The Explore agent is a "Fast agent" optimized for codebase exploration,
not deep reasoning. The skill-stocktake V4 design requires holistic AI
judgment (actionability, scope fit, uniqueness, currency) which needs
the full reasoning capability of the conversation's main model.
Additionally, the Agent tool has no `model` parameter — specifying
`model: opus` was silently ignored, causing the evaluation to run on
the lightweight Explore model. This resulted in all skills receiving
"Keep" verdicts without genuine critical analysis.
Changing to `general-purpose` agent ensures evaluation runs on the
conversation's main model (e.g., Opus 4.6), enabling the holistic
judgment that V4 was designed for.
Replace the 5-dimension numeric scoring rubric with a checklist + holistic
verdict system (Save / Improve then Save / Absorb into [X] / Drop).
Key improvements:
- Explicit pre-save checklist: grep skills/ for duplicates, check MEMORY.md,
consider appending to existing skills, confirm reusability
- 4-way verdict instead of binary save/don't-save: adds "Absorb into [X]"
to prevent skill file proliferation, and "Improve then Save" for iterative
refinement
- Verdict-specific confirmation flows tailored to each outcome
- Design rationale explaining why holistic judgment outperforms numeric
scoring with modern frontier models
- Pin installation to specific commit hash (full SHA) to mitigate
supply-chain risk (cubic-dev-ai feedback)
- Add "When to Use", "How It Works", "Examples" sections to match
repo skill format conventions (coderabbitai feedback)
- Add review-before-update instructions for safe version upgrades
- Emphasize zero-runtime-risk: pure Markdown, no executable code
Claude Code sends tool output as `tool_response` in PostToolUse hook
payloads, but observe.sh only checked for `tool_output` and `output`.
This caused all observations to have empty output fields, making the
observer pipeline blind to tool results.
Adds `tool_response` as the primary field to check, with backward-
compatible fallback to the existing `tool_output` and `output` fields.
Extract BIOME_CONFIGS and PRETTIER_CONFIGS as shared constants to eliminate
duplication between PROJECT_ROOT_MARKERS and detectFormatter(). Unify the
biome/prettier branches in resolveFormatterBin() via a FORMATTER_PACKAGES
map. Remove redundant path.resolve() in quality-gate.js.
Handle Windows .cmd shim resolution via spawnSync with strict path
validation. Removes shell:true injection risk, uses strict equality,
and restores .cmd support with path injection guard.
Invoke hook scripts directly via require() when they export a
run(rawInput) function, eliminating one Node.js process spawn per
hook invocation (~50-100ms).
Includes path traversal guard, timeouts, error logging, PR review
feedback, legacy hooks guard, normalized filePath, and restored
findProjectRoot config detection with package manager support.
- Add get_anomaly_attr() helper that handles both dict and object
anomalies. The SDK's send_message() returns dicts, so getattr()
was silently returning defaults -- critical blocking never triggered.
- Fix field name: "detail" -> "details" (matches SDK schema).
- Make fail-open/fail-closed configurable via INSAITS_FAIL_MODE env var
(defaults to "open" for backward compatibility).
- Include exception type name in fail-open log for diagnostics.
- Normalize severity comparison with .upper() for case-insensitive matching.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
cubic-dev-ai P2: dev_mode now defaults to "false" (strict mode).
Users opt in to dev mode by setting INSAITS_DEV_MODE=true.
cubic-dev-ai P2: Move null-status check above stdout/stderr writes
in wrapper so partial/corrupt output is never leaked. Pass through
original raw input on signal kill, matching the result.error path.
coderabbit major: Wrap insAItsMonitor() and send_message() in
try/except so SDK errors don't crash the hook. Logs warning and
exits 0 (fail-open) on exception.
coderabbit nitpick: write_audit now creates a new dict (enriched)
instead of mutating the caller's event dict.
coderabbit nitpick: Extract magic numbers to named constants:
MIN_CONTENT_LENGTH=10, MAX_SCAN_LENGTH=4000, DEFAULT_MODEL.
Also: added env var documentation to module docstring.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Defaults to true (no API key needed) but can be disabled by setting
INSAITS_DEV_MODE=false for production deployments with an API key.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
P1: Log non-ENOENT spawn errors (timeout, signal kill) to stderr
instead of silently exiting 0. Separate handling for result.error
and null result.status so users know when the security monitor
failed to run.
P1: Remove "async": true from hooks.json — async hooks run in the
background and cannot block tool execution. The security hook needs
to be synchronous so exit(2) actually prevents credential exposure
and other critical findings from proceeding.
P2: Remove dead tool_response/tool_result code from extract_content.
In a PreToolUse hook the tool hasn't executed yet, so tool_response
is never populated. Removed the variable and the unreachable branch
that appended its content.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Critical fixes:
- Convert hook from PostToolUse to PreToolUse so exit(2) blocking works
- Change all python references to python3 for cross-platform compat
- Add insaits-security-wrapper.js to bridge run-with-flags.js to Python
Standard fixes:
- Wrap hook with run-with-flags.js so users can disable via
ECC_DISABLED_HOOKS="pre:insaits-security"
- Add "async": true to hooks.json entry
- Add type annotations to all function signatures (Dict, List, Tuple, Any)
- Replace all print() statements with logging module (stderr)
- Fix silent OSError swallow in write_audit — now logs warning
- Remove os.environ.setdefault('INSAITS_DEV_MODE') — pass dev_mode=True
through monitor constructor instead
- Update hooks/README.md: moved to PreToolUse table, "detects" not
"catches", clarify blocking vs non-blocking behavior
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Use local node_modules/.bin/biome binary instead of npx (~200-500ms savings)
- Change post-edit-format from `biome format --write` to `biome check --write`
(format + lint in one pass)
- Skip redundant biome check in quality-gate for JS/TS files already
handled by post-edit-format
- Fix quality-gate to use findProjectRoot instead of process.cwd()
- Export run() function from both hooks for direct invocation
- Update tests to match shared resolve-formatter module usage
Extract project-root discovery, formatter detection, and binary
resolution into a reusable module. Caches results per-process to
avoid redundant filesystem lookups on every Edit hook invocation.
This is the foundation for eliminating npx overhead in format hooks.
- Add insaits-security-monitor.py: real-time AI security monitoring
hook that catches credential exposure, prompt injection,
hallucinations, and 20+ other anomaly types
- Update hooks.json with InsAIts PostToolUse entry
- Update hooks/README.md with InsAIts in PostToolUse table
- Add InsAIts MCP server entry to mcp-configs/mcp-servers.json
InsAIts (https://github.com/Nomadu27/InsAIts) is an open-source
runtime security layer for multi-agent AI. It runs 100% locally
and writes tamper-evident audit logs to .insaits_audit_session.jsonl.
Install: pip install insa-its
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
PR #344 replaced the blocking dev-server hook with auto-tmux-dev.js
which transforms commands into tmux sessions (exit 0) instead of
blocking them (exit 2). Updated 2 tests to match the new behavior.
The "Traction & Distribution" section contained internal business
context (sponsor-call checklists, partner reporting instructions)
that doesn't belong in a user-facing README.
The "Traction & Distribution" section contained internal business
context (sponsor-call checklists, partner reporting instructions)
that doesn't belong in a user-facing README. Moved to docs/business/.
Tatsuya Shimomoto