os.fchmod is Unix-only; the Windows os module has no fchmod (only
chmod). Passing mode= (e.g. 0o600 when saving the Hindsight config
during `hermes memory setup`) crashed on Windows with:
AttributeError: module 'os' has no attribute 'fchmod'
Guard the fchmod fast-path with hasattr(os, "fchmod"). Skipping it on
Windows is safe: mkstemp already creates the temp file as 0o600, and
the existing post-replace os.chmod(real_path, mode) — already wrapped
in try/except — applies the final mode durably (as far as Windows
honors it).
Adds regression tests: one simulating a Windows os module without
fchmod (must not raise), and one asserting the durable 0o600 mode on
POSIX.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Subway2023's #14639 blocks write_file/patch to ~/.hermes/config.yaml, but
the terminal side was only partially paired: echo>/tee/cp/mv to config.yaml
already tripped the project-config pattern, while `sed -i` and direct edits
slipped through with auto-approve. An unpaired write_file deny is theater per
SECURITY.md — the agent could flip approvals.mode=off via `sed -i` and the
mtime-keyed config cache reloads it mid-session.
config.yaml IS the security policy (approvals.mode/yolo/permanent allowlist
live there), so it warrants real pairing, not a half-door. Add a
_HERMES_CONFIG_PATH fragment mirroring _HERMES_ENV_PATH, fold it into
_SENSITIVE_WRITE_TARGET (covers tee/>/>>/cp/mv), and add sed -i coverage for
both config.yaml and .env. Pins 9 regression tests including no-regression
guards (reads pass, /tmp writes pass).
Co-authored-by: sbw2025 <subw3@mail2.sysu.edu.cn>
A stream that drops mid-response after tokens are delivered (peer-closed
connection, stale-stream reconnect) is converted into a synthetic
finish_reason="length" stub. The conversation loop treated that network
stall as a max-output-tokens truncation: when the dropped content was a
tool call it retried exactly once, then hard-failed with "Response
truncated due to output length limit" — even on large-output models that
never hit any cap (e.g. Opus).
- Tool-call truncation now retries up to 3 times (was 1) with a
progressive max_tokens boost, and is stub-aware: a PARTIAL_STREAM_STUB_ID
stall prints "Stream interrupted mid tool-call — retrying (n/3)" instead
of the false "model hit max output tokens", and the give-up message
distinguishes a network drop from a real truncation.
- Length-continuation retries preserve the original request's output cap
as a floor, so a high provider/model default isn't silently downshifted
to 8K/12K on retry.
- Added _requested_output_cap_from_api_kwargs() helper.
Tests: stub-stall mid-tool-call recovery within 3 retries; continuation
preserves a large provider-default output cap.
Fixes#26425. Salvages the substance of #26427 (cap floor) and #9525
(retry bump), adapted to the post-refactor conversation_loop.py which
handles all three api_modes uniformly.
Co-authored-by: LeonSGP43 <cine.dreamer.one@gmail.com>
Co-authored-by: ygd58 <ygd58@users.noreply.github.com>
* feat(dashboard): backend API for MCP, pairing, webhooks, credential pool, memory, gateway lifecycle
Adds REST endpoints so a remote admin can manage these without CLI access:
- MCP servers: list/add/remove/test (config.yaml parity with hermes mcp)
- Pairing: list/approve/revoke/clear-pending messaging codes
- Webhooks: list/subscribe/remove (hot-reloaded JSON store)
- Credential pool: list/add/remove rotation keys (via CredentialPool API)
- Memory provider: status/select/disable/reset
- Gateway lifecycle: start/stop (restart+update already existed)
Secrets redacted on read; usable values only reach the agent at session start.
All endpoints sit behind the existing dashboard auth gate.
* feat(dashboard): backend API for ops + skills hub
- Ops actions (spawned, log-tailed via /api/actions): doctor, security audit,
backup, import, checkpoints prune
- Ops reads (structured JSON): hooks list + allowlist status, checkpoints list
with per-session size
- Skills hub actions (spawned): install / uninstall / update
- Registers new action log files for all spawn-based endpoints
All gated by the existing dashboard auth middleware.
* feat(dashboard): admin pages for MCP, pairing, webhooks, and system ops
Adds four new dashboard pages + nav entries so a remote admin can manage
Hermes without CLI access:
- MCP: list/add/remove/test MCP servers
- Webhooks: list/create/delete subscriptions (one-time secret reveal)
- Pairing: approve/revoke/clear messaging pairing codes
- System: gateway start/stop/restart, memory provider + reset, credential
pool add/remove, ops (doctor/audit/backup/import/skills update) with a
live action-log viewer, checkpoints prune, shell-hooks status
api.ts: client methods + types for all new endpoints.
App.tsx: routes + sidebar nav (plain labels, no i18n key required).
Verified: tsc -b clean, production build succeeds, new pages lint clean,
zero new eslint errors in App.tsx.
* test(dashboard): cover admin API endpoints
20 tests across MCP, credential pool, memory, pairing, webhooks, ops, plus
an auth-gate parametrize that asserts every admin endpoint requires the
session token. Asserts request contract + CLI-config parity, not catalog
values (per the no-change-detector-tests rule).
* docs(dashboard): document MCP, Webhooks, Pairing, and System admin pages
Adds Pages sections for the four new admin tabs and an Admin-endpoints table
to the REST API reference. Updates the page description to reflect the
dashboard's expanded role as a full administration panel.
* feat(install): --no-skills flag for blank-slate default profile
Add an install-time --no-skills flag so the default ~/.hermes profile can
be created with zero bundled skills, matching what
`hermes profile create --no-skills` already does for named profiles.
The flag writes $HERMES_HOME/.no-bundled-skills and skips the install-time
seed. sync_skills() now honors that marker with an early return
(skipped_opt_out=True), so neither the installer, a later `hermes update`,
nor a direct sync re-injects bundled skills into a profile that opted out.
Previously the marker was only checked by seed_profile_skills() (named
profiles); the default profile had no opt-out and `hermes update` would
re-seed it every time.
Tests: TestNoBundledSkillsOptOut covers marker-present (no-op) and
marker-absent (normal seed) paths.
* feat(skills): hermes skills opt-out / opt-in for existing profiles
Adds an interactive counterpart to the install-time --no-skills flag so
an already-installed profile (default or named) can toggle the
.no-bundled-skills marker without reinstalling.
- `hermes skills opt-out` writes the marker (stop future seeding). Safe
by default: nothing on disk is touched.
- `hermes skills opt-out --remove` ALSO deletes already-present bundled
skills, but ONLY ones that are manifest-tracked AND byte-identical to
their origin hash. User-edited bundled skills, hub-installed skills, and
hand-written skills are never removed. Previews + confirms before
deleting (--yes to skip).
- `hermes skills opt-in [--sync]` removes the marker and optionally
re-seeds immediately.
Core logic lives in tools/skills_sync.py (set_bundled_skills_opt_out,
is_bundled_skills_opt_out, remove_pristine_bundled_skills) reusing the
existing manifest origin-hash machinery for the safety check.
Tests: TestOptOutToggleAndRemove covers marker toggle idempotency and
proves user-modified + non-bundled skills survive --remove.
* docs: blank-slate skills — install --no-skills + opt-out/opt-in
- features/skills.md: new 'Starting with a blank slate' section covering
the install flag, profile-create flag, and runtime opt-out/opt-in, with
a safe-by-default note.
- reference/cli-commands.md: document the new skills opt-out / opt-in
subcommands + examples.
- reference/profile-commands.md: fix the marker filename (was .no-skills,
actually .no-bundled-skills) and cross-link the runtime commands.
Validated with a full docusaurus build (exit 0); the three edited pages
compile clean with no new warnings.
Two related changes to the skill curator:
1. Built-in pruning. New curator.prune_builtins config (default on) lets the
curator archive bundled built-in skills after the inactivity period, not
just agent-created ones. A .curator_suppressed list tells the update-time
re-seeder (tools/skills_sync) to leave pruned built-ins archived, so the
prune is durable across `hermes update`. Built-ins are seeded with a
baseline record on first sight, so the inactivity clock starts at upgrade
time -- no mass-prune on the first run. Hub-installed skills are never
pruned regardless of the flag. Restoring a built-in clears its suppression.
2. Usage tracking for all skills. Telemetry (view/use/patch) was wrongly gated
behind curation-eligibility, so built-ins were tracked only when prunable
and hub skills never. Telemetry is observability and is now decoupled from
curation: every skill accrues usage counts regardless of provenance, while
lifecycle mutators (set_state/set_pinned/mark_agent_created) stay
curation-gated. New usage_report() + provenance() expose all skills with an
agent/bundled/hub tag.
Gateway /undo was wired into every platform but still ran the old
single-turn hard-truncate. Now it matches the CLI/TUI: /undo [N] backs
up N user turns (default 1, clamps to oldest), soft-deletes the
truncated rows on disk (active=0, kept for audit, hidden from re-prompts
and search) via SessionDB.rewind_to_message, evicts the cached agent so
the next turn rebuilds from the active-only transcript (the gateway's
equivalent of the CLI's in-place history surgery + memory invalidation),
and echoes the backed-up message text so the user can copy/edit and
resend — platforms have no editable composer to prefill.
- gateway/session.py: SessionStore.rewind_session(session_id, n) wraps
the soft-delete primitive; load_transcript already returns active-only
- gateway/run.py: _handle_undo_command parses [N], calls rewind_session,
evicts the agent, echoes target text; confirm-prompt detail is count-aware
- locales: undo.removed gains {turns}; new undo.invalid_count, all 16 langs
- tests: tests/gateway/test_undo_rewind_session.py (6 cases)
The skill security scanner blocked legitimate community skills on three
intrinsic false-positive patterns:
- read_secrets_file matched `cat > file.env <<` heredocs (writing the
user's own keys into their own local .env), not just `cat file.env`
reads. Exclude output redirections.
- allowed-tools frontmatter is REQUIRED by the agent-skill spec; every
compliant skill declares it. Drop from HIGH privilege_escalation to a
LOW informational finding so it no longer drives the verdict.
- python_os_environ flagged `os.environ.get("CONFIG_VAR")` config reads
as HIGH exfiltration. Exempt non-secret `.get()` reads; add a dedicated
CRITICAL python_environ_get_secret pattern so secret-named reads
(OPENAI_API_KEY etc.) are still caught.
Also: scan_skill() now honors a skill-provided .skillignore / .clawhubignore
(gitignore-style) so dev/docs artifacts shipped in a skill root are excluded
from both structural checks and pattern scanning. SKILL.md is never ignorable.
80 tests pass (64 existing + 16 new).
The setup-mode chooser showed two bare labels ('Quick Setup (Nous
Portal) — OAuth login, model & messaging' / 'Full setup — configure
everything') that didn't explain what Quick Setup actually is. Expand
both labels inline so each choice line carries a concise explanation:
Quick Setup (Nous Portal) — free OAuth login, no API keys, model + tools
Full setup — configure every provider, tool & option yourself (bring your own keys)
Single-file change to the choice labels; no new plumbing.
Two pre-existing failures on main, unrelated to each other:
- test_model_catalog: website/static/api/model-catalog.json was stale vs
_PROVIDER_MODELS — minimax/minimax-m2.7 was renamed to minimax/minimax-m3
without regenerating the committed manifest. Ran scripts/build_model_catalog.py.
- test_gui_command: the macOS relaunchable-signing fixup
(_desktop_macos_relaunchable_fixup) makes two subprocess.run calls (xattr +
codesign) on darwin before launch. The two darwin GUI tests set
sys.platform='darwin' and mock subprocess.run with a 2-element side_effect
(pack + launch), so the fixup's calls drained the iterator -> StopIteration.
Mock out the fixup in those two tests so the subprocess accounting stays
focused on pack/launch.
The on_session_switch fan-out passed rewound=rewound unconditionally,
injecting rewound=False into every provider's **kwargs on the common
/resume, /branch, /new, and compression paths. Providers that capture
extra kwargs into an 'extra' dict (and the exact-dict-equality tests
guarding them) broke. Forward rewound only when truthy; /undo sets it
explicitly, everyone else stays clean.
Extends the existing /undo command from a single in-memory exchange
removal into a full rewind: back up N user turns (default 1), soft-delete
the truncated rows in SessionDB (active=0, kept for audit, hidden from
re-prompts and search), notify memory providers, and prefill the composer
with the backed-up message text for editing — CLI and TUI.
Reuses the SessionDB rewind primitives, the on_session_switch(rewound=True)
memory hook, and the TUI command.dispatch prefill payload from SaguaroDev's
#21910 work, wired to /undo [N] instead of a separate /rewind picker.
- cli.py: undo_last(n, prefill) — in-memory truncate + SQLite soft-delete
+ agent surgery (system-prompt invalidate, flush-index reset) + memory
notify + editable buffer prefill; /undo dispatch parses optional count;
checkpoint-rollback caller passes prefill=False
- tui_gateway/server.py: command.dispatch undo branch (was rewind) parses
count, picks Nth-from-last user turn, clamps to oldest
- commands.py: /undo gains [N] args_hint
- tests: rename + expand TUI suite (multi-turn, clamp, invalid-count)
- release.py: AUTHOR_MAP entry for SaguaroDev
Co-authored-by: SaguaroDev <74339271+SaguaroDev@users.noreply.github.com>
Adds the TUI half of the /rewind feature so the Ink terminal UI gets
the same affordance as the prompt_toolkit CLI.
Python side (tui_gateway/server.py):
- /rewind added to _PENDING_INPUT_COMMANDS so slash.exec rejects it
and the TUI falls through to command.dispatch (the only path with
access to live session state + memory hooks).
- New command.dispatch branch for name == "rewind":
v1 auto-picks the most recent user turn (Claude-Code-style single-
step undo), calls SessionDB.rewind_to_message, refreshes the
in-memory history, fires _memory_manager.on_session_switch with
rewound=True, and returns the new "prefill" payload.
- A dedicated picker overlay (multi-step rewind) is tracked as a
follow-up to #21910.
TS side (ui-tui/src/):
- New "prefill" variant on CommandDispatchResponse + asCommandDispatch
validator. Mirrors "send" but does NOT auto-submit; the client drops
the message into the composer for editing.
- createSlashHandler renders the optional notice via sys() and calls
ctx.composer.setInput(d.message), letting the user edit-and-resubmit
the rewound turn — the core UX promised by the issue.
Tests:
- 7 new tui_gateway tests covering prefill payload shape, in-memory
history truncation, DB soft-delete, memory-provider notification
(rewound=True), busy-session refusal, missing-session error, and
registry placement in _PENDING_INPUT_COMMANDS.
- Extended asCommandDispatch vitest covering the new prefill variant
(with + without notice, and rejection of malformed payloads).
Out of scope for v1 (tracked as #21910 follow-up):
- Dedicated picker overlay in Ink (the multi-step rewind UI). v1 auto-
picks the most recent user turn, matching the most common case.
- Gateway platforms (Telegram, Discord, etc.) — issue scopes v1 to
CLI + TUI only.
Self-review of the code-block masking fix: the cleanup path ran
media_pattern.sub('') over the _mask_protected_spans() copy of the text and
assigned that back to 'cleaned', so whenever a real MEDIA: tag was delivered
(if media: branch), every fenced code block / inline code / blockquote in the
reply was blanked to whitespace in the user-visible text.
Now mask only a length-equal copy of 'cleaned' to locate the real tag spans,
then delete those spans from the unmasked 'cleaned' — masking is a locator,
not a text rewrite. Protected spans survive verbatim. Strengthens the existing
mixed-code test (it only asserted 'Done.' survived, not the code block) and
adds an inline-code-survives regression test. Both fail on the old sub-based
code and pass now.
extract_media() scanned the full response text without distinguishing
live delivery tags from example paths in fenced code blocks, inline code
spans, and blockquotes. This caused false positives where the agent's
explanation of MEDIA: syntax (or tool output containing example paths)
was stripped from user-visible text and the path was added to the media
delivery list.
Added _mask_protected_spans() helper that replaces protected regions
with equal-length whitespace before regex matching, preserving match
offsets. The helper skips backtick-quoted paths in MEDIA: tags to
maintain existing path extraction behavior.
Fixes#35695
Self-review of #34375 fix: the cleanup path ran media_pattern.sub('') over
the JSON-masked copy of the text, which baked the masking spaces into the
user-visible 'cleaned' string — a serialized tool result like
{"old":"MEDIA:/x.png"} came back as {"old":" "}.
Now mask only a length-equal copy of 'cleaned' to locate the real tag spans,
then delete those spans from the unmasked 'cleaned'. Real tags are stripped;
JSON-embedded MEDIA: text reads back verbatim. Masking 'cleaned' (not the
original 'content') keeps offsets valid after the [[audio_as_voice]] /
[[as_document]] directives are removed. Adds two cleaned-text regression tests.
Serialized tool results frequently embed a prior reply's text, e.g.
{"result": "MEDIA:/path/stale.png"}. The bare-path branch of
MEDIA_TAG_CLEANUP_RE matched these and re-delivered stale files (#34375).
Adds BasePlatformAdapter._mask_json_string_media, which blanks (offset-
preserving) only MEDIA:<bare-path> tokens that sit inside a JSON value-
context string (opened by : , { or [). Legitimate tags at line start,
after prose, indented, MEDIA:"quoted" form, and two-line TTS output are
all left untouched.
Reworked from the approach in #34388 (a line-start regex anchor), which
no longer applied to current main and regressed same-line/indented tags.
Co-authored-by: kshitijk4poor <82637225+kshitijk4poor@users.noreply.github.com>
Adds nicsequenzy@gmail.com -> polnikale to AUTHOR_MAP so the
check-attribution gate passes for the Playwright headless_shell browser
discovery fix (#35717).
* feat(desktop): drop files anywhere in the chat area
File drops were only wired to the composer input. Add a reusable
useFileDropZone hook (enter/leave depth counting + capture-phase reset so
the affordance clears even when the composer claims the drop) and a
pointer-events-none ChatDropOverlay, wired onto the conversation viewport.
Drops funnel through the existing onAttachDroppedItems; composer drops keep
their own inline-ref behavior.
* fix(desktop): chat-area drops insert inline @file refs, not attachment cards
Match the composer-input drop behavior — funnel dropped paths through
droppedFileInlineRef + the composer insert bus so they render as inline
ref chips instead of attachment cards.
* fix(desktop): don't render bare file paths as tool images (404)
vision_analyze reports its input image as a local filesystem path, which
toolImageUrl handed straight to <img src>. In the renderer that resolves
against the dev-server origin and 404s. Restrict inline tool images to
fetchable sources (data: URLs and remote http(s)); bare paths now fall
back to the tool's codicon.
When an unauthenticated SPA fetch hit a gated /api/* endpoint (e.g.
GET /api/analytics/models?days=30 fired from ModelsPage on mount or
after a session expiry), the gated middleware stamped the request's
own path into next= on the 401 envelope's login_url. The SPA's global
401 handler in web/src/lib/api.ts full-page-navigated to that URL,
the PKCE cookie carried the encoded /api/* value through the OAuth
round trip to Portal, and /auth/callback's _validate_post_login_target
accepted it as same-origin and redirected the user to the raw JSON
endpoint instead of the dashboard.
Symptom Ben reported: after the OAuth screen he kept landing on
$DOMAIN/api/analytics/models?days=30 (raw JSON) rather than /models.
The bug was deterministic per page — whichever /api/* call ModelsPage,
AnalyticsPage, or SessionsPage fired first owned the redirect race.
Fix: both validators now reject /api/* targets in addition to the
existing /login, /auth/, /api/auth/ exclusions:
- _safe_next_target in middleware.py drops the value before it ever
enters login_url, so the SPA's 401 handler navigates to a bare
/login (which the SPA itself can return-from via its own
sessionStorage["hermes.lastLocation"] fallback that was already
saving the actual browser location).
- _validate_post_login_target in routes.py drops it as second-line
defence at the callback boundary, so a legacy cookie, a regressed
middleware, or an attacker-crafted /auth/login?next=/api/... value
can't smuggle the redirect through. Either layer alone is enough;
pairing them means a regression in one is caught by the other.
The match is anchored: ``decoded == "/api"`` or
``decoded.startswith("/api/")``. SPA route lookalikes like /apidocs
or /api-keys remain valid landing targets — tests pin that.
Test additions in test_dashboard_auth_401_reauth.py:
- TestApi401Envelope: rewrote test_login_url_carries_next_for_deep_
api_path (which asserted the pre-fix behaviour) as
test_login_url_drops_next_for_deep_api_path, plus added the
specific analytics-models repro case from Ben's report.
- TestNextSameOriginValidation: rejects-api-paths + does-not-reject-
api-prefix-lookalikes (covers /apidocs, /api-keys).
- TestAuthCallbackNext: end-to-end test_callback_with_api_next_
lands_at_root drives /auth/login?next=/api/... through to the
callback and asserts the user lands at "/", not the API URL.
- TestValidatePostLoginTarget: new class covering the callback-side
validator directly, including the URL-encoded ``%2Fapi%2F...``
form the PKCE cookie actually carries.
Mutation-tested: reverting both validators causes exactly the 5 new
or rewritten /api/*-related assertions to fail (each fix layer is
independently tested), while the 31 other assertions in the file
remain green. Full tests/hermes_cli/ suite (288 files, 5,938 tests)
passes with the fix applied.
The desktop rename dialog sent PATCH /api/sessions/{id}, but the backend
only defined GET and DELETE for that path — FastAPI returned 405 Method
Not Allowed, surfaced to the user as "Rename failed". Add the PATCH route
backed by SessionDB.set_session_title (handles sanitization, uniqueness,
and clearing the title when empty).
Also fix a misleading notification: any 405 was summarized as an unrelated
"does not support that audio endpoint" message. Make it a generic 405 hint.
The targeted data-volume chown in stage2-hook.sh only covers hermes-owned
*subdirectories*; loose state files living directly under $HERMES_HOME
(auth.json, state.db, gateway.lock, gateway_state.json, …) are missed.
When created or rewritten by `docker exec <container> hermes …` (root
unless `-u` is passed) they land root-owned, and the unprivileged hermes
runtime then hits PermissionError on next startup, producing a gateway
restart loop.
Fix: reset ownership of an explicit allowlist of hermes-owned top-level
files on every boot. The list mirrors the top-level file entries of
hermes_cli.profile_distribution.USER_OWNED_EXCLUDE plus the runtime lock
files.
This uses a targeted allowlist rather than the originally-proposed blanket
`find $HERMES_HOME -maxdepth 1 -user root` sweep, preserving the
targeted-ownership contract from #19788 / PR #19795: a bind-mounted
$HERMES_HOME may contain host-owned files Hermes does not manage, and
those must never be chowned. Verified end-to-end: allowlisted root-owned
files are reset to hermes on restart while a non-allowlisted host file
keeps its root ownership.
Co-authored-by: x1am1 <2663402852@qq.com>
Shiki's github-light-default colors comments #6e7781 (~4.2:1 on the code
card background), which is borderline unreadable at the 11px code font
size — and worst for shell snippets, where a single `#` turns the rest
of the line into one long comment span. Remap light-mode comments to
GitHub's darker muted gray (#57606a, ~6.4:1) via per-theme
colorReplacements. Dark mode (~6.1:1) reads fine and is left untouched.
s6-overlay images (e.g. hermes-agent:latest) use /init as PID 1 and exec
/run/s6/basedir/bin/init during stage0 startup. The Docker terminal backend
unconditionally added Docker --init and mounted /run as noexec, which broke
those images in two ways: --init created a second competing PID-1 init, and
the noexec /run made s6 stage0 fail with "exec: /run/s6/basedir/bin/init:
Permission denied" (exit 126), so the container died and terminal commands
reported a generic "container is not running" error.
Detect images whose entrypoint is /init via 'docker image inspect' and, for
those images only, skip Docker --init and mount /run with exec. All other
images keep the hardened --init + noexec defaults. Detection is best-effort:
any inspect failure falls back to the safe defaults.
#34192 reports Hostinger's 'Hermes WebUI' catalog crashes on startup
with:
/usr/bin/tini: No such file or directory
The image moved from tini to s6-overlay as PID 1 (/init) earlier in
2026. Orchestration templates that still pin /usr/bin/tini as the
entrypoint \u2014 like the Hostinger Hermes WebUI catalog \u2014 have no
binary to exec and the container crashes immediately.
Hermes has no control over the Hostinger catalog template, but we can
make the image backward-compatible by symlinking /usr/bin/tini -> /init
during the s6-overlay install step. External wrappers that exec
/usr/bin/tini will land on the same s6-overlay reaper they would have
landed on if they'd used the canonical /init entrypoint.
The image's own ENTRYPOINT continues to be /init verbatim \u2014 the shim
is purely for legacy external wrappers, not for the image's own
runtime path. Once affected catalogs are updated, the symlink can be
removed.
Other issues #34192 raises that are NOT addressed by this PR:
* Problem #2 (UID 1024 vs 10000 mismatch): already fixed by #33148
(S6_KEEP_ENV=1) and #32412 (with-contenv shebangs). The Hostinger
template likely needs to update its env-var propagation.
* Problem #3 (incompatible session formats): RFC for pluggable
SessionDB is tracked in #23717.
* Problem #4 (Telegram polling conflict): an operations problem on
Hostinger's side, not in this codebase.
This PR is scoped to the one issue that can be fixed inside
Dockerfile: the missing /usr/bin/tini binary.
Tests (3 in test_dockerfile_tini_compat_shim.py):
- test_tini_compat_symlink_present
Guard: the symlink line must exist in Dockerfile.
- test_tini_compat_comment_explains_why
The #34192 anchor comment must be present so future readers know
why the shim is there (avoid accidental removal).
- test_entrypoint_still_init_not_tini
Sanity check: ENTRYPOINT remains /init (s6-overlay). The shim is
only for external wrappers.
Refs: #34192
Partial fix: addresses the immediate tini-binary crash. Catalog-side
fixes still needed by Hostinger for the UID and session-format
problems documented in the issue.
Co-authored-by: Cursor <cursoragent@cursor.com>
Fixes#34107. When Hermes runs in Docker with HERMES_UID=1000 /
HERMES_GID=911, the entrypoint chowns the top-level HERMES_HOME once
at startup — but subdirectories created at runtime by
ensure_hermes_home() (especially for profile namespaces under
profiles/<name>/ spawned by kanban workers) were landing as root:root
and blocking subsequent uid-mapped worker invocations with:
PermissionError: [Errno 13] Permission denied:
'/opt/data/profiles/charles/logs/curator'
Fix: add _resolve_hermes_uid_gid + _chown_to_hermes_uid helpers that
read the env vars and apply chown after mkdir. Invoke from _secure_dir
which already runs after every directory creation in the home-init path,
so all newly-created subdirs (including the profile namespaces) get the
right ownership.
Safety properties:
- No-op when HERMES_UID/HERMES_GID unset (the dominant non-Docker path)
- No-op on Windows (os.chown doesn't exist; AttributeError swallowed)
- No-op when running as non-root (EPERM swallowed — the entrypoint's
startup chown -R picks it up on next restart, and in most cases the
dir was already correctly-owned by the calling user)
- Uses -1 sentinel for missing field so only the set value applies
- Empty-string env vars treated as unset
Adds 14 tests across:
- TestResolveHermesUidGid (7) — env-var parsing
- TestChownToHermesUid (5) — chown helper invariants
- TestSecureDirChown (2) — end-to-end through _secure_dir
Co-authored-by: Cursor <cursoragent@cursor.com>
Add MiniMax-M3 to the minimax, minimax-oauth, and minimax-cn curated
lists (these are hardcoded — the native Anthropic-format endpoint has no
/v1/models listing and the providers aren't in _MODELS_DEV_PREFERRED, so
new models don't auto-pull). Add a DEFAULT_CONTEXT_LENGTHS key
'minimax-m3' -> 1,000,000 so M3 resolves to its 1M context on every
surface (native ID + OpenRouter/Nous slug) via longest-key-first
substring match, while the M2.x series stays at 204,800.
On macOS the desktop app is built locally and ad-hoc signed (no Developer ID
on the user's machine). An ad-hoc bundle has no stable Designated Requirement,
so when the self-updater rebuilds it in place with a fresh build (new cdhash)
— plus the com.apple.quarantine flag inherited from the downloaded installer
process chain — Gatekeeper/LaunchServices treats the changed code as tampering
and macOS reports "Hermes is damaged and can't be opened," and the app fails to
relaunch. First launch works (fresh registration); the in-place update relaunch
is what breaks.
Fix: after building the desktop app locally, strip quarantine xattrs and
re-apply a clean deep ad-hoc signature (omitting the hardened-runtime flag,
which an ad-hoc build can't satisfy). Applied in both build entry points:
- hermes_cli/main.py cmd_gui (the `hermes desktop --build-only` path the
updater drives) — so the fix ships via `hermes update` (git), no installer
re-download needed.
- scripts/install.sh install_desktop (first install) for parity.
Both are no-ops on non-macOS and when a real signing identity (CSC_LINK /
APPLE_SIGNING_IDENTITY) is configured, so signed/notarized builds are untouched.
The docker_forward_env build loop only consulted the ~/.hermes/.env disk
fallback when a key was unset (value is None), not when it was present
but empty (""). A transient empty value in os.environ was therefore
forwarded into the sandbox container as `-e KEY=`, clobbering the correct
value on disk. Sandboxed workloads then read a zero-length secret and
failed auth (observed as intermittent Linear API 401s) with no gateway
restart and no .env rewrite.
Treat empty-string like unset (`if not value:` on the fallback) and never
forward a blank secret (`if value:` on the guard).
Fixes#35580
Adds me@simontaggart.com → SiTaggart to AUTHOR_MAP so the
check-attribution gate passes for the docker_forward_env empty-secret
fix (#35583, fixes#35580).
Read the Portal's tool_access claim (JWT + /api/oauth/account) into NousToolAccessInfo and gate managed Tool Gateway access on it: tool_gateway_entitled (paid OR live pool) and per-category tool_gateway_entitled_for(). The pool funds web/image/tts/browser but not video, so per-backend availability, the charge picker (ensure_nous_portal_access coverage_category), and managed defaults all respect coverage.
Setup: rebuild prompt_enable_tool_gateway as a per-tool checklist that renders whenever the pool is enabled, lists only pool-covered tools (video excluded for free-pool users), and is framed as the free tool pool for $0 subscribers rather than a paid subscription. get_gateway_eligible_tools now gates and filters off the entitlement snapshot.
The thin installer (apps/bootstrap-installer) drives install.sh stage-by-stage,
each in its own process. The `desktop` stage never called check_node, so the
Hermes-managed Node provisioned earlier (at $HERMES_HOME/node/bin) wasn't on
PATH. install_desktop's `command -v npm` check then failed and the build was
skipped — yet the stage still reported {"ok":true,"skipped":false}, so the
installer showed "Installation Complete" and only failed at the end with
"Couldn't find a built Hermes desktop ... the desktop build step may have been
skipped or failed."
Fix:
- Call check_node in the `desktop` stage (mirrors every other Node-dependent
stage) so the managed Node is on PATH (or installed).
- Make install_desktop self-provision via check_node and hard-fail (return 1)
if npm is still unavailable, instead of a silent `return 0`. The desktop
stage only runs when a build is explicitly requested (--include-desktop), so
an unavailable toolchain is a real failure, not graceful degradation.
Verified on macOS arm64: the `desktop` stage now builds
release/mac-arm64/Hermes.app, which matches resolve_hermes_desktop_exe, so the
installer's "Launch Hermes" succeeds.
