EN + zh-Hans) updated to
mention model selection. New regression test asserts `_run_portal_one_shot`
delegates to `_model_flow_nous`.
Verified live: `hermes portal` now shows the 27-model curated picker, 'Skip
(keep current)' preserves prior provider/model.
