Document the actual transport split: legacy editable drafts by default, optional rich drafts, persistent rich final sends, and in-place rich final edits for edit-based streams.
New tests/skills/test_authoring_standards.py parametrizes every bundled +
optional SKILL.md (1148 checks) against the mechanically-verifiable subset
of the hardline standards:
- required frontmatter fields (name/description/version/author/license/
platforms) + tags
- frontmatter name == directory name
- description <= 60 chars, ends with period, no marketing words
- related_skills resolve in-repo
- no machine-local paths
- <= 100k chars
Grandfather dict for legacy debt ships EMPTY — all pre-existing violations
fixed in this PR:
- 13 frontmatter names canonicalized to their directory names (the install
identifier); all related_skills references updated (comfyui -> stable-
diffusion). Fixes the class behind PR #42788's report; also fixes
here.now's invalid dot-name.
- optional-skills/devops/cli -> inference-sh-cli (dir was the generic
'cli'; fm name was right) incl. docs pages (en + zh-Hans), catalog row,
sidebar entry.
- pytorch-fsdp: 157k generated 'Quick Reference' dump moved to
references/common-patterns.md; SKILL.md 159k -> 2.5k with a pointer.
- research-paper-writing: 31.7k Phase 5 drafting section moved to
references/phase5-paper-drafting.md; SKILL.md 103k -> 71k.
Docs regenerated with scope discipline.
Per review: upscaling should be the default behavior (like the original
flux-2-pro chain), not agent opt-in. Policy: every image model whose
native output is below ~2MP now sets upscale=True in its catalog —
users never silently get low-res images. Native hi-res models
(Seedream 5 Pro/Lite, Krea 2 Large) stay off to avoid paying to
upscale already-large output.
- FAL catalog: 16 models flipped to upscale=True (klein, z-image,
nano-banana pro/2/2-lite, gpt-image 1.5/2, ideogram v3/v4, recraft
v4/v4.1, qwen image/3, krea-2 medium on FAL, MAI 2.5 pro).
- Krea plugin: per-model upscale defaults (medium + medium-turbo ON at
1.5K native; large OFF at 2K native), precedence explicit kwarg >
image_gen.krea.upscale config > catalog default.
- The 'upscale' tool param remains as a per-call override in both
directions (false = fast draft, true = force on hi-res/edits).
- Video unchanged: opt-in only (default-on would double every video's
cost and latency).
- Sibling tests updated: routing/payload tests pass upscale=False where
the assertion targets the generation submit; catalog test now pins
the native-resolution policy instead of the flux-2-pro snapshot.
The generated-media surface previously had almost no upscaler coverage:
only fal-ai/flux-2-pro chained Clarity Upscaler (hardcoded catalog
default), every other image model returned ~1MP output with no high-res
path, and video had no upscaler at all. Krea's API treats the enhancer
as a standard second pass; this brings the same shape to Hermes.
- image_generate: new optional 'upscale' boolean in the tool schema.
Explicit true chains the backend upscaler on ANY model (including
edits); explicit false disables flux-2-pro's automatic default;
omitted keeps per-model catalog behavior. Response now reports
'upscaled' so the agent knows which resolution it got.
- FAL image path: explicit flag overrides the catalog 'upscale' default
(Clarity Upscaler, 2x). Failure falls back to the native image.
- Krea plugin: upscale=true chains Krea Enhance
(/generate/enhance/krea/enhance, 2x, prompt-guided) through the same
BYO/managed base URL + auth as generation, with a best-effort poll
loop that never fails a successful generation.
- video_generate: new optional 'upscale' boolean; FAL video plugin
chains ByteDance SeedVR2 (fal-ai/seedvr/upscale/video, 2x factor
mode). Providers without upscalers ignore the kwarg per the ABC
contract (documented in both ABCs).
Validation: targeted suites green (123 tests across 6 files, including
new coverage for override-wins/default-kept/failure-fallback on all
three paths); live E2E on direct FAL verified both chains end-to-end
(klein 9b + Clarity upscaled image; pixverse-v6 1s 360p + SeedVR2
upscaled video).
- description 210 -> 57 chars; author credits Ben Barclay (benbarclay) first
- optional-skills/creative/ (marketing vertical, narrowest audience of
the batch)
- dropped phantom 'image-generation-workflow' ref; visuals via the
image_generate tool
- honest handoff language: platforms without connectors end at approved
drafts marked handed-off, never claimed as published
- tests (10) incl. phantom-ref and honest-handoff guards
- docs regen scoped: per-skill page + one catalog row + one sidebar line
Skill polish:
- description 208 -> 57 chars; author credits Ben Barclay (benbarclay) first
- connector framing (google-workspace, obsidian, notion, email-inbox-triage)
- modern section order; boilerplate folded into step-local rules
Blueprint wiring (completes the batch's recipe integration):
- weekly-review blueprint loads weekly-review-planning; prompt follows the
skill's seven-section shape, drafts-only
- morning-brief blueprint loads google-workspace; prompt points at
references/daily-brief.md when connected
- important-mail blueprint loads email-inbox-triage
- blueprints index regenerated
Tests: 13 skill tests + two catalog invariants (every blueprint skills=
entry resolves to a real bundled skill; the four task blueprints are wired
to their procedure skills). 32 green across both files.
Opt-in via compression.codex_responses_native (default: false). When enabled,
gpt-5.6-family models on the direct OpenAI API (api.openai.com) or a ChatGPT
Codex subscription send context_management=[{type: compaction,
compact_threshold: N}] on Responses requests. OpenAI compacts server-side and
returns an encrypted compaction output item; Hermes captures it into the
existing codex_reasoning_items sidecar and replays it on later turns in place
of the pruned history — inheriting persistence, session replay, the
cross-issuer guard, and the encrypted-replay kill switch with zero new state.
Scope is deliberately hard-gated (agent/native_compaction.py, re-checked per
request): gpt-5.6 family only — gpt-5.1/5.2 fail server-side on the field
(HTTP 500 / stream stall, no structured rejection; live-verified) — and
direct OpenAI/Codex routes only; xAI, GitHub/Copilot, OpenRouter, relays,
and local servers never see the field.
Hermes' local compression stays armed as the fallback owner: the native
threshold is clamped ~8K tokens below the local trigger so the server
compacts first, and a structured provider rejection of context_management
disables native compaction for the session and retries without it
(one-shot guard in TurnRetryState).
Live-verified E2E on api.openai.com/gpt-5.6: server compaction fired at a
4K threshold, checkpoints captured and replayed, recall preserved across
3 turns; gpt-5.1 with the flag enabled stays clean (field never sent).
Direction credit: PR #76950 by @laryhorb explored native Responses
compaction; this is a minimal reimplementation on current main.
The bundled docx, xlsx, powerpoint, and pdf skills were adapted from
Anthropic's document skills and carried their proprietary LICENSE.txt
(no derivatives, no redistribution). Flagged as critical license
findings by the SkillEvaluator Tier 1 scan of our skill tree.
This replaces all four with clean-room rewrites:
- Authored from scratch against library knowledge only (python-docx,
openpyxl, python-pptx, pypdf/reportlab/pdfplumber — all MIT/BSD) by
isolated subagents given functional specs, with an explicit
prohibition on reading the prior skill content or anthropics/skills;
session transcripts retained as provenance evidence.
- MIT licensed (LICENSE file per skill), author: Nous Research.
- Each skill: SKILL.md to house standards + argparse helper scripts
with UTF-8-explicit I/O + its own e2e pytest suite (fixtures built
on the fly, non-ASCII round-trips run under LC_ALL=C).
- All four pass SkillEvaluator Tier 1 pii+unicode+lint 3/3.
tests/skills/test_office_document_skills.py rewritten against the new
contracts: MIT/no-Anthropic-text invariants, scripts documented in
SKILL.md, argparse CLI shape, and a no-locale-default-open() check
(which caught and fixed a real gap: pdfplumber text reads are fine,
but the invariant scan now guards every future script).
Docs pages regenerated for the four skills (scoped; unrelated
generator drift excluded).
Honest capability deltas vs the old versions are documented per
SKILL.md (e.g. tracked-changes accept/reject and OOXML XSD validation
are not reimplemented; form flattening limits stated).
Salvage of PR #17973 by @TKCen (Sebastian Hänisch), re-implemented on
current main to preserve speed/instructions/provider params,
prepare_spoken_text normalization, OPUS_VOICE_PLATFORMS, is_write_denied
path security, microsecond timestamps, and the streaming-TTS gate.
- Split long TTS text into provider-safe chunks instead of truncating
- Pack generated audio against platform upload limits (Discord 10MB,
Telegram 50MB, configurable via tts.delivery_profiles)
- Combine chunks with ffmpeg (OGG/Opus re-encoded, MP3 stream-copied)
- Multi-file delivery when combination fails or would exceed limits
- Remove hard [:4000] truncation from all callers (cli.py, voice.py,
gateway/run.py, gateway/platforms/base.py)
- Gemini TTS raises ValueError instead of silently truncating when
composed prompt exceeds the provider limit
Simplify-code fixes: removed dead all_touched_paths set, added
try/finally for scratch file cleanup on exception, clean error response
on chunk failure instead of leaking stale file_path.
The published tarball ships lib/binding/napi-9-darwin-unknown-arm64 on every
platform, so a real Windows host has both it and the downloaded win32 binding
— the classify-everything gate threw on the darwin dir and killed every
Windows pack. Stage only bindings naming the target platform (classify still
rejects impostors), stop copyGlobByExt from recursing into lib/binding, and
add a version tripwire so a get-windows bump fails the build until the
lib/windows.js rewrite is re-verified.
Also from review: the renderer answers window.read.respond with empty text
when the IPC invoke rejects (older shell / main-side throw) instead of
stalling the tool's 30s timeout; the tool schema discloses that sibling
Hermes windows are skipped; docs gain read_window_below in both references.
The brief is single-connector (every command comes from google-workspace),
so it ships as references/daily-brief.md with a pointer + load trigger in
SKILL.md — progressive disclosure instead of a new skill-index entry.
Contributor's procedure preserved (half-open day windows, mail-to-meeting
linking with fuzzy-match discipline, 7-section brief, bounded actions);
credit noted in the reference header. Tests (8) guard the wiring and
disciplines. Version 1.1.0 -> 1.2.0.
The coverage warning listed bare page ranges, which tells the agent
WHERE the gaps are but not WHAT they contain — its only options were
guessing or OCRing everything. Each gap is now labeled with the last
text extracted before it (usually a section divider page), so the agent
can decide which gaps it actually needs and render/OCR only those.
Gap list capped at 20 entries with a summary line for pathological
alternating documents.
Rewrote from a sibling-skill routing table into a skill that carries its
own procedure, and folded in generalized rules from maintainer practice:
- full-thread reads (gh issue view --comments; newest comment = live state)
- duplicate-PR sweep (issue number + keyword variants) before any code
- design-intent check via git log -p -S alongside premise reproduction
- fix the class: sweep sibling call sites into the same PR
- sabotage run: prove the regression test fails without the fix
- open the PR immediately (PR dispatches CI; CI latency is the long pole)
- close the loop: comment the issue with the PR link
Also: description 205 -> 59 chars, author credits Ben Barclay first,
modern section order, boilerplate trimmed, tests (10) incl. a
router-pattern guard, scoped docs regen.
The non-streaming /v1/responses path built function_call and
function_call_output output items with no status field (and no item id),
while the SSE streaming path correctly emits status in_progress ->
completed. Spec-strict OpenAI clients reading the non-streaming output
array could interpret the status-less function_call items as pending
calls the CLIENT must execute — but these tools were already executed
server-side by the Hermes agent and are replayed for structured tool UI
only. Reported by a community user whose GPT-5.6 client concluded 'a
server should not tell an OpenAI client to execute a tool the server
already executed itself'.
- _extract_output_items now stamps status: completed and spec-shaped
item ids (fc_/fco_) on replayed items, matching the streaming path
- test updated to pin status + id shape
- docs example updated + explicit note that output tool calls are
replayed, never pending
- description 219 -> 58 chars
- author credits Ben Barclay (benbarclay) first
- modern section order; trimmed template safety boilerplate into
step-local rules and a skill-specific verification checklist
- tests at tests/skills/test_email_inbox_triage_skill.py (9 passing)
- docs regen scoped: per-skill page + one catalog row + one sidebar line
--resume latest resolves the most recent session through the same
workspace-scoped MRU lookup as -c (TUI source first under --tui, with
classic-CLI fallback). --in DIR chdirs before session resolution so the
lookup keys off DIR's workspace, and pins the session there by skipping
the recorded-cwd restore.
Requested by @Jeff9James: hermes --tui --resume latest --in ./dir
Fleet audit showed these task skills are commonly needed across users;
shipping bundled per Teknium's direction. Docs and tests follow the
bundled paths.
Adds a per-server `trust: full|untrusted` config key
(mcp_servers.<name>.trust). On an untrusted server, every write-capable
tool call — any tool whose discovery-time annotations do not carry
readOnlyHint=True — routes through the existing approval surface
(tools.approval.request_elicitation_consent, same lazy-import +
surface-routing pattern the MCP elicitation handler uses) before the RPC
fires. Denied/cancelled/errored approvals fail closed: the RPC never
runs, including the lazy first-use server spawn.
Design points:
- Classification happens at CALL TIME from metadata captured at
DISCOVERY (_record_tool_trust_metadata in _register_server_tools and
the lazy cache-registration path). No toolset/schema mutation, so the
toolset stays byte-stable and prompt caching is preserved.
- readOnlyHint is a server-supplied HINT: on an untrusted server a lying
server can at most skip approval for tools it claims read-only — it
can never widen access. Trust tiering itself is operator config.
- Missing/malformed annotations => write-capable (fail closed).
- Unrecognized trust values => untrusted (fail closed); missing key =>
full (backward compatible, documented in mcp-config-reference).
- The schema cache now persists readOnlyHint so lazy-registered servers
gate identically on next startup without spawning.
Tests: tests/tools/test_mcp_trust_gating.py (11 tests, TDD red->green):
approval invoked + accept proceeds, deny/cancel blocks RPC, readOnlyHint
=true skips gate, trusted/unconfigured servers skip gate, explicit
readOnlyHint=false gated, approval exception fails closed, trust
normalization, discovery-time capture (SDK objects and cached dicts).
Ported from: cloudflare-os classifyTool() (Apache-2.0), corroborated by
Claude Cowork (idea-level).
Two small config-gated features:
1. Kanban orphaned-card reconciliation (kanban.reconcile_orphans, default
true, config.yaml): a running card with broken claim bookkeeping
(claim_lock or claim_expires NULL — crash mid-claim, manual SQL, DB
restore) is invisible to all existing recovery paths
(release_stale_claims requires claim_expires NOT NULL,
detect_crashed_workers requires host-local lock + pid,
detect_stale_running is config-disabled by default) and shows Running
forever. New reconcile_orphaned_running() pass in kanban_db.py runs
each dispatch_once tick: requeues orphans to ready with an explanatory
comment, closes any leaked run, emits a 'reconciled' event, and defers
when the recorded PID is still alive on this host (never requeue
beside a live worker). Surfaced via DispatchResult.reconciled_orphans.
2. Per-server MCP identity header (mcp_servers.<name>.identity_header,
config.yaml): optional {name, value_from: static|profile, value}
mapping; the header is attached to that server's HTTP/SSE transport
requests. 'static' sends the config value; 'profile' resolves the
active Hermes profile name once at connect time (no per-call
mutation). Explicit per-server headers of the same name (any casing)
win. Invalid blocks warn-and-ignore; stdio servers warn-and-ignore.
Tests: tests/gateway/test_kanban_reconcile_orphans.py (9),
tests/tools/test_mcp_identity_header.py (13), all written first (RED)
then implemented (GREEN). No new HERMES_* env vars.
Inspired by: openai/symphony tracker reconciliation (Apache-2.0) +
Poke per-user MCP identity (idea-level).
Validate a job's configuration BEFORE any agent machinery is constructed:
- missing provider API key (AuthError from a read-only
resolve_runtime_provider probe; skipped when a fallback_providers chain
is configured, since auth-fallback may rescue the run)
- attached skill not ready (skill_view readiness_status=setup_needed —
missing required env vars / commands / credential files)
- delivery platform unknown or unconnected (deliver=local/origin/all are
never checked; gateway-config load failures fail open)
On a failing check run_job returns a [blocked_config]-marked error without
constructing AIAgent/MCP/etc, so a misconfigured job never burns an LLM
call. run_one_job records last_status='blocked_config' and delivers the
alert exactly ONCE across ticks (persisted preflight_alerted bit — the
alert-once shape from the #73506 dead-pin auto-pause); the next healthy
run clears the marker so a future break re-alerts. Every preflight check
fails open: only an affirmative misconfiguration verdict blocks.
Config: cron.preflight (default true); `cron.preflight: false` restores
the old fail-during-run behavior. Documented in the cron user guide and
config defaults.
mark_job_run gains an optional status= override (unblocked call shape
unchanged) and drops preflight_alerted on any successful run.
Tests: tests/cron/test_preflight_config.py (blocked_config + no agent +
single alert across two ticks, healthy job unaffected, recovery clears
dedup, fallback-chain rescue, opt-out restores old behavior, skill
readiness miss, unknown delivery platform, deliver=local never loads
gateway config). Full tests/cron/ + cronjob tool suite green (525 tests).
Ported from: paperclipai/paperclip execution-semantics §5 (MIT);
in-repo precedent: #27948, #73506
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).
External hosts speak this protocol directly, so the parameter that
rewrites a session's stored transcript should not be folklore. Document
what each truncation field means, that an ordinal without
confirm_truncate is refused, and that a client must never hold the
ordinal in state across ordinary submits.
The timeout table already lists the stale non-stream detector, but the
prose read as if it only guarded the interactive path. Spell out that it
bounds the inline cron / delegated-subagent calls too, and name the
accepted-then-silent failure mode it recovers from.
- matrix/dingtalk: extract deps-only installers (ensure_matrix_deps,
ensure_dingtalk_deps) and register THOSE as ensure_deps_fn — the prior
check_*_requirements combined credential env checks with the install,
so a platform configured via PlatformConfig.extra (which is_connected
accepts) would pass enablement, reach create_adapter(), and have the
'installer' veto on env-var grounds before installing anything —
re-creating the #79812 deadlock for extra-configured setups. The
combined deps+credentials functions remain for setup/status callers.
- matrix/feishu passive probes: use the existing lazy_deps.is_available()
instead of hand-rolling 'not feature_missing(...)' (reuse finding).
- teams: module docstring no longer recommends bare system pip (the
PEP 668 trap purged everywhere else); docs troubleshooting row updated
to match the new hint text.
- wecom_callback: drop dead 'global ET, DEFUSEDXML_AVAILABLE'
(ensure_and_bind mutates the module dict directly; nothing assigns).
- tests: parametrized wiring contract for all 8 lazy-installable
platforms — ensure_deps_fn present and distinct from check_fn
(behavior contract, not identity snapshot, so renames don't churn it).
- gateway/config.py: rewrite the stale enablement-pass header comment that
still described check_fn as 'the single source of truth for are-my-env-
vars-set' / 'lazy-installs it' — both false under the new contract.
- teams: check_requirements docstring wrongly claimed credential checks
(body checks only SDK/aiohttp presence); derive install_hint from the
canonical LAZY_DEPS pins + sys.executable instead of hardcoding
'~/.hermes/hermes-agent/venv/bin/pip' and version pins (wrong under
HERMES_HOME overrides / profile installs; pins go stale on CVE bumps);
connect() fatal-error hints now point at the venv pip instead of bare
system pip (the PEP 668 trap the docs warn about).
- teams docs: drop exact version pins from the two manual-install commands
(LAZY_DEPS is the source of truth; unpinned installs still work and the
text can't go stale).
- hermes_cli/status.py: per-entry exception guard around check_fn so one
raising probe can't abort the listing of all remaining plugin platforms
(aligns with the other three call sites).
- tests: rename test_register_check_fn_is_active_lazy_installer ->
test_register_splits_passive_probe_from_active_installer (name said the
opposite of what it verifies).
mzkarami
teknium1
kshitij
Jack Maloney
Brooklyn Nicholson
Ken Weiner
kinsolee
GodsBoy
HexLab98
kshitij
texasich