initialize() snapshots OPENVIKING_* into the provider once, so /reload
(which only updates os.environ) leaves viking_* tools running against
stale auth — users have to restart hermes to pick up keys added to
~/.hermes/.env after startup.
Add _ensure_client(), which re-resolves the connection settings via the
same _resolve_connection_settings/_load_hermes_openviking_config path
initialize() uses and rebuilds + health-checks the client only when an
OPENVIKING_* value actually changed; otherwise it reuses the cached
client so the hot path stays at one dict comparison with no network
calls. Every `if not self._client:` guard in system_prompt_block,
queue_prefetch, sync_turn, on_session_end, on_memory_write and
handle_tool_call now goes through it.
Refreshing is gated behind a flag set at the end of initialize() so the
baseline is established before any env re-resolution happens — callers
that wire up a client directly keep the existing client untouched.
Refs #21130
(cherry picked from commit b694d21b7c)
`_write_env_vars` in the OpenViking memory provider interpolates each
secret straight into a `KEY=VALUE` line, but the values only ever pass
through `_clean_config_value`, whose `value.strip()` trims surrounding
whitespace and leaves internal CR/LF intact. Because the file is strictly
line-oriented and is re-read via `read_text().splitlines()`, a value that
carries an embedded newline spills onto a second physical line, and the
tail is re-parsed as an independent `KEY=VALUE` entry on the next round
trip. A secret pasted with a trailing record (e.g. an `OPENVIKING_API_KEY`
copied with an extra line) therefore injects an arbitrary additional
variable into the persisted credentials file and silently corrupts it.
The fix neutralizes the line terminators at the single chokepoint where
values reach the file. A small `_env_line_safe` helper strips `\r`, `\n`,
and the NUL byte from each value, and both write sites in `_write_env_vars`
(the existing-key update branch and the appended-key branch) route through
it, so a value can only ever occupy the single line it is written on.
## What does this PR do?
Hardens the OpenViking memory provider's `.env` writer so a malformed or
pasted secret value can no longer break out of its `KEY=VALUE` line and
inject a rogue variable into the profile-scoped credentials file.
## Related Issue
N/A
## Type of Change
- [x] 🐛 Bug fix (non-breaking change that fixes an issue)
## Changes Made
- `plugins/memory/openviking/__init__.py`: add `_env_line_safe()` which
removes `\r`, `\n`, and `\x00` from a value, and apply it to both the
updated-key and appended-key write branches in `_write_env_vars()`.
- `tests/plugins/memory/test_openviking_provider.py`: add two regression
tests covering a fresh write and an in-place key update with embedded
CR/LF, asserting no injected line survives the read-back.
## How to Test
1. Run the targeted tests:
`pytest tests/plugins/memory/test_openviking_provider.py -k env_writer -q`
2. Reverting the `_env_line_safe` sanitization makes
`test_openviking_env_writer_strips_embedded_newlines_in_values` and
`test_openviking_env_writer_strips_newlines_when_updating_existing_key`
fail with a rogue `INJECTED_KEY=`/`ROGUE=1` line appearing in the file,
confirming the tests pin the bug.
3. `ruff check plugins/memory/openviking/__init__.py` and
`python scripts/check-windows-footguns.py plugins/memory/openviking/__init__.py`
both pass.
## Checklist
### Code
- [x] I've read the Contributing Guide
- [x] My commit messages follow Conventional Commits
- [x] I searched for existing PRs to make sure this isn't a duplicate
- [x] My PR contains only changes related to this fix
- [x] I've run the relevant tests and they pass
- [x] I've added tests for my changes (required for bug fixes)
- [x] I've tested on my platform: macOS 15 (Darwin 25.5)
### Documentation & Housekeeping
- [x] I've updated relevant documentation (docstrings) — or N/A
- [x] I've updated `cli-config.yaml.example` if I added/changed config keys — N/A
- [x] I've updated `CONTRIBUTING.md` or `AGENTS.md` if I changed architecture or workflows — N/A
- [x] I've considered cross-platform impact (strips CR as well as LF) — done
- [x] I've updated tool descriptions/schemas if I changed tool behavior — N/A
(cherry picked from commit f29dd2df84)
Audited every reference against ground truth (COMMAND_REGISTRY, argparse
--help output, TOOLSETS dict, provider profiles, DEFAULT_CONFIG) and
corrected drift the restructure had carried over verbatim:
- slash-commands.md: rebuilt from the live registry. Removes the phantom
/skill command (never existed — skills load via /skills, hermes -s, or a
skill's own /<name> command) and the wrong /q alias on /quit (/q belongs
to /queue) — both reported by @liuhao1024 in #50608 and @gauravsaxena1997
in #50613. Adds ~25 real commands that were missing (/learn, /memory,
/pet, /hatch, /bundles, /moa, /suggestions, /blueprint, /whoami, …) with
correct aliases and CLI/GW scoping.
- cli-reference.md: verified flags (-z/--oneshot, --tui/--cli, --safe-mode),
real subcommand sets for config/setup/sessions/skills/gateway/mcp/profile/
auth/webhook/cron/skin/pets, added missing top-level commands (fallback,
project, moa, logs, console, hooks, security, backup); condensed tables.
- providers-and-models.md: rebuilt from the 36 shipped provider profiles
with their actual env vars (was 21 rows, several stale); folds in the
alias system and fallback-chain command.
- configuration.md: toolset table matches TOOLSETS (adds coding,
computer_use, video_gen, context_engine, project; drops nonexistent
messaging/rl), config sections match DEFAULT_CONFIG, STT/TTS provider
lists match the shipped set.
- background-systems.md: curator verbs match the argparse surface.
- troubleshooting/security-privacy/windows-quirks/contributor-guide: fixed
dangling 'section above/below' references from the body split; moved the
reset-permissions playbook to security-privacy.md (its routing home).
Reported-by: liuhao1024 <sunsky.lau@gmail.com>
Reported-by: gauravsaxena1997
Flips the default fan-out cadence from per_iteration (advisors re-run on
every tool iteration, multiplying advisor spend by tool-loop depth) to
user_turn (advisors run once on the first message of each user turn; the
acting aggregator works the rest of the tool loop with that turn's
advice). Until per-mode benchmarks justify a costlier default, MoA
defaults to the cheapest, lowest-impact cadence (#67199).
One default for everyone — no split legacy/new-preset semantics; presets
that want per-step advising set fanout: per_iteration explicitly. All
three modes (user_turn / per_iteration / every_n:N) remain selectable;
every_n:1 still collapses to per_iteration (semantic identity), while
unparseable values now fall to user_turn (the default).
Docs updated with a default-change note; the per-iteration rerun test
pins its mode explicitly.
Co-authored-by: skyer-flyyy <188930297+skyer-flyyy@users.noreply.github.com>
Every SKILL.md description over 60 chars was silently truncated to
57 chars + '...' in the system-prompt skill index
(extract_skill_description, agent/skill_utils.py), destroying the
routing signal for 69 of 179 skills — some descriptions ran to
1,005 chars.
Rewrites follow the authoring standard: <=60 chars, one sentence,
ends with a period, trigger front-loaded, no marketing words, no
skill-name repetition. Excess detail already lives in each skill's
body.
Includes the touchdesigner-mcp trim from PR #32361 (credit:
@JeliTron) and aligns with the router-precision direction of PR
#48780 (@John-Lussier). Docs catalogs + per-skill pages regenerated
via website/scripts/generate-skill-docs.py.
Co-authored-by: JeliTron <287797501+JeliTron@users.noreply.github.com>
The skill-authoring guide and curator prompt both reference
descriptions as the primary discovery mechanism but never mentioned
the 57-char system prompt truncation. Add explicit guidance:
- Authoring guide: frontmatter docs, template comment, size limits,
pitfall #3 with good/bad examples, verification checklist
- Curator prompt: parenthetical noting the 57-char window when
writing umbrella skill descriptions
When a skill is created or edited with a description longer than
SKILL_PROMPT_DESC_LIMIT (60 chars), the tool response now includes a
system_prompt_preview field showing exactly what the system prompt
skill index will display. This gives the agent immediate feedback to
self-correct truncated trigger phrases.
Also adds tool schema guidance about the 57-char window and fixes a
stale docstring in skill_commands.py that incorrectly claimed the
system prompt renders the full description.
The system prompt skill index truncates long descriptions to 57 chars,
but this limit was a hardcoded magic number. Extract it as a named
constant and factor the normalization logic into a shared private
helper so the extraction function and the new truncation predicate
cannot drift.
No behaviour change — pure refactor.
Follow-up on the salvaged commit: the xurl skill loads even when x_search
isn't registered (check_fn-gated on xAI credentials), so per the
cross-toolset reference rule the skill must not name it. Replaced the
skill's x_search routing block and workflow step with skill-native
wording (raw engageable posts, authenticated context, write-evidence
rule). The cross-surface comparison stays on surfaces where both are
known to exist: x-search feature docs, toolset description, tools-config
setup note, and the x_search tool schema (kept generic, no tool names).
Rewrote the routing tests to pin the placement contract, including that
the skill never names credential-gated surfaces.
Make the x_search / xurl boundary explicit in the skill, feature docs,
toolset metadata, setup note, and reference pages, while keeping the
model-facing x_search schema generic (no static xurl name).
Regression tests assert behavioral routing invariants rather than frozen
prose snapshots. Drop the stale CI-only plugin/hangup hunks already on
main so this rebases cleanly.
skills/computer-use/SKILL.md sat at the root of the bundled skills
tree as an uncategorized single-skill directory. Move it to
skills/autonomous-ai-agents/computer-use/ alongside claude-code,
codex, opencode, and hermes-agent.
- git mv skills/computer-use -> skills/autonomous-ai-agents/computer-use
(history preserved)
- root-level-skill docstring examples (commands.py,
generate-skill-docs.py) now use a generic placeholder instead of
naming a specific skill, since categorizing root-level skills is
ongoing
- docs: move the bundled page to autonomous-ai-agents-computer-use,
update sidebars.ts, skills-catalog.md, and the two skill-path
references in features/computer-use.md
E2E validated: fresh sync_skills() copies to the new nested path;
existing installs with the old flat copy keep it untouched (manifest
name-keyed, hash unchanged -> skipped, no duplicate);
_get_category_from_path resolves 'autonomous-ai-agents'; docusaurus
build green; 396 targeted tests pass.
The 'queues an Enter-submitted draft while compaction is active' test
pastes a large message to push the session over the fixture's 22k
threshold_tokens. At repeat(500) the payload is only ~4k tokens — the
other ~18k came from the ambient system prompt (tool schemas + skills
index + memory), leaving the trigger margin-less. The hermes-agent
skill hub restructure (e3d524b482) shrank the bundled skills index by
~160 tokens and dropped the total just under threshold: compaction never
started, waitForHeldCompletion() hung, and the test timed out at 90s on
every branch since — including main (first red run: 9a4d1a0130).
Bump the payload to repeat(1500) (~12.4k tokens, total ~30k) so the
test crosses the threshold on its own weight with ~8k tokens of margin,
and document the invariant so the next prompt-weight change doesn't
resurrect this.
A direct run (cronjob action='run' / webhook-triggered manual fire) takes
the job's fire_claim and advances next_run_at. When it races the external
provider's scheduled fire for the same occurrence, Chronos loses the claim
and — by design — does not re-arm (the winner owns the re-arm). But the
direct-run winner never notified the provider, so the NAS one-shot for the
consumed occurrence was left stale forever and the recurring job silently
stopped firing.
Observed in production: a managed 1-minute review job stalled for 20 hours
because a GitHub-webhook direct run claimed the job 2s before the Chronos
fire arrived; every subsequent occurrence was orphaned while /api/status
stayed green.
Fix: after a *claimed* direct execution completes (success or failure —
next_run_at advances at claim time either way), call
_notify_provider_jobs_changed_safe() so the active provider re-arms the
post-run next_run_at. No-op for the built-in ticker; claim-lost direct
runs still never notify (the winning scheduler owns the re-arm).
Claude Code's /simplify was renamed away, community-restored, and rebuilt
since our 3-agent port (v2.1.63 -> v2.1.154+). This brings simplify-code
in line with the current upstream design, re-expressed in our own wording
and layered onto our existing risk-tier/confidence machinery:
- New Reviewer 4 (Altitude): flags band-aid fixes layered on shared
infrastructure — special cases, symptom patches with unfixed sibling
sites, workaround stacks — and points at the deeper mechanism fix.
Matches our AGENTS.md 'fix the class, not the site' rubric.
- Explicit cleanup-vs-bug-hunt boundary: this skill improves working
code; correctness review stays with requesting-code-review.
- Inline single-pass fallback when delegate_task is unavailable (leaf
subagents, delegation disabled) — previously the skill just broke;
now all four angles run sequentially with honest disclosure.
- Finding format gains a concrete-cost field.
- Efficiency reviewer adds closure-capture scope-retention leaks.
- Pitfalls: fan-out cap 3 -> 4, band-aid-vs-deliberate-boundary caveat,
no-bug-hunting drift guard.
Kept our value-adds upstream lacks: SAFE/CAREFUL/RISKY tiers,
Chesterton's Fence via git blame, dry-run/focus/scope modifiers.
With the built-in x_search tool (xAI-backed, returns a synthesized
answer) now shipping, the xurl skill's plain 'searching posts' wording
no longer disambiguated the two search surfaces. Describe xurl search
in its own terms — authenticated X index query returning raw post JSON
with IDs suitable for immediate engagement — so agents route correctly
whether or not x_search is enabled. No cross-toolset reference added,
per the schema/skill cross-reference rule.
Add two pitfalls discovered when running the skill against a fresh
Jupyter server:
- Pitfall #9: When the websocket reply channel hangs on every execute
even though the kernel actually ran (REST shows execution_state=idle
and execution_count increments), force zmq transport with
--transport zmq. The zmq transport uses jupyter_client directly and
sidesteps the broken websocket layer.
- Pitfall #10: A fresh ServerApp rejects POST /api/sessions with
"_xsrf argument missing from POST" unless you start it with
--ServerApp.disable_check_xsrf=True. Needed for REST-only flows
where no browser/cookie is establishing the XSRF token.
The yuanbao skill is platform-specific guidance for Tencent Yuanbao
group chats — niche for the default bundled set. Per the
'when in doubt, optional' rule, ship it as an optional skill.
Install via: hermes skills install official/yuanbao/yuanbao
The nano-pdf description repeated the skill name ('via nano-pdf CLI'),
violating the skill authoring standard, and spent its char budget
restating the name instead of distinguishing the skill from the
structural pdf skill. New description: 'Edit text in existing PDFs via
natural-language prompts.' (56 chars) — no name repetition, and
'text in existing PDFs' contrasts cleanly with pdf's
'Create, merge, split, fill, and secure PDF files.'
Propagated to the auto-generated docs pages and zh-Hans locale
(catalog row + per-skill page) for locale parity.
Per the 'when in doubt, optional' rule — zero-shot image segmentation
via SAM is a niche computer-vision capability with heavy deps
(torch, transformers), not something most users load.
Install via: hermes skills install official/mlops/segment-anything
Both are heavy-dep, GPU-bound local music-generation skills (8-16GB VRAM,
torch stacks, manual source patches) that were bundled in two different
categories (media/ and mlops/models/). Per the 'when in doubt, optional'
rule they now sit side by side in optional-skills/creative/, and the
bundled songwriting-and-ai-music skill points at them for local generation.
- git mv skills/media/heartmula -> optional-skills/creative/heartmula
- git mv skills/mlops/models/audiocraft -> optional-skills/creative/audiocraft
- cross-link related_skills both ways + songwriting-and-ai-music
- new section 10 in songwriting-and-ai-music with install commands
- docs: bundled pages -> optional pages (en + zh-Hans), catalogs, sidebar
Install via:
hermes skills install official/creative/heartmula
hermes skills install official/creative/audiocraft
The hermes-agent skill body was a 51KB monolith loaded in full on every
trigger. It is now a lightweight hub (~12KB): identity, quick start, the
tmux orchestration guide (kept in-body — the autonomous-ai-agents category
contract), surface orientation, and hard invariants, with a routing table
into 18 reference files that carry the depth.
Absorbed four Hermes-specific skills as first-class references so their
content gains a discoverable home and room to grow without bloating any
primary body:
- skills/hermes-themes -> references/themes.md + templates/skin.yaml
- skills/hermes-desktop-plugins -> references/desktop-plugins.md + templates/plugin.js
- skills/productivity/tui-widgets -> references/tui-widgets.md + templates/clock.mjs
- skills/productivity/petdex -> references/petdex.md
New references extracted from the old body: cli-reference, slash-commands,
providers-and-models, configuration, project-context-files,
security-privacy, background-systems, windows-quirks, troubleshooting,
contributor-guide (native-mcp and webhooks already existed). Also commits
delegate-task-concurrency-diagnosis and portal-auth-for-third-party-apps,
which the old body referenced but the repo never shipped.
Description updated to cover the widened scope:
'Use, configure, theme, extend, and orchestrate Hermes Agent.' (60 chars)
skills/dogfood/SKILL.md sat at the root of the bundled skills tree,
making it one of the few uncategorized skills (Discord /skill
autocomplete listed it under 'uncategorized'; hermes_cli/commands.py
cited it as the example). Move it to
skills/software-development/dogfood/ alongside the other QA/testing
skills (test-driven-development, systematic-debugging,
requesting-code-review).
- git mv skills/dogfood -> skills/software-development/dogfood
(history preserved)
- fix test fixture path in tests/tools/test_browser_console.py
- update root-level-skill docstring examples (commands.py,
generate-skill-docs.py) to cite computer-use, which is still root-level
- drop 'dogfood' from the category list in hermes-agent-skill-authoring
SKILL.md (en + zh-Hans docs mirrors)
- docs: regenerate/move the bundled page to
software-development-dogfood (en + zh-Hans), update sidebars.ts,
skills-catalog.md, and the adversarial-ux-test related-skills link
E2E validated: fresh sync_skills() copies to the new nested path;
existing installs with the old flat copy keep it untouched (manifest is
name-keyed, hash unchanged -> skipped, no duplicate);
_get_category_from_path resolves 'software-development'; docusaurus
build green.
Maintainer scoping decision for the #51226 salvage: document that
select_context() is for engines that must REPLACE per-request context
(retrieval/routing) — pre_llm_call is inject-only by documented cache
design; that observation-only plugins should implement a MemoryProvider
(sync_turn) rather than a context engine, with on_turn_complete scoped
as the observation mirror for engines that already select; and that a
non-no-op select_context naturally changes the prompt-cache prefix on
turns where the selection changes — engines should return stable
selections when nothing changed.
Verification follow-up for the #51226 salvage: the host call site guarded
select_context with hasattr(), but the ABC defines a default on every
engine, so the built-in ContextCompressor (and any non-implementing
engine) still paid per-request shallow copies of the conversation
history plus a hook call on every provider request. Identity-check the
bound method against ContextEngine.select_context and return the
request untouched — mirroring the existing base-method short-circuit in
_notify_context_engine_turn_complete — so the default path does zero
work, not just produces an identical result.
Adds two pins: the base no-op is never invoked (patched-to-raise base
stays silent), and ContextCompressor.__dict__ contains neither new verb.
Also registers the contributor email mapping for @chaos-xxl.
- website guide: on_turn_complete() now carries the same best-effort coverage
caveat as the ABC docstring (fires from the finalization seam; abnormal
early-return paths bypass it) — removes the doc/code inconsistency.
- test: finalization seam emits on_turn_complete with usage=None + the
interrupted flag for an interrupted finalized turn. Docstring records that
the negative early-return-bypass half is best-effort and deferred to a
shared-seam follow-up rather than pinned via a full run_conversation harness.
Addresses the hermes-sweeper review on #51226:
- _apply_context_engine_selection now passes shallow copies of the read-only
conversation_messages / incoming_message to the hook, so an engine mutating
them in place cannot corrupt persisted transcript state (enforces the
request-only contract, not just documents it). Adds a mutation-regression
test asserting persisted history + incoming message are untouched.
- on_turn_complete docstring: scope the coverage claim to the standard
finalization seam. Some abnormal early-return paths (content-policy block,
provider terminal failure) currently persist+return without finalization and
don't emit the hook; documented as best-effort with a shared-seam follow-up,
rather than over-promising a guaranteed callback for every early exit.
- _apply_context_engine_selection: reject an empty list. all([]) is True, so
a [] returned by a failing/buggy engine previously replaced a valid request
with an empty message list the downstream sanitizers can't restore; now it
falls open to the unmodified request (honors the fail-open contract).
Thanks @johnnykor82 for catching this on #41918's review.
- test: empty list keeps the original request (fail-open regression).
- docs: document select_context()/on_turn_complete() in the public
context-engine plugin guide (were still describing only the old contract).
- context_engine.py: document that select_context() runs before cache-control
and all request sanitizers, so (a) replacements still pass host validation
and (b) the no-op default keeps the request byte-stable (AGENTS.md prompt-
cache invariant). Note the hook is evaluated per provider request.
- tests: no-op path is byte-stable for cache-control; a role-unusual
replacement is passed through for the existing downstream sanitizers to
normalize (select_context does structural validation only).
The on_turn_complete() observation hook is the engine's post-turn signal,
so it should receive the completed turn's canonical token usage when the
host has it, not a hardcoded None. Per @johnnykor82's #41918 contract: the
engine uses prompt/completion + cache_read/write/reasoning buckets to judge
how large/expensive the selected context was before the next select_context().
- conversation_loop.py: stash the most recent provider response's usage_dict
(the same canonical shape fed to update_from_response) on the agent as
_last_turn_usage; reset to None at turn start so turns that never reach a
provider response (early failure / interrupt) forward None, not a stale
prior turn's usage.
- turn_finalizer.py: forward agent._last_turn_usage instead of usage=None.
- context_engine.py: document the usage param contract on the ABC hook.
- tests: cover both ends through the real finalize_turn path — completed turn
forwards the full canonical bucket set intact; no-response turn forwards None.
Co-authored-by: johnnykor82 <johnnykor82@users.noreply.github.com>
Adds the post-turn observation verb as the companion to select_context():
an optional, no-op-default on_turn_complete() called once after the
assistant/tool loop finishes, with the finalized transcript snapshot. Lets
an engine ingest/index/summarize the completed turn to inform the next
select_context(). Wired via _notify_context_engine_turn_complete() from
turn_finalizer.finalize_turn(); fail-open, base no-op short-circuited so
non-implementing engines (incl. the built-in compressor) pay nothing.
This is the request-assembly + observation pair from #41918; with this
commit the PR fully subsumes #41918's two hooks (prepare_request_messages
-> select_context, on_turn_complete) rather than only the selection half.
Co-authored-by: johnnykor82 <johnnykor82@users.noreply.github.com>
Adds an optional, no-op-default select_context() hook to the ContextEngine
ABC, called every turn after the request messages are assembled and before
provider dispatch — independent of should_compress(). Lets an engine select
or replace which context enters the prompt for a single request (retrieval,
topic routing, role/branch switching) without mutating persisted history,
removing the need to abuse should_compress()=True as a per-turn callback.
The host call site (_apply_context_engine_selection) is fail-open: a missing
hook, an exception, or an invalid return value leaves the assembled request
untouched. Additive and non-breaking: the built-in compressor and every
existing engine are unaffected.
Consolidates the per-turn request-assembly surface proposed across #41918,
Related: #36765#41918#24949#47109#50053#23837#25115#29370
Routine automatic compression stays silent-by-design on chat platforms
(default unchanged, byte-identical). New opt-in config key
compression.progress_notices (bool, default false) opens a gate on the
gateway noise filter (_prepare_gateway_status_message) that lets ROUTINE
compression progress statuses through to chat surfaces.
- Membership is derived from the #69550 status template constants in
agent/conversation_compression.py (compiled to a literal-escaped regex,
never re-inlined wording), so unrelated noisy statuses (aux failures,
provider retry/rate-limit chatter) stay suppressed even when enabled.
- The compaction completion notice (COMPACTION_DONE_STATUS, #69546
lifecycle 'compacted' edge) already flows through the status path and
passes the filter — no new emit site needed.
- Config wired everywhere: hermes_cli/config.py DEFAULT_CONFIG,
cli-config.yaml.example, gateway raw-YAML read (live, mtime-cached),
gateway hot-reload cache-busting key list, website configuration docs.
- Failure notices and manual /compress feedback remain always-visible;
VISIBLE_COMPRESSION_MESSAGES and emit sites untouched.
Design by @havok-training (issue #52995).
_render_tool_calls only handled dict-shaped entries; a SimpleNamespace-
shaped tool_call (SDK-style stream-stitched responses) rendered as
'[called tool: tool]', silently losing the function name and arguments
from the advisory view. Handle both shapes (including a namespace-shaped
nested function inside a dict entry).
One-hunk hardening salvaged from closed#59712.
Co-authored-by: SquabbyZ <601709253@qq.com>
Reference models may have a smaller context window than the aggregator
(e.g. kimi-k2.7-code @ 262K advising a glm-5.2 @ 1M conversation).
Without context-length protection, a reference whose window is exceeded
gets a hard HTTP 400 from the provider, which _run_reference's
try/except silently converts to a [failed: …] note — the MoA turn
silently degrades to fewer references (#60345).
Redesigned implementation of #60387:
- Estimate AFTER the advisory system prompt is prepended, so the
request that is actually sent is what gets budgeted.
- Reserve output headroom: the preset's reference_max_tokens when set,
else an 8192-token constant, plus a 10% estimator-error fraction.
- Trim on advisory-view boundaries (text-only user/assistant turns; no
tool-result frames to orphan), preserving the system prompt, the
user-first invariant after every pop (never assistant-first), and the
trailing synthetic user turn.
- Cache get_model_context_length per (provider, model) in a per-fan-out
dict shared across the worker threads, so a turn resolves each
window once instead of probing metadata sources
per-reference-per-iteration (failures are cached too).
Co-authored-by: webtecnica <75556242+webtecnica@users.noreply.github.com>
Follow-ups for salvaged #56344:
- A reference that completes between the interrupt check and the reap
keeps its REAL output and accounting (the provider call billed) instead
of being zeroed with a placeholder.
- A reference still in flight at interrupt time gets a placeholder in the
results, but its future now carries a done-callback that folds the
eventual real usage/cost into the facade's pending accounting
(late_accounting_sink -> _record_late_reference_accounting), so billed
spend is never silently dropped. Pending totals are folded (not
overwritten) and guarded by a lock since done-callbacks fire on
executor worker threads.
- Interrupted placeholder results are no longer written into the facade's
turn-scoped reference cache: a cache HIT never re-runs references, so
caching a partial snapshot would replay '[skipped: interrupted by
user]' notes for the rest of the turn. The cache is left empty and the
next create() re-runs the fan-out.
agent/tool_executor.py's concurrent tool batch checks agent._interrupt_requested
and aborts the wait early; agent/moa_loop.py's _run_references_parallel had
no equivalent, so a MoA-enabled turn blocked on ThreadPoolExecutor.result()
until every reference model finished or hit its own individual
auxiliary.moa_reference timeout -- there was no way for the user to abort a
live turn mid-fanout.
Thread an optional `agent` parameter through aggregate_moa_context ->
_run_references_parallel (used when MoA references run alongside the main
model) and MoAClient/MoAChatCompletions (used when the MoA preset itself is
the acting model), then poll concurrent.futures.wait() in
_REFERENCE_POLL_INTERVAL_S slices instead of blocking on future.result() per
reference, checking agent._interrupt_requested each cycle.
Deliberately scoped to interrupt/cancel only -- no new or changed timeout
value, so this doesn't overlap open PRs #53784/#53875 (which lower the
per-reference timeout default but don't add interrupt support). `agent` is
optional and defaults to None, so any caller that doesn't pass it keeps
today's uninterruptible blocking behavior unchanged.
Extends the all-references-failed short-circuit (#56975) to the
persistent `provider: moa` facade path: MoAChatCompletions.create()
previously attached 'use the reference responses below' guidance built
entirely from failure sentinels and called the aggregator with it. Now
an all-failed turn attaches either the sanitized unavailability notice
(loud policy) or nothing (silent policy), and the aggregator — which IS
the acting model — simply acts alone. Advisor accounting for the failed
fan-out is still recorded.
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
When every MoA reference model returns a failure (HTTP error, timeout,
etc.) or is skipped by the recursion guard, the one-shot aggregator
synthesis call is now skipped entirely. Previously it would try to
synthesise a wall of failure sentinels, which could block for the full
provider timeout (observed ~6 min on SenseNova) before returning a
non-retryable error that left the session hanging.
The early return carries the sanitized unavailability notice (never raw
provider error text, per the failed-reference containment) so the main
agent loop can still act in single-model mode.
Salvaged from #56975, reworked atop the _is_failed_reference helpers.
Follow-ups for salvaged #53784:
- reference_timeout now defaults to None = no per-preset override, so the
reference fan-out inherits auxiliary.moa_reference.timeout (900s default)
via call_llm's own per-task timeout resolution. The PR's 30.0s default
would have cut off long-thinking advisors mid-response, and its 300s max
cap capped legitimate explicit values — both removed. Explicit per-preset
values are still honored as-is.
- _is_failed_reference also treats '[skipped: …]' recursion-guard notes as
internal sentinels, keeping them out of both aggregator prompts.
- Dashboard/desktop TS types updated to number | null; web_server validator
accepts null/empty as 'inherit'.
