A stalled compression summary never raises, so the auxiliary client's
exception-path fallback is unreachable from it. When the progress-aware
timeout aborts a stalled worker, re-run the summary once pinned to the
first auxiliary.compression.fallback_chain entry before degrading to
continue-without-compression.
The pin is a single-use ContextVar consumed by the context compressor's
summary call, so it cannot leak into the detached stalled worker or the
compressor's own main-model retry. A fresh fence is minted through the
host factory so a /stop during the retry still admits against the live
commit boundary.
The legacy tail budget scales as threshold×target_ratio, which was designed
around 128K windows at a 50% trigger (~13K tail). On modern big-window
models with raised thresholds it silently hoards: a 1M-window session at
threshold 0.85 keeps a 170K-token verbatim tail (255K soft ceiling) out of
EVERY compaction, so a 540K manual /compress lands at ~290K and every
subsequent turn re-ships the hoard. Nobody chooses this; it is an artifact
of the formula outside its design envelope.
Lean mode (#87326, compaction-v2) was built for exactly this and its recall
was validated in the before/after eval (evals/compaction/results/): clamped
2.5%-of-window tail (10K floor / 25K cap), continuity carried by the
upgraded summary (digests, anchor index, verbatim user messages,
session_search recovery pointers). This flips the DEFAULT to lean; explicit
'tail_mode: legacy' in config keeps the old behavior exactly.
Also fixes a latent bug the flip exposed: update_model() re-assigned the
LEGACY formula directly when recomputing budgets, silently reverting a lean
compressor to the hoard on every mid-session model switch. The recompute
now routes through the mode-aware tail_token_budget property (regression
test included).
Surfaces: context_compressor.py defaults + getattr fallbacks, agent_init
parse default, DEFAULT_CONFIG, gateway _CACHE_BUSTING_CONFIG_KEYS gains
compression.tail_mode (mode changes now evict cached gateway agents like
target_ratio changes do), user + developer docs. Tests: 3 new default
contracts, legacy tests pinned explicitly, feasibility-skip scenario pinned
to legacy (under lean its payloads correctly become compressible).
E2E counterfactual (real imports, 1M window @ 0.85):
main default: legacy, tail 170,000 (ceiling 255,000)
head default: lean, tail 25,000 (ceiling 37,500)
head legacy: 170,000 (opt-out intact)
update_model to 400K: 10,000 (lean preserved across switch)
Background processes started by subagents (task_id sa-*) route their
notify_on_complete / watch_pattern notifications to the parent
conversation (b95ec1cb5) because anything outliving the child needs a
durable consumer. In practice these 'npm ci finished' walls are noise
mid-conversation — the child's consolidated delegation result is the
deliverable.
- New config key delegation.surface_child_process_notifications
(default false = suppress). Flag true restores the previous behavior
exactly (delivery with subagent attribution line).
- drain_notifications drops (never requeues) completion/watch_match/
watch_disabled events whose task_id starts with 'sa-' when the flag
is false, logging at debug with session_id+task_id for diagnosis.
Requeueing would pin them forever — children never drain notifies.
- async_delegation events are NEVER suppressed (they ARE the result).
- watch_disabled emitters now carry task_id so sa- sessions' safety
events follow the same suppression as their other events.
- Config read errors fall back to the default (suppress) and never
crash the drain loop.
- Docs: delegation.md + configuration.md.
Review finding on #94633: _sanitize_label_value is lossy ('team/workspace'
and 'team_workspace' both sanitize to 'team_workspace'; >63-char keys
truncate identically), and container reuse is label-keyed — so two teams
with DIFFERENT shared keys could silently attach to one running container
(filesystem, processes, env) while their host sandboxes stayed separate.
Shared-key labels now carry a sha256 digest suffix of the raw key
(deterministic across processes; plain profile labels unchanged for
backward compat). Docs also state the first-creator-wins rule for image/
mounts on a shared container. Adds adversarial collision tests.
web_extract stopped using an auxiliary LLM long ago (deterministic
truncate-and-store), but browser snapshots still routed oversized
accessibility trees through the auxiliary web_extract model, keeping a
dead-looking aux slot alive across every config/picker surface.
- tools/browser_tool.py: remove _extract_relevant_content and
_get_extraction_model; oversized snapshots always truncate at line
boundaries, store the full tree to cache/web, and append a read_file
pointer (element refs beyond the cut live in the file)
- tools/browser_camofox.py: same — no LLM path
- Remove auxiliary.web_extract slot: config_defaults (removal note, same
pattern as session_search/PR #27590), cli.py defaults + env bridge,
gateway/run.py bridged keys, hermes config display, hermes model picker,
dashboard REST slots, desktop + web AUX_TASKS, i18n labels (en/zh/
zh-hant/ja/ar)
- Docs: env-vars, configuration, fallback-providers, browser + zh-Hans
mirrors (web-search zh-Hans was stale on the old LLM pipeline — synced
to truncate-and-store truth)
- Tests updated: aux bridge uses approval slot, browser tests assert the
LLM path is gone and stored files are secret-redacted
GSC page-level data shows configuration (171K impr, pos 5.3), quickstart
(135K, 5.3), providers (123K, 5.3), web-dashboard (68K, 6.5), docker
(37K, 6.3) and the desktop app page (98K, 3.9) all losing ranking
headroom because their title/H1 are bare nouns instead of the terms
people search.
- Frontmatter title + H1 now carry the query terms on all six pages
- Desktop docs page links back to the new marketing /desktop product
page, joining the two official properties Google sees for the query
Done by Hermes Agent (deepseek-v4-pro via nous), Nous Research.
Follow-up to the #65422 salvage:
- startup_orphan_reap joins _RECOVERABLE_END_REASONS (kept distinct from
ws_orphan_reap for forensics): every recovery fence
(find_latest_gateway_session_for_peer, unarchive_recoverable_session,
promote_to_session_reset) now treats a startup-swept row as an
accidental end, so a sweep never makes a session unresumable.
- Config key moves from sessions.orphan_reaper to
dashboard.startup_orphan_sweep in DEFAULT_CONFIG, next to its siblings
ws_ping_interval / ws_ping_timeout / ws_orphan_reap_grace_s; the raw
loader in tui_gateway.server reads the new key (fail-open on missing).
cli-config.yaml.example and website/docs/user-guide/configuration.md
follow the dashboard.* documentation pattern.
- New regression test: a stranded 'active' row (ended_at NULL, no live
runtime) is swept AND still recoverable via peer-keyed lookup and fully
revivable via reopen_session afterward.
Close session rows left ended_at IS NULL when the in-process websocket
orphan timer dies with the process (#65194). Dual-clock staleness
(started_at AND newest message), desktop included, live in-memory
sessions excluded, scheduled once from both entry.main and the WS
sidecar so desktop/dashboard boots also run the sweep.
Builds on @fattchris resolve_turn_limit salvage (#67696): flips the default
from a numeric cap to unlimited across all construction paths (CLI, agent_init,
run_agent subagents), adds inf/infinity/null to the unlimited spellings, and
sets DEFAULT_CONFIG agent.max_turns to null. The turn cap caused more problems
than it solved (silent mid-task truncation).
Addresses teknium1 sweeper review on PR #67696:
1. Gateway bridge: Skip str(None) bridging when YAML value is Python None
(from or bare ). Previously str(None) → None → unlimited
instead of default 90. Now clears stale env var so resolver applies default.
2. TUI: Route _cfg_max_turns through resolve_turn_limit instead of bare
int(). Old code crashed on none/unlimited and swallowed 0 via
. HERMES_TUI_MAX_TURNS env var also routed through
resolver.
3. Docs: Document unlimited spellings (none/unlimited/infinite/0/-1) in
configuration.md.
4. Tests: Add TestGatewayBridgeNullHandling (4 tests) and TestTUIResolver
(8 tests) covering null handling, string spellings, env var override,
and legacy root-level config.
All 50 tests pass.
When the chosen/keyed backend fails a web_search or web_extract call
(bad key, upstream outage, 5xx, raised exception), that single call
retries on the keyless free-tier ring instead of erroring. The next
call attempts the chosen backend again — no sticky failover, no state.
Resolves the keyed half of #78984/#32159 (keyless half landed in the
ring PR).
- tools/web_tools.py: _rescue_eligible (keyed ring vendors + non-ring
backends eligible; keyless-mode calls excluded — they already walked
the ring), _rescue_search/_rescue_extract (search annotates
rescued_from + backend_error naming the original failure and the
retry-next-call semantics; extract rescues only whole-batch failures,
partial failures pass through untouched; rescue failure preserves the
ORIGINAL backend error with the rescue note appended)
- both dispatchers wrap the provider call: failure-results AND raised
exceptions rescue; ineligible paths re-raise unchanged
- web.keyless_rescue config key (default true; implicitly off when
keyless_fallback is off); docs updated
Live E2E: keyed Tavily with an invalid key 401'd and the call was
served by the real ring with the rescue annotation; a second call
re-attempted Tavily first (statelessness proven); whole-batch extract
rescue returned real page content. 13 new tests; 67 green across the
keyless suites.
Fresh installs with zero web credentials now rotate web_search/
web_extract across FIVE vendors' public free tiers — Exa, Parallel,
Tavily, Firecrawl, Keenable — instead of a 2-vendor 50/50 split, with
next-in-line ring failover on rate limits (multi-hop until a vendor
serves or the ring is exhausted; served_by marks the actual vendor).
- plugins/web/keenable/: new bundled provider (search via /v1/search,
fetch via /v1/fetch; keyed Bearer or keyless with the mandatory
X-Keenable-Title app header). Credit: integration proposed by
Ilya Gusev (Keenable) in #49758; Free/Paid picker rows included.
- keyless_mcp: tavily/firecrawl/keenable keyless search+extract
wrappers, _KEYLESS_RING + per-process round-robin cursor (seeded by
the random session id, advances per unpinned request), pinned-vendor
entry (pin = start there; rotation off), paid-pinned vendors excluded
from the ring entirely.
- Tavily/Firecrawl providers route keyless traffic through the ring;
both are now default-on ring members (no longer selection-gated).
- web_tools/registry: keenable in backend sets, auto-detect, availability
probes; _keyless_preference() delegates to the ring cursor.
- KEENABLE_API_KEY in OPTIONAL_ENV_VARS; docs updated (ring semantics).
Live E2E: all 10 vendorXcapability paths (5 search + 5 extract) served
real results keyless; rotation cycled all five vendors over 5 dispatch
calls; double-throttle failover walked exa->parallel->tavily.
- Updated the Tavily API key description to clarify that it is optional and keyless access is supported.
- Modified the Tavily plugin and provider to handle requests with or without an API key, using Bearer authentication when the key is provided.
- Enhanced documentation to reflect the new keyless functionality and updated environment variable descriptions.
- Added tests to ensure correct behavior for both keyed and keyless requests.
Composio eval traces showed Hermes wasting turns re-issuing identical tool
calls (same tool, same args, same result — 3x/4x in one run) and ending
turns by announcing an action it never took. Two conservative, config-gated
guards (agent.stall_guards, default true):
- Identical-call loop breaker: ToolCallGuardrailController.observe_identical_call
tracks the consecutive streak of (tool, canonical args, result-hash); on
the 3rd identical call a compact one-line notice is appended to that tool
RESULT at construction time (cache-safe — tool results are append-only).
Never blocks the call. Pollers (process, *_get_result, *_poll) are exempt
via STALL_GUARD_REPEATABLE_TOOLS. Streak resets on any different call,
changed result, or new turn. Observed on the raw result before the
tool-loop warning suffix so its changing count can't defeat matching.
- Said-continue-but-stopped recovery: trailing_continue_intent() detects a
short reply ENDING on an announced next action ('Let me now…', 'I will
now…', 'Next, I…'); the conversation loop feeds it into the EXISTING
intent-ack continuation path (same interim-assistant + user-nudge
mechanism, same codex_ack_continuations cap of 2), preserving message
alternation — no parallel recovery machinery.
Config: agent.stall_guards in DEFAULT_CONFIG; docs in configuration.md;
unit tests for streak/allowlist/reset/gate and detector pos/neg cases.
Composio-style MCP servers return un-paginated 22-47K-char payloads that
sail under the generic 100K per-result spillover threshold, bloating
context and ballooning per-turn reasoning time on long conversations.
Competitors cap harder (OpenCode/pi 50KB, Claude Code 30K, Codex ~10K
tokens). Three changes:
- mcp_* tools spill at a tighter 50K default (BudgetConfig.mcp_result_size,
config-overridable via tool_budget.mcp_result_size_chars; pinned and
per-tool overrides still win; capped by the context-scaled default).
- The persisted-output preview now teaches recovery: page the saved file
with read_file or process with execute_code instead of re-requesting the
same data from the remote API.
- Untrusted/MCP string results are scanned (bounded, first 64KB) for
provider-side elision markers ('...N more items', "has_more": true,
'saved to sandbox', data_preview) and get ONE cache-safe incompleteness
notice appended at result-construction time, before untrusted wrapping —
so the model stops treating provider-elided enumerations as complete.
- Hard 2M-char allocation cap in mcp_tool.py (text, error, and
structuredContent paths) so a pathological multi-MB server payload is
bounded before it propagates, while ordinary large results reach
spillover intact. Distilled from #56060/#56072/#56511 (issue #56059);
supersedes their 50K lossy truncation with spillover-friendly semantics.
Docs: configuration.md spillover-budget section + cli-config.yaml.example.
Co-authored-by: Stoltemberg <215755014+Stoltemberg@users.noreply.github.com>
Co-authored-by: AlexFucuson9 <295703459+AlexFucuson9@users.noreply.github.com>
Co-authored-by: Tranquil-Flow <66773372+Tranquil-Flow@users.noreply.github.com>
Unpinned zero-credential installs now pick Exa or Parallel by the
parity of the per-process random session id (stable within a process,
even split fleet-wide) instead of always favoring Parallel. An explicit
hermes tools selection (web.backend / per-capability keys) bypasses the
split entirely; the runner-up vendor stays in the walk as fallback.
Live E2E: 6 fresh processes split 3/3 between vendors, each performed
a real keyless search via its picked endpoint; explicit pin verified.
Un-fences OPENAI_MODEL_EXECUTION_GUIDANCE from the gpt/codex/grok substring
check and gives it its own injection gate, independent of
tool_use_enforcement, controlled by config.yaml `agent.execution_guidance`
(auto/true/false/list — same semantics as tool_use_enforcement). The "auto"
list (EXECUTION_GUIDANCE_MODELS) now also covers deepseek, kimi, qwen, glm,
minimax, mimo, and mistral.
Composio agentic-eval traces showed Hermes+DeepSeek/Kimi failing where
competitors passed: financial math done in prose, no read-back after
external writes, malformed identifiers "repaired", completeness claimed
despite count mismatches. The discipline block existed but those models
never received it.
The block is extended with compact clauses distilled from that analysis:
- external-write read-back (tool-call success is not task success; internal
file edits already confirmed by the tool are not re-verified)
- count reconciliation (declared totals/has_more are hard assertions)
- literal preservation (never normalize identifiers that fail a stated
format; lookup success does not validate a malformed token)
- retry-differently (empty/partial/suspiciously narrow results get a
broader retry before concluding)
- completion gated on verification (done = every named acceptance
criterion verified, never a plausible subset)
The todo tool description now encourages enumeration-as-checklist for
"all N items" tasks and gates completed status on verified work, never
intent.
Guidance is chosen once at session start keyed on model name, so the
system prompt stays byte-stable for the life of a conversation.
Supersedes/absorbs prior contributor proposals: #20588, #35087, #41874
(MiMo), #53847 (GLM tool-calls-as-text stall).
Co-authored-by: Mat-London <56627804+Mat-London@users.noreply.github.com>
Co-authored-by: intelac <8803887+intelac@users.noreply.github.com>
Co-authored-by: 6ylqq <51219463+6ylqq@users.noreply.github.com>
Co-authored-by: tauros1983 <267660491+tauros1983@users.noreply.github.com>
Exa and Parallel now each render as two picker rows in hermes tools —
'Free (keyless)' and 'Paid (API key)'. Selection persists to
web.provider_tier.<name>:
- free: always the anonymous public endpoint, even with a key set
- paid: always the keyed SDK path; missing key errors instead of
silently downgrading to the free tier (is_keyless_available also
returns False so the auto-fallback walk can't route there)
- unset: auto (key present -> paid, else keyless)
Mechanism: get_setup_schema() gains a 'variants' list the picker
flattens into sibling rows sharing one web_backend; selection writes
the tier via both _write_provider_config sites; active-row detection
matches the tier (auto mirrors use_keyless). Routing goes through a
single use_keyless() chokepoint shared by search+extract in both
providers.
Live E2E: tier=free with a fake key present searched keyless OK (a
keyed call would have 401'd); tier=paid without key errored naming
PARALLEL_API_KEY; picker rows verified for both vendors x both tiers.
With zero web credentials configured, web_search/web_extract previously
resolved to the nonfunctional firecrawl sentinel and errored. Now the
backend resolution walks a strictly-last keyless tier: Parallel's and
Exa's public anonymous MCP endpoints (the same free tiers opencode ships
as its default search path).
- plugins/web/keyless_mcp.py: minimal JSON-RPC tools/call client for
mcp.exa.ai + search.parallel.ai (SSE + plain JSON parsing, typed
errors, per-process random session id, no user identifiers)
- WebSearchProvider.is_keyless_available(): separate weaker tier that
never leaks into is_available(), so keyed setups are never pre-empted
- Exa/Parallel providers: route to keyless endpoints when their key is
absent; keyed SDK path unchanged
- registry + _get_backend(): keyless walk (parallel -> exa) strictly
after every keyed/importable candidate; check_web_api_key() lights
the tools up on zero-credential installs
- web.keyless_fallback config key (default true) to disable the tier
- docs: web-search.md + configuration.md
E2E-verified against both live endpoints from an isolated HERMES_HOME
(search + extract via the real dispatchers, disable-flag negative path).
Live incident 2026-08-17: the source checkout was parked on a stale feature
branch (claude-code-inspired/local-terminal-memory-limit, days behind main),
left there by earlier tooling. 'hermes update' autostashed, refreshed lazy
backends, synced skills, and printed '✓ Code updated!' / '✓ Update complete!'
while the checkout stayed on the stale branch with none of main's new code.
Two sessions burned time on 'the fix is missing' confusion.
- Parked-branch guard: auto-switch back to the update target ONLY when the
parked branch is clean and fully merged (git cherry origin/<target> shows
nothing unmerged); the checkout then STAYS on the target instead of being
re-parked. Otherwise: loud CODE UPDATE SKIPPED block naming the branch,
behind-count, and resolution commands; exit 1; branch untouched.
- The up-to-date (commit_count == 0) path no longer switches back to a
fully-merged parked branch either.
- Post-pull gate additionally refuses to print '✓ Code updated!' when HEAD
ends up attached to a non-target branch.
- Summary lines now carry the actual branch + HEAD short-sha:
'✓ Update complete! [main @ 30fcf9580]' — drift visible at a glance.
- New config toggle updates.auto_switch_parked_branch (default true).
- Real-git-fixture regression tests (init/clone/branch, no subprocess
mocks): clean+merged auto-switch, dirty skip, unmerged skip, cherry-picked
equivalence, config opt-out, unverifiable ref, on-main fast path,
up-to-date no-repark, summary branch/sha assertions.
Follow-ups on top of the salvaged CommandCode provider plugin (PR #32909):
- hermes_cli/config_defaults.py: COMMANDCODE_API_KEY setup-wizard entry
- hermes_cli/doctor.py: add key to the doctor env-var scan list
(health check comes free via the pluggable-profile loop)
- hermes_cli/dump.py: include commandcode in debug-dump api_keys
- docs: provider table row, fallback-provider table + supported lists
- tests: doctor dedicated-skip test now uses exact-name checks so
Bearer-authed Anthropic-COMPATIBLE gateways (CommandCode (Anthropic))
are allowed in the generic loop while native anthropic stays skipped
E2E verified with real imports: profile registration, aliases,
PROVIDER_REGISTRY auto-extension, bearer-auth host match
(positive + negative), live /models fetch (55 models).
The 'Configure auxiliary models' menu under 'hermes model' now includes a
Delegation entry so the delegate_task subagent model is discoverable and
configurable interactively, instead of requiring hand-edited
delegation.provider / delegation.model keys in config.yaml.
Delegation is not an auxiliary_client task — subagents are full child
agents resolved via tools/delegate_tool.py — so the picker entry writes to
the top-level delegation.* section rather than auxiliary.*. 'auto' (inherit
the parent agent) is persisted as empty strings, never the literal 'auto',
which delegate_tool would try to resolve as a provider name. 'Reset all to
auto' clears only the four delegation routing fields and preserves
non-routing settings like max_concurrent_children.
#87326 shipped the lean-compaction capability on the compressor; this adds
the config.yaml surface (compression.tail_mode: legacy|lean, default
legacy), DEFAULT_CONFIG entry, and docs on both the dev-guide compression
page and the user-guide configuration page, with zh-Hans parity.
Completes the duplicated-chrome class fix:
- Focus-in (CSI I) now routes through the same rate-limited full-redraw
recovery as Ctrl+L//redraw, clearing ghost prompt/composer copies after
Alt+Tab / tab switches (focus-regain variant reported on #60920, #25337)
- Document display.cli_rebuild_scrollback_on_redraw in configuration.md
- Register the new default in hermes_cli/config_defaults.py (moved from
the pre-refactor config.py location the salvaged commit targeted)
- Collapse the duplicated discord LoginFailure/PrivilegedIntentsRequired
classification (name-match + isinstance blocks repeated the same code/
message tuples) into a single _is() helper — one message per failure.
- Replace the user-facing HERMES_RECONNECT_ATTENTION_AFTER_SECONDS env var
with agent.reconnect_attention_after in config.yaml (default 7200, 0
disables), bridged internally like gateway_timeout. .env is for secrets.
- Use _float_env for robust parsing instead of bare int(os.getenv(...)).
- Document terminal classification + needs_attention escalation in
website/docs/user-guide/configuration.md.
Adds delegation.worktree_isolation (default: false). When enabled, each
delegate_task child gets its own git worktree branched from the repo's
current HEAD under <repo>/.worktrees/subagent-<id>, its terminal session
starts there, and its goal message carries the isolation contract
(work + commit in the worktree; parent reviews/merges the branch).
- tools/subagent_worktree.py: clean-room implementation from Muse Code's
documented --subagent-worktree-isolation behavior (create per-child
worktree, finalize/inspect after run, auto-prune clean no-commit
worktrees, keep anything holding work).
- tools/delegate_tool.py: config gate + per-child setup in
_run_single_child; result entries gain a "worktree" field (path,
branch, commits, dirty, pruned) only when isolation engaged — the
default-off wire shape is byte-identical.
- Git-only + local-terminal-backend-only; non-git dirs, remote backends,
or any worktree failure degrade silently to shared-workspace behavior.
- Tests: tests/tools/test_subagent_worktree.py (15 tests, real git
repos) + E2E through _run_single_child with a real repo verified
parent-checkout isolation, branch reviewability, prune, and
default-off shape pinning.
- Docs: delegation feature page section + configuration.md key.
OpenRouter's /v1/models entries advertise reasoning capability
(supported_parameters + reasoning.mandatory/supported_efforts). Use that
metadata as the primary gate in _supports_reasoning_extra_body instead of
the hand-maintained vendor-prefix allowlist, which went stale one vendor at
a time (nvidia/ missing -> #75386). Also clamp the requested effort to the
nearest LOWER catalog-supported level in the OpenRouter profile so ultra/max
against a high-capped route no longer 4xxes.
Cache-only on the hot path: capabilities parse for free out of the existing
fetch_openrouter_models() payload, a background warmer covers cold starts,
and unknown models/offline catalogs fall back to the static prefix list
unchanged.
Adds a pre_transcription transform hook (prompt/language/model mutable,
file_path read-only, last-writer-wins per the transform_* convention)
fired before any STT backend, threads prompt to faster-whisper
(initial_prompt) and OpenAI/Groq/Mistral/DeepInfra (prompt), adds an
optional stt.prompt config key on the same plumbing, and keeps the
no-hook dispatch path byte-identical. Fixes#64168.
Documents the new surface for users: a "Transcription prompt
(vocabulary hints)" subsection in the configuration guide (composition
order, per-provider support matrix, length contract, privacy warning),
a pre_transcription entry in the hooks reference, and the mirrored row
in the plugins hook table.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AG6LyYMvHC2o6HbVUozmVR
* fix: make verify_on_stop opt-in everywhere (default False, not auto)
The verify-on-stop nudge was already judged more noise than signal: the
v31 migration flips existing installs off, the v32 migration catches the
baked-in literal-true population, and the docs tell users to 'treat off
as the effective default and opt in explicitly'. But DEFAULT_CONFIG still
shipped the "auto" sentinel, so exactly one population kept getting the
nudges: fresh installs (and any config missing the key), where "auto"
resolves ON for CLI/TUI/desktop surfaces. Live symptom: repeated
'[System: You edited code ... run verification]' interruptions the user
never asked for and had to hunt down in source to disable.
- DEFAULT_CONFIG: agent.verify_on_stop "auto" -> False (opt-in).
- verify_on_stop_enabled(): missing/unrecognized value now falls back
OFF instead of surface-aware; explicit "auto" still selects the
legacy surface-aware behavior, explicit bools unchanged, and the
HERMES_VERIFY_ON_STOP env override is untouched.
- No migration needed: v31/v32 already normalized existing installs,
and this only changes the merged default for configs without the key.
- Docs updated; default-path E2E test now asserts OFF, plus a new
missing-value regression test. Also added the standard win32 skip
marker to the symlink-based temp-dir test (pre-existing Windows
failure, same class as tests/cron/test_cron_script.py).
* test: update config goldens — verify_on_stop=False is now stripped as default
With the DEFAULT_CONFIG flip to False, the migration-write invariant
(_persist_migration / save_config strip_defaults) no longer materialises
verify_on_stop: false to disk unless the user explicitly set the key:
- V20 floor fixture (agent: {} on disk): v31's write is stripped —
agent stays {} and load_config() supplies False at read time.
- V12 floor fixture (explicit verify_on_stop: true on disk): the key is
a user-set path, so the v32 flip stays materialised as false.
- Partial-write and _persist_migration regressions now assert the key is
absent from disk and (for the merge case) that the merged view still
resolves False.
Behavior verified with a one-shot migrate_config run against both
fixture shapes.
Two bugs reported on the docker terminal backend (desktop app, sandboxed
profiles with container_persistent: false):
1. A NEW chat's container inherited the PREVIOUS session's workspace,
bind-mounted rw at /workspace, because the mount source was the
process-global TERMINAL_CWD env var (written by the workspace picker,
outliving its session) and all sessions shared one 'default' container.
2. Every command failed with exit 126 because the desktop gateway recorded
the HOST launch directory as the session cwd, and each command was
prefixed with 'cd /Users/<user>/...' inside the container.
Fixes (class-wide, single owners):
- container_persistent: false + docker now keys containers PER SESSION:
fresh container per chat, removed at session close/idle. delegate_task
children share the parent's container via an explicit alias registry.
container_persistent: true keeps the documented ONE-long-lived-container
contract unchanged.
- _resolve_task_host_cwd() is the single owner of the cwd->/workspace mount
policy across all four env-creation sites; under isolation it refuses
process-global cwd sources and mounts only the session's own attached
workspace (tui_gateway now tags overrides with cwd_source).
- _resolve_command_cwd() gains the same host-path guard the env-creation
sites already had (#50636/#54447 sibling site): a recorded host cwd is
discarded on container backends instead of cd-ing every command into a
nonexistent path.
E2E-tested against real Docker: distinct containers per session, no stale
mount in a fresh session, no exit 126 from host cwd records, containers
removed at session teardown.
Record why the cache needs a third bound and what the pressure pass will and
will not shed, so an operator tuning agent.agent_cache knows which knob to
reach for. Adds the config keys to the session-lifecycle appendix and a user
guide section covering the "auto" cgroup-derived budget.
The docs promised 'frees ~370MB RAM' — measured behavior on macOS/CPU
is that ctranslate2's allocator keeps the freed pages (RSS doesn't
visibly shrink); the concrete win is VRAM release on CUDA hosts and
process-internal reuse on CPU. Say exactly that instead.
The local faster-whisper model singleton (_local_model) is loaded once
and never released — the 'base' model holds ~370 MB of RAM/VRAM for
the entire lifetime of the process, even when no voice messages arrive
for hours or days. On long-running gateway processes (especially with
local LLMs competing for the same GPU) this is wasteful.
Add a config-driven idle unload: after stt.local.unload_after_idle_seconds
(default 0 = never) of no transcription activity, a lightweight daemon
thread sets _local_model = None so the Python GC can reclaim the
ctranslate2 objects. The next voice message reloads the model
transparently (the existing lazy-load path handles it).
The watcher:
- Checks every 30s whether idle time exceeds the configured threshold
- Acquires _local_model_lock before unloading (prevents races with
concurrent transcriptions that are mid-load)
- Exits immediately if the model is already None (unloaded by another
path, e.g. the CUDA fallback eviction)
- Is restarted by each transcription with the current config value,
so changing stt.local.unload_after_idle_seconds in config.yaml takes
effect on the next voice message without a process restart
Default is 0 (never unload) — zero behavior change for existing users.
Recommended value for gateway processes: 300 (5 minutes).
15 tests: config resolution (garbage/negative/None fallbacks), unload
safety (already-None, lock acquisition), touch timestamp, watcher
lifecycle (unload after timeout, no unload within timeout, exits when
model already None, stopped on new start). Existing STT test suite
unchanged.
The review-fold commit added the input-duration gate but the user-facing
docs and config example still implied every cloud clip gets trimmed.
One-line additions to both.
Local faster-whisper gets Silero VAD (bf8004e3a) so silence never
reaches the model. Cloud providers got no such protection: the raw
file uploads untouched, so every second of silence in a voice note is
paid for twice — upload time and per-audio-minute billing — and cloud
Whisper hallucinates junk tokens on silent stretches exactly like
local Whisper did before the VAD hardening. A 13s voice note with two
long pauses is billed as 13s of audio to transcribe ~6s of speech.
Close the gap client-side: before uploading to a built-in cloud
provider (groq/openai/mistral/xai/elevenlabs/deepinfra), collapse long
pauses with ffmpeg's silenceremove filter, keeping
stt.cloud_trim_keep_ms (default 300) of every pause so word boundaries
and natural pacing survive. Uses ffmpeg, already a dependency of this
exact path via _transcode_audio_for_stt — no new dependency.
The trim is strictly best-effort — ALL of these upload the original
untouched, transcription never fails because of the trim:
- stt.cloud_trim_silence: false
- ffmpeg/ffprobe missing, trim failure, or timeout
- trimmed result ~empty (mostly-silence clip: the provider, not a
client-side dB heuristic, decides whether it contains speech)
- trim saves <10% (re-encoding for nothing)
Command-type and plugin providers are deliberately NOT trimmed: they
may wrap local CLIs that want the original bytes or run their own VAD.
E2E (real ffmpeg + faster-whisper): 13.2s voice note with 7s pause ->
6.2s upload (-53%); transcript of trimmed audio matches the original
on both utterances. Dense-speech and all-silence WAVs correctly fall
back to the original. 22 unit+E2E tests; STT/voice suite failures
identical to upstream/main baseline (all pre-existing).
Shaun Eccles