The `text_to_speech` tool schema accepted only `text` and `output_path`,
so style direction (tone, emotion, pacing, whispering) could never reach
the OpenAI backend — even though `gpt-4o-mini-tts` (Hermes's OpenAI
provider default) treats `instructions` as its primary voice-design
control.
This plumbs an optional `instructions` argument through the tool schema,
the handler lambda, and `text_to_speech_tool()` into
`_generate_openai_tts`, where it is forwarded to
`client.audio.speech.create()` only when truthy. Empty/None values still
omit the key entirely, preserving behavior on `tts-1`/`tts-1-hd` and
strict OpenAI-compatible servers.
The same passthrough unblocks self-hosted OpenAI-compatible voice-design
servers (Qwen3-TTS-VoiceDesign on oMLX, etc.) that are already wired in
via `tts.openai.base_url` — the established convention per #9004 and the
TTS config docs — without inventing a new provider backend.
Tests: `tests/tools/test_tts_instructions.py` covers backend passthrough,
tool-level threading, schema declaration, and the empty-string/absent
omission cases. `tests/tools/test_tts_max_text_length.py` fake_openai
signature widened to accept the new kwarg.
Refs NousResearch/hermes-agent#14196
The Buzz Desktop section covered discovery only. The combination that
actually needs stating: the hermes-acp toolset carries terminal and
execute_code, and buzz-acp answers session/request_permission itself with
allow_once instead of surfacing it. A Hermes agent in Buzz runs shell
commands on the host unattended.
Buzz defaults every agent to owner-only and that default holds through to
the spawned process env, so nobody reaches the open state by accident.
But Anyone is one dropdown change away with no warning shown, and it
hands channel-wide shell access to the host.
Also record that the two obvious mitigations do not work: approvals.mode
manual raises the request but Buzz auto-approves it anyway, and
platform_toolsets.acp does not narrow the ACP toolset. Both verified by
running rm -rf through the ACP path under each setting.
Amend the Approvals section too — it promised prompts route back to the
editor, which is only true for hosts that choose to surface them.
en + zh-Hans.
Signed-off-by: SHL0MS <SHL0MS@users.noreply.github.com>
Existing installs predate the install.sh hermes-acp launcher, and hermes
update never re-runs setup_path, so ACP hosts (Zed, JetBrains, Buzz)
still resolve Hermes as unavailable until a reinstall. _ensure_acp_launcher()
writes the launcher next to an existing hermes command in ~/.local/bin or
/usr/local/bin during hermes update — delegating to the sibling launcher so
it is correct for every install layout. Never follows symlinks (#21454),
skips unwritable dirs, no-op on Windows (venv Scripts is already on PATH).
Docs: add a Buzz Desktop section to the ACP page (en + zh-Hans).
Per-job cron inference pins are now user-owned: the agent-facing cronjob
tool schema no longer exposes model/provider/base_url, and the registered
handler ignores them even if a model hallucinates the old parameters.
Users set pins via the dashboard, hermes cron create/edit --model/--provider,
or jobs.json directly — and once set, a pin sticks until the user changes it.
Existing agent-era pins are grandfathered untouched.
New cron.model / cron.model_provider config keys give the cron fleet its
own default model, independent of the chat model. Fire-time resolution:
per-job pin > cron.model > HERMES_MODEL > model.default. An axis covered
by the explicit cron-fleet default is deliberate routing, not drift, so
the #44585 fail-closed guard skips it — switching your chat model with
/model or hermes model no longer breaks unpinned cron fleets.
- tools/cronjob_tools.py: drop model param from agent schema + handler;
remove now-dead _resolve_model_override
- cron/scheduler.py: cron.model/model_provider resolution + per-axis
drift-guard skip
- cron/jobs.py: snapshot resolution mirrors the new precedence
- hermes_cli/subcommands/cron.py + hermes_cli/cron.py: --model/--provider
on hermes cron create/edit
- hermes_cli/config.py: cron.model / cron.model_provider defaults
- docs: cron.md model-resolution tip rewritten
The grant_spent notice fired for every subscription user with top-up
funds the moment their cap was reached and camped in the CLI/TUI status
bar and desktop toasts with no action to take — the account keeps
working off top-up. Remove it everywhere:
- agent/credits_tracker.py: drop grant_cond + the emit/clear block;
the dev fixture state now (correctly) produces no notice
- TUI: keep the turn-start clear of credits.grant_spent as back-compat
for older backends that still emit the key
- Desktop: drop the demo step and stale comment references
- Docs/config comments: remove grant-spent from credits_notices text
- Tests updated: policy/cold-start now assert the key never fires
Usage bands, depleted, and restored notices are unchanged; /usage still
reports the full balance breakdown.
An open chat window took a session-cap slot at session.create/resume time.
Every desktop tile paint and every background reconnect-resume opens one, so
on a websocket-flappy host they accumulated: five parked desktop tabs filled a
5-slot cap and locked the messaging gateway (which shares the cap) out for
fourteen minutes while running no agents at all.
A slot held that way is invisible everywhere. An unprompted draft has no DB row
and the sidebar filters it out with min_messages=1, so the only way to diagnose
it was reading runtime/active_sessions.json by hand.
Claim on the first turn instead, mirroring the lazy contract
_ensure_session_db_row already uses for the row itself. Capacity now means an
agent can run rather than that a window exists, and anything holding a slot is
something the user can see.
Also reclaim leases whose session skipped teardown. _prune_dead only fires when
the owning pid dies, and a dashboard/serve backend runs for days, so a leaked
lease was held until restart. The owning process reconciles against the leases
it still holds, which is exact and needs no heartbeat write on the turn path.
- Change await self._send_read_receipt to asyncio.create_task to avoid
blocking message dispatch on slow bridge responses (matches BlueBubbles
pattern). Up to 5s per-message latency eliminated.
- Update tests: assert_called_once_with instead of assert_awaited_once_with
since the receipt is now scheduled, not directly awaited.
- Add send_read_receipts documentation to whatsapp.md following the
BlueBubbles docs pattern.
When an operator changes the global model/provider config, warn that
unpinned cron jobs with stored snapshots will fail-closed on their next
run. Adds a cron.model_drift_guard config opt-out (default true) for
fleets that should deliberately track changing global defaults.
Addresses #59031. Original PR #59177 by @doncazper.
Figma's mcp.figma.com register endpoint is a client_name allowlist
(Claude Code / Codex succeed; Hermes Agent 403s) and returns a client
secret while advertising auth_method=none, then requires the secret on
token exchange. Auto-set client_name + client_secret_post for Figma
hosts, pass oauth cfg through login/add paths, force interactive OAuth
for hermes mcp login from non-TTY desktop shells, and ship a catalog
entry. Proven: hermes mcp login figma → 26 tools.
Gating this behind an opt-in was the wrong call. A dev server already
executes arbitrary local JS — vite's module graph, every postinstall in
node_modules — so a loopback debugging port does not meaningfully widen
what a `npm run dev` session can already do, and `perf:serve` has opened
one unconditionally all along.
Requiring the variable also defeated the point: the tooling exists to be
reached for mid-task, and a capability you must remember to enable before
launching is one you don't have when you need it.
So the port opens on 9222 — the same port scripts/eval.mjs and
scripts/perf/lib/cdp.mjs already default to — for any dev-server run.
HERMES_DESKTOP_CDP_PORT stops being an on-switch and becomes an
override: a different port, or `off` to disable.
The hard gate is unchanged and still checked first: a packaged build
never opens the port, and no env value talks it into it. Neither does an
unpackaged `electron .` against dist/, which is how the packaged app
gets smoke tested.
Refusals only log when they contradict something the developer asked for
(a typo'd port, an explicit `off`). Packaged and dist runs are closed by
design and stay quiet.
Saying EXACTLY a configured stop phrase (default: 'stop') and nothing
else now ends the voice conversation instead of being sent to the agent
as a prompt. Match is deliberately strict — whole utterance,
case-insensitive, surrounding punctuation stripped — so 'stop doing
that and try X' still reaches the agent.
- tools/voice_mode.py: is_voice_stop_phrase() + voice.stop_phrases
config loader (default ['stop'], [] disables, malformed config falls
back safely).
- hermes_cli/voice.py: shared continuous loop (TUI + desktop) halts on
a stop phrase exactly like the silent-cycle limit (fires
on_silent_limit so every UI turns voice off); stop_continuous
force-transcribe path swallows the phrase without counting a silent
cycle.
- cli.py: classic CLI push-to-talk/continuous path and barge-in
utterance path disable voice mode on a stop phrase.
- Config default voice.stop_phrases: ['stop']; docs updated.
Tests: tests/tools/test_voice_stop_phrase.py (27 tests,
sabotage-verified: loop test fails when detection is disabled);
voice suites green (110 + 44 passed).
Whisper auto-detection frequently misidentifies short/accented clips,
which users experience as voice notes transcribed in the wrong language
(Teknium + CTO both hit this). The unified resolver from #73067 made a
global hint possible; this makes it the DEFAULT so stock installs stop
guessing. Non-English users set stt.language once; '' restores
auto-detect for multilingual use.
Deep-merge gives existing configs the new default automatically (no
_config_version bump needed); any explicit per-provider or global
language setting still wins.
The renderer is a Chromium page, and apps/desktop already carries a whole
CDP toolkit for it — scripts/eval.mjs, scripts/perf/lib/cdp.mjs with its
shared SELECTORS map, and the diag-*/probe-* family. None of it can
attach to `hgui` or `npm run dev`, because neither passes
--remote-debugging-port. The only launcher that opens one is
`npm run perf:serve`, which is a separate isolated instance rather than
the app you're looking at.
Add HERMES_DESKTOP_CDP_PORT. When set, the shell opens a CDP port on
loopback so that existing tooling can read the live DOM: computed
styles, geometry, which rule actually won.
Three independent gates, all required, resolved by a pure function in
electron/dev-cdp.ts so the policy is testable without an Electron app:
1. not packaged — a shipped build never opens the port, and this is
checked first so no env combination can talk it into doing so;
2. HERMES_DESKTOP_DEV_SERVER present — an unpackaged `electron .`
against dist/ is how the packaged app gets smoke tested, so it
behaves like the packaged app here;
3. the port explicitly requested and a valid integer.
Default `npm run dev` is unchanged and silent: no port, no nag. An
opt-in that gets refused always logs why, so nobody loses an hour
wondering what isn't listening.
The address is pinned to 127.0.0.1 rather than left to Chromium's
default, and is deliberately not configurable — there's no reason to
expose a renderer debugger off-host and offering the knob invites
someone to try.
scripts/eval.mjs hardcoded :9222 and threw a raw ECONNREFUSED stack when
nothing was there. It now honours the same variable and explains itself.
Class-level fix for the 'STT transcribes the wrong language' issue family
(#55551, #50181 and siblings). Previously language handling was per-provider
chaos: local honoured stt.local.language, Groq/OpenAI/Mistral/DeepInfra sent
no language hint at all, xAI silently forced 'en', ElevenLabs used its own
language_code key, and there was no global setting.
- New _resolve_stt_language() helper: stt.<provider>.language >
stt.language (new global key) > HERMES_LOCAL_STT_LANGUAGE > auto-detect.
- Threaded through ALL providers: local, local_command, groq, openai,
mistral, xai, elevenlabs, deepinfra (shared OpenAI handler), command
providers, and plugin dispatch.
- xAI no longer forces English when nothing is configured (auto-detect).
- Mistral Voxtral now receives a language hint when configured.
- stt.groq.model is now honoured from config (previously env-only).
- DEFAULT_CONFIG gains stt.language, stt.groq, stt.xai, stt.mistral.language.
- Tests: tests/tools/test_stt_language_resolution.py (11 tests, sabotage-
verified) + full transcription suite green (236 passed).
Builds on cherry-picked contributor work from #19786 (@zombopanda),
#23161 (@materemias), #50684 (@BlackishGreen33).
- Normalize stt.groq.language: cast to str, strip, treat
empty/whitespace as unset (parity with xAI's str().strip()).
- Clarify "blank = auto-detect" inline comments in 4 docs/configs to
reflect the env-var fallback (HERMES_LOCAL_STT_LANGUAGE).
- Document that HERMES_LOCAL_STT_LANGUAGE also drives the local
faster-whisper provider, not just the CLI fallback.
- Add unit tests covering: omitted language when unset, config-supplied
language, env fallback, config-over-env precedence, whitespace
normalized to unset.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Read stt.groq.language from config.yaml (with HERMES_LOCAL_STT_LANGUAGE
env fallback) and forward it to the Groq Whisper API to skip
auto-detection on known-language audio. Omit when unset so Groq
auto-detects, preserving today's behavior. Bonus: swap xAI's hardcoded
env literal for the LOCAL_STT_LANGUAGE_ENV constant for consistency.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add optional stt.openai.language config for OpenAI transcription. Forward non-empty hints to the API while preserving auto-detection when unset. Document the config-only setting, add its default, and cover configured and unset request arguments.
Three mechanisms to detect and notify when gateway sessions stall silently:
1. Mid-turn activity heartbeats stamped to SessionDB so hermes sessions list
and hermes status show progress during long turns without new message rows.
2. Stall watchdog: when a busy session has pending inbound and the shared
activity clock is idle past agent.session_stall_timeout (default 300),
log a WARNING and notify the user once to try /new. Notify-only; does
not kill the turn.
3. Compaction timeout: fenceless compress_context callers get a progress-aware
host budget (compression.context_timeout_seconds default 120 idle,
compression.context_total_ceiling_seconds default 600 ceiling). On timeout,
cancel via commit fence, skip compaction without dropping messages, and
continue the turn.
Closes#72016 (slices 1-3; slice 4 cumulative SSE stream-retry deadline
remains a follow-up).
Cherry-picked from PR #72424 by @fangliquanflq.
Per Teknium: the caps should bound a single agent loop, not accumulate
over the whole session. Rename SessionCapConfig -> LoopCapConfig and the
config section session_caps -> loop_caps; move the counters into
reset_for_turn (invoked per turn via turn_context) so each turn starts
with a fresh budget; retune defaults 200 -> 50 (a single turn issuing 50
web searches / spawning 50 subagents is already pathological). Block
codes session_*_cap -> loop_*_cap and messages updated to drop the
/new-resets-the-budget guidance (irrelevant now that it resets per turn).
Tests flipped: the old persists-across-turn-resets assertion becomes
resets-each-turn.
Add per-session lifetime caps on web_search calls and subagent spawns
(defaults 200/200, matching Claude Code v2.1.212). Unlike the existing
per-turn tool-loop guardrails, these count over the whole session and
reset only when a fresh agent is built (/new, /clear). Hitting a cap
blocks the offending call and halts the turn cleanly.
- agent/tool_guardrails.py: SessionCapConfig + session counters on the
controller (in __init__, not reset_for_turn, so they persist across
turns). before_call() enforces caps first, independent of
hard_stop_enabled. delegate_task batches count each task.
- hermes_cli/config.py: tool_loop_guardrails.session_caps defaults.
- docs + tests (unit + E2E validated against a real AIAgent).
Inspired by Claude Code 2.1.214, which added permission prompts for
container-CLI commands (including the Podman shim) carrying
daemon-redirect flags (--url, --connection, --identity, remote mode)
that previously ran without one.
A daemon redirect makes a local-looking command operate on a different
(often remote) daemon, silently acting on production infrastructure.
Any container-CLI invocation carrying a redirect now requires approval
regardless of subcommand:
- -H/--host and --context global flags (value required, global-flag
position only — bare -h help and run-level -h <hostname> stay allowed)
- context use (persistently switches the default daemon)
- podman --url/--connection/--identity and -r/--remote
- DOCKER_HOST=/DOCKER_CONTEXT=/CONTAINER_HOST=/CONTAINER_CONNECTION=
environment prefixes
Sibling-site widening: the existing container lifecycle rules matched
only the verb directly adjacent to the binary name, so a global flag or
a compose -f file flag slipped past the guard, and the legacy hyphenated
compose binary was never covered. They now tolerate global flags — the
same treatment the 'hermes ... gateway' rule already has — and match the
hyphenated compose binary.
Validation: 33 new tests; 339 pass in test_approval.py + new file;
442 pass across the adjacent guard suites; E2E battery of 12 dangerous
+ 17 safe commands via real imports, hot path ~330us/call.
Widen the cherry-picked /diff base (#4839 by @SHL0MS) into one
cross-surface implementation, folding in the review feedback and the
best ideas from the two sibling PRs (#22703, #53527):
- tools/working_diff.py: shared git collection layer — unstaged
(default), staged, and all (vs HEAD) modes; untracked files folded in
via `git diff --no-index` so new files appear as additions (Codex
/diff parity); shlex-split arguments preserve quoted paths.
- CLI: handler moved to hermes_cli/cli_commands_mixin.py per the
current god-file decomposition (dispatch stays in cli.py), renders
through the rich console with a 400-line terminal-flood guard.
- Gateway: _handle_diff_command in gateway/slash_commands.py + dispatch
in gateway/run.py; fenced ```diff output truncated to 60 lines /
3000 chars before the platform senders apply their own per-platform
message clamps (tool-progress-style layered truncation). Localized
strings in all 17 locale catalogs.
- /diff session (from #53527): cumulative checkpoint-baseline diff of
everything Hermes changed, via new CheckpointManager.session_diff();
docstring records the retained-baseline approximation caveat from
review. Works on both surfaces; degrades with an actionable message
when checkpoints are off.
- Slack: /diff routed via /hermes diff (50-slash cap; keeps
telegram-parity test green and /version native).
- Registry: cross-surface CommandDef with staged|all|session
subcommands; docs: slash-commands reference (CLI + gateway tables +
both-surfaces list) and hermes-agent skill reference.
- Tests: tests/tools/test_working_diff.py (real git repos),
tests/hermes_cli/test_diff_command.py (real git + stubbed checkpoint
manager), tests/gateway/test_diff_command.py (end-to-end handler,
real checkpoint store), TestSessionDiff in
tests/tools/test_checkpoint_manager.py.
Salvaged from the /diff PR cluster #4839 + #22703 + #53527.
Co-authored-by: Ninso112 <ninso112@proton.me>
Co-authored-by: Harshkamdar67 <harshkamdar67@gmail.com>
Extends the cherry-picked /context command (PR #52184) and prompt-size
attribution helpers (PR #66656) into one visual context view across
surfaces, and absorbs the per-component budget-visibility goal of the
/tokens proposal (PR #48470):
- agent/context_breakdown.py: pure renderers over the existing payload —
a 5x20 glyph block grid (1 cell ~= 1% of the model window), an
'Estimated usage by category' table with free space, and expanded
per-skill / per-toolset listings via compute_context_details(), which
reuses the prompt-size attribution mechanism (skills index-line bytes +
registry tool->toolset map) converted to the same chars/4 heuristic.
- cli.py: /context [all] renders grid + category table (+ expanded
listings) from the live agent and in-memory conversation history.
- gateway/slash_commands.py: /context appends the plain-text category
table (no grid — monospace not guaranteed on messaging platforms);
/context all adds the expanded listings. Fail-open: breakdown errors
never break the gauge.
- hermes_cli/commands.py: /context gains the 'all' subcommand; /version
demoted to /hermes version on Slack to keep the 50-slash cap.
- tests: renderer unit tests against synthetic payloads, registry test,
gateway /context + /context all + failure-degradation handler tests.
- docs: slash-commands reference + CLI guide entries.
Read-only and locally computed: no provider calls, no prompt-cache impact.
Co-authored-by: RemyFevry <29257684+RemyFevry@users.noreply.github.com>
Co-authored-by: joelbrilliant <joelbrilliant1@gmail.com>
Co-authored-by: CharlesMcquade <6466275+CharlesMcquade@users.noreply.github.com>
The Nous Portal docs claimed routing 'happens through OpenRouter under
the hood' with OpenRouter-equivalent failover, and that the catalog
'mirrors OpenRouter's model list'. That is not the Portal's contract:
some models route through OpenRouter, others through proprietary or
secondary providers, and per-model routing can change over time.
The stale wording licensed users to expect OpenRouter-proprietary
request extensions (top-level cache_control, session_id sticky
routing, provider preferences) to work through the Portal, producing
misfiled bug reports like #71576. Reworded both pages (en + zh-Hans)
and added an explicit note that OpenRouter-specific extensions are not
part of the Portal API contract.
After N consecutive guardian denials in a session the deny message escalates to a hard-stop instruction. Inspired by ChatGPT Work auto-review circuit breaker.
Maps CLAUDE.md/AGENTS.md, permission allowlists, MCP servers, skills, and memories into their Hermes equivalents. Follows the openclaw migration pattern. Inspired by ChatGPT Work import-from-another-agent onboarding.
The default max_iterations/agent.max_turns budget was set when long
agentic runs were rare; complex tasks now routinely exceed 90 tool
calls. Raise the default to 500 across every surface that hardcodes
the fallback: AIAgent constructor, DEFAULT_CONFIG, CLI resolution
chain, gateway env bridge, cron scheduler, and TUI gateway. Explicit
user config values are unaffected (deep-merge preserves them; no
_config_version bump needed).
Docs (en + zh-Hans), CLI help text, tips, and pinned tests updated
to match.
Renaming a Telegram bot's @username in BotFather silently stopped the
gateway from answering in groups.
PTB caches getMe() in Bot._bot_user and only rewrites it inside get_me(),
so after a rename the adapter kept comparing mentions against the OLD
handle. The exclusive-mention gate then saw the new @handle, failed to
match itself, and concluded the message was addressed to a different bot
— dropping it before the reply and wake-word fallbacks could run. Native
replies to the bot were discarded too. Polling mode recovered on the next
90s heartbeat; webhook mode never calls get_me() again, so it stayed dead
until restart.
Separately, the bot-handle pattern assumed every bot username ends in
"bot". Collectible (Fragment) usernames can be assigned to bots and drop
that suffix (@jarvis, @pic), so such a bot could not recognise its own
handle in the entity-less fallback and was suppressed by any message that
also named another bot.
- Route every mention comparison through _current_bot_username(), which
prefers the last observed handle over PTB's cache.
- Learn the live handle from inbound updates: Telegram stamps the current
username on our own messages and on reply_to_message. Guarded by user
id, so another account's handle is never adopted.
- Re-check identity out of band (TTL-bounded, one getMe per 5 min) when
the exclusive gate is about to drop a message — the exact stale-handle
symptom — so the mistake self-corrects instead of persisting.
- Refresh identity in webhook mode via a dedicated low-frequency loop,
cancelled on the same teardown fence as the heartbeat.
- Match our own handle by identity rather than shape. Foreign handles keep
the deliberate "...bot" narrowing so human @handles still never act as
routing hints (the intent behind ce4d857021).
Validation: 11 regression tests; sabotage runs confirm each behavioral
test fails with the fix reverted. 157 tests green across the Telegram
gating, reconnect, and topic-mode suites.
The include/exclude filter matched exact names only — glob-style entries
('*_radar_*') silently matched nothing, so a Cloudflare flat-mode config
meant to trim 3,320 tools to ~1,900 actually registered 3,319. Unmatched
patterns produced no warning.
- matches_name_filter(): exact membership first (O(1) for literal lists),
then fnmatch.fnmatchcase for entries containing * ? [ — same pattern
semantics as approvals.deny. Entries without metacharacters stay
strictly literal ('docs' never matches 'docs_search').
- _should_register() uses it for both include and exclude (symmetric)
- hermes mcp tools picker (mcp_config.py) pre-selection uses the same
matcher so the UI agrees with runtime registration
E2E against the live Cloudflare capture with the real exclude list:
3,320 -> 1,905 surviving (1,415 excluded); radar/DLP gone,
purge_cache/dns_records kept. 220/220 mcp_tool tests (4 new).
Teknium review changes on the tiered policy:
1. threshold_pct default 10 -> 5 (listing budget = min(5% of context,
listing_max_tokens)); unknown-context fallback 20K -> 10K.
2. Tier 2 no longer leaves the model blind: when even names-only doesn't
fit, the bridge description now carries a one-line-per-server summary
('cloudflare (3320 tools)') plus an instruction to search FIRST rather
than substitute a generic tool or claim the capability is missing —
the measured tier-2 failure mode (core-tool substitution) at zero
meaningful token cost (~50 tokens/server).
3. Listing degradation is now PER SERVER, largest first: one oversized
server (Cloudflare) collapses to its summary line while small
co-attached servers (Linear) keep their full per-tool listings
('mixed' form). Previously global: attaching Cloudflare next to
Linear silently cost Linear its listing. Greedy fit is deterministic
(size then label) so the rendered block stays byte-stable per catalog
— prompt-prefix cache safe.
E2E on real captures (defaults, 200K ctx): linear alone -> tier 1 full;
unreal alone -> tier 2 groups (5% budget) / tier 1 names at 1M;
cloudflare alone -> tier 2 groups; linear+cloudflare -> tier 1 MIXED
(linear fully listed, cloudflare summarized). 48/48 tests.
Tier 0: no MCP/plugin tools -> everything eager (pass-through).
Tier 1: deferred tools whose catalog listing fits min(threshold_pct%
of context, listing_max_tokens) -> bridge + skills-style
listing, degrading to names-only over budget.
Tier 2: listing over budget even names-only (Cloudflare's flat API
surface: 3,320 tools, names alone ~32K tokens) -> bare bridge,
discovery through tool_search only.
The old activation threshold (defer only when schemas > threshold_pct
of context) let mid-size catalogs ride eager and pay full schema cost;
with servers like Cloudflare (~597K tokens of schema, would not even
fit a 200K window) the binary gate is the wrong shape. Activation is
now driven purely by deferrable-tool presence; threshold_pct is
repurposed as the listing budget's context-relative leg.
- AssemblyResult gains tier + listing_form for observability
- listing_max_tokens default 4000 -> 20000 (cap 60000) so an 830-tool
catalog keeps a names-only listing while Cloudflare-scale drops to
bare bridge
- E2E verified against real captures: Linear 24 tools -> tier 1 full,
Epic UE 5.8 830 tools -> tier 1 names-only, Cloudflare 3,320 tools
-> tier 2 (both 200K and 1M context)
Deferred MCP/plugin tools become invisible once the tool_search bridge
activates — live benchmarking (48 runs, Claude Haiku 4.5) showed models
substituting visible core tools (terminal/web_search/browser) for deferred
capabilities or declaring them nonexistent instead of searching: 16/24
task success vs 24/24 with eager loading.
Skills never had this failure mode because every skill keeps a ~21-token
name+description listing line in the system prompt. This ports that exact
pattern to the tool bridge: when tool_search activates, a grouped manifest
of every deferred tool (name + first sentence of description, clipped to
60 chars, grouped per MCP server / toolset) is embedded in the tool_search
bridge description.
- tools/tool_search.py: build_catalog_listing() with deterministic
ordering (byte-stable across assemblies -> prompt prefix stays
cacheable); token-budget fallbacks full -> names-only -> legacy bare
count; bridge_tool_schemas(listing=...) embeds it and instructs the
model to skip tool_search when the exact name is visible (one fewer
round-trip per use)
- config: tools.tool_search.listing auto|on|off (default auto),
listing_max_tokens (default 4000, clamped 200..20000); legacy bool
shapes keep working
- tests: 8 new tests (config parsing/clamps, short-desc clipping,
deterministic rendering, budget fallbacks, bridge embedding, assembly
on/off paths); full file green (47 passed)
- docs: tool-search.md config table + rationale
- scripts/tool_search_livetest2.py: benchmark harness v2 with real
per-call token accounting (normalize_usage spy) and a third 'listing'
mode for A/B/C comparison
The background write guard decided ownership from `isinstance(usage_rec, dict)`,
so a local skill with NO usage record passed. That successful write called
bump_patch(), which created a `created_by: null` record — and the identical
write was refused from then on. "Allowed exactly once, then never" is a race
with our own bookkeeping, not a policy. Reproduced on main: patch #1 succeeds,
patch #2 with the same arguments is refused.
Option B from the issue. Option A (split `session_review` from
`scheduled_curator` and let the session fork patch user-owned skills it
consulted) would widen autonomous write permission onto skills the user owns
with no user present to consent — wrong direction for a no-user-present actor.
- skill_manager_tool: missing and explicit-null records now resolve
IDENTICALLY, both fail closed. The refusal names the reason and points at
`hermes curator adopt <name>`.
- background_review: both review prompts told the reviewer to patch any skill
consulted in the session and claimed pinned skills could be improved, while
enforcement refused both. Prompts now list pinned, external, and user-owned
skills as protected, and tell the reviewer to RECOMMEND adoption instead of
attempting a write that will be refused.
- skill_usage: document that `created_by` is a curator-management policy flag,
not a provenance claim, and add `is_curator_managed()` so call sites read as
the question they ask. Field name retained — it is on disk in every
`.usage.json` and renaming would strand those records.
- curator CLI: `hermes curator list-unmanaged` itemizes unmanaged skills with
the reason each is unmanaged (completes the #67139 spec).
Foreground writes are untouched: a user-directed edit to a user-owned skill
still works, including on pinned skills.
Sibling tests: 9 failures in test_skill_manager_tool.py were fixtures that
created record-less skills to exercise OTHER guards (consolidation-delete,
read-before-write) and relied on ownership falling through. Fixed at the
fixture, since the real curator only ever operates on managed sediment. One
test asserted the old "manually authored" wording; rewritten to assert the
behavior contract instead of the string.
Validation: 274 targeted tests + all 7 background-review files (60 tests) pass.
E2E on a temp HERMES_HOME (30 checks) covers the flip, foreground writes,
adoption unblocking, pin semantics, prompt/enforcement parity, and the new verb.
Each new test sabotage-verified: revert the fix, confirm it goes red.
Fixes#67140
`hermes curator status` reported only the skills it manages, staying silent
about curation-eligible skills it can never touch. On a 237-skill library that
meant 112 skills were invisible to every automatic transition with no signal
anywhere — the curator looked broken when it was working as designed.
A skill becomes curator-managed only when `created_by: agent` lands on its
usage record, and only the background review fork writes that marker. Two
populations therefore never qualify: records written before the marker existed
(no key at all, authorship unknowable) and every foreground
`skill_manage(create)` (unset by design — those skills belong to the user).
- skill_usage: `list_unmanaged_skill_names()` / `unmanaged_report()` enumerate
the blind spot, tagging each row with `has_provenance_key` so the two causes
are distinguishable. `adopt_skill()` writes the marker on user declaration
and refuses bundled, hub-installed, external, and protected built-ins.
Adoption never resets the inactivity clock.
- curator CLI: status prints an `unmanaged (no provenance marker)` block on
BOTH the managed and no-managed-skills paths; new `adopt` verb takes names or
`--all-unmanaged`, with `--dry-run` and a confirmation prompt on bulk.
- skills_sync: `_backfill_optional_provenance()` matched candidates by
repo-derived path only, so a skill installed at `mlops/chroma` that upstream
later moved to `mlops/vector-databases/chroma` was skipped forever and
`hermes skills repair-optional` could not fix it. Falls back to an
unambiguous name match, still gated on identical content, and records the
ACTUAL install path.
Provenance stays a declaration, never an inference: a high patch count proves
the agent MAINTAINS a skill, not that it authored one, since Hermes edits
user-written skills on the user's behalf routinely. An "looks agent-made"
heuristic would eventually archive hand-written work.
Validation: 347 targeted tests pass. Each new test verified via sabotage run
(revert the fix, confirm the test goes red) — one initially passed for the
wrong reason because the fixture pinned `prune_builtins` off, masking the guard
under test; fixed to force the shipped default on.
The Codex image backend rejected our own request shape for every account, and
we then translated that rejection into "Image generation is not enabled for the
current Codex account. Switch the image provider to OpenAI API key, FAL, or
xAI." — telling every affected user to abandon a provider that had never
actually been tried. That message is why this reads as a setup failure rather
than a bug: the wire error was replaced with a confident, wrong diagnosis.
Removes the classifier and its exception, so any HTTP failure surfaces
verbatim. The paired request-shape fix (previous commit) is what makes the
400 stop happening; this commit makes the next one diagnosable.
Also fixes error-body truncation: bodies were head-truncated at 500 chars, and
Codex error payloads can carry hundreds of bytes of leading metadata, so the
user got a wall of padding and no message. _summarize_error_body() prefers the
parsed error.message and falls back to a truncated raw body.
Docs: drop the unqualified image-to-image claim for the Codex backend and note
that the hosted tool call cannot be forced, so it is best-effort.
Verified E2E against a local fake Codex backend: success path writes a real PNG
with no tool_choice on the wire; the 400 path now returns api_error carrying
"Tool choice 'image_generation' not found in 'tools' parameter" (148 chars)
instead of the entitlement message. Sabotage run confirms all 4 regression
tests fail when the old behavior is restored.
Refs #19505, #49008, #31335.
Some OpenAI-compatible endpoints — notably Tencent Copilot
(copilot.tencent.com) — only accept streaming chat requests; any
non-streaming call returns HTTP 400 (code 11101, 'Non-stream chat
request is currently not supported'). The main conversation loop already
streams, so interactive chat works, but every auxiliary task (title
generation, compression, web extraction) used the non-streaming path and
failed on each call.
_provider_requires_stream() detects stream-only endpoints
(copilot.tencent.com built in, plus user-configurable
auxiliary.stream_only_base_urls substring markers in config.yaml).
Matching sync auxiliary calls route through _create_with_progress
(force_stream=True) and async calls through the new
_acreate_with_stream, aggregating the chunk stream — including tool-call
deltas and reasoning deltas — into a complete response via the shared
_ChatStreamAccumulator.
Salvaged from PR #60686 by @kudi88 onto the progress-aware streaming
machinery from #71508, addressing both sweeper-review gaps: the async
path now consumes the stream with 'async for' (awaiting create() and
iterating synchronously raised on AsyncOpenAI streams), and tool-call
deltas are reassembled instead of dropped (MCP passes tools= through
call_llm). Under force_stream there is no silent non-streaming retry —
a stream-only provider rejects those by definition, so the original
error surfaces to the normal recovery chains.
The gateway's pre-agent session-hygiene compression killed the summary
call at a fixed 30s wall-clock deadline (compression.hygiene_timeout_seconds),
regardless of whether the summary model was hung or merely slow. A reasoning
model happily streaming a large summary was cut off mid-generation, the user
got '⚠️ Context compression timed out after 30.0s', and a 300s failure
cooldown left the session oversized — a doom loop for slow-but-healthy
auxiliary models.
Timeouts are now liveness-based instead of wall-clock-based:
- agent/auxiliary_client.py: new thread-local aux_progress_hook. When
installed (only by context compression today), the primary call_llm
attempt streams (stream=True) and aggregates chunks back into a complete
response, ticking the hook per chunk. The configured timeout then acts
per stream read (idle) instead of as a total budget. Providers that
reject streaming fall back to the plain non-streaming call; auth/payment/
rate-limit/transport errors propagate unchanged into the existing
recovery chains. Codex Responses (per SSE event) and Anthropic Messages
(per stream event, via the new create_anthropic_message on_stream_event
callback) tick the same hook from inside their wire adapters.
- agent/conversation_compression.py: CompressionCommitFence gains
touch_progress()/seconds_since_progress(); compress_context() installs
fence.touch_progress as the progress hook around the compress call.
- gateway/run.py: the hygiene wait loop treats hygiene_timeout_seconds as
an INACTIVITY budget — while the fence reports fresh progress the wait
extends, bounded by the new compression.hygiene_total_ceiling_seconds
(default 600s, clamped >= the idle budget) so a degenerate trickle
stream still dies. The timeout warning now says the summary model
produced no output, which is the only case that still triggers it.
- config/docs: hygiene_total_ceiling_seconds added to DEFAULT_CONFIG and
configuration.md; hygiene_timeout_seconds documented as inactivity-based.
Tests: tests/agent/test_aux_progress_streaming.py (hook plumbing, stream
aggregation incl. tool-call deltas and reasoning deltas, rejection
fallback, ceiling kill, fence progress surface); two new gateway tests
prove a slow-but-streaming worker survives past the fixed timeout
(sabotage-verified: fails with the old fixed deadline) and a
forever-trickling worker is still cut off at the ceiling.
