菜单
Execution trace capture and token baselining. Use when user says trace, extract trace, count tokens, token baseline, or /harness-trace. Not for harness optimization (use harness-mechanic).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Release prep and tagging with pre-flight verification. Use when user says release, tag, ship it, push to production, create release, or bump version. Not for editing a version string outside a release, a lockfile or a changelog entry by hand, and not for reading what changed in a past release: those are ordinary edits, just do them.
Mine LEARNING.md retrospectives across all repos for recurring failure-modes, then propose harness changes (hooks, rules, skills, checklist items) with falsifiable change-contracts. Also runs the doc-gardening pass over the governance docs (stale rules, dead cross-references, skill-table drift). Use when user says learning loop, mine learnings, recurring failures, what should I fix in my process, cross-repo retrospective, doc gardening, stale docs, or run learning-loop. Runs on human schedule, never autonomously. Not for single-project retrospectives (use learning-docs) or trace/token analysis (use harness-mechanic).
Full contractor-mode loop: sub-protocols (LOCALIZE, BENCH-BASELINE, REPRODUCE, DRIFT), review routing by file pattern, blast radius, UAT, parallelism and effort caps, escalation, just-do-it, goal-backed runs. Load it to RUN that loop, once you have an approved plan and are entering step 1, or when a step needs its detail. It is a protocol, not a way to do the work: it never replaces the language, review or planning skill the task itself calls for, and a task small enough to just edit (SKIP_SET) needs none of it. The always-on spine lives in rules/orchestrator-protocol.md. Not for trace capture (harness-trace) or harness optimization (harness-mechanic).
Ruby on Rails with service-oriented architecture, Dry-validation, Sidekiq/Solid Queue, Hotwire. Use for Rails API, Rails services, Rails forms, RSpec, ActiveRecord, Rails migrations. Not for standalone Ruby gems (use ruby skill), and not for judging whether a migration is safe to run in production (that is the database-reviewer agent).
Get three independent second opinions (isolated Claude + isolated Gemini + isolated DeepSeek) on a problem Claude is analyzing. Use when user says second opinion, another perspective, challenge this approach, or asks gemini/deepseek about a decision, approach, or diagnosis. Also invoke on your own initiative, without being asked: on a complex task before planning, after 2 failed root-cause attempts, on a 🔴 decision (rewrite, core logic, security), or when the review loop is stuck below 80 after 2 fix rounds. Gathers context, writes a focused prompt, calls all reviewers in isolated Docker containers, and synthesizes all viewpoints. Not for reviewing code or changes, even phrased as 'ask gemini to review' (use gemini-review).
Assess test suite quality using Farley's 8 Properties and Tautology Theatre detection. Use when user says review tests, test quality, are my tests good, test assessment, or test design review. Not for writing tests (use language skills) or code review (use gemini-review).
| name | harness-trace |
| description | Execution trace capture and token baselining. Use when user says trace, extract trace, count tokens, token baseline, or /harness-trace. Not for harness optimization (use harness-mechanic). |
| compatibility | Python >=3.11, uv |
Structured execution trace capture and token baselining for the claude-forge harness. Based on Meta-Harness (arxiv 2603.28052): measure before optimizing.
cd skills/harness-trace
# Install
uv sync
# Extract traces from a session JSONL
uv run -- harness-trace extract ~/.claude/projects/PROJECT/SESSION.jsonl
# Count tokens in harness files
uv run -- harness-trace count-tokens rules/
# Generate full baseline
uv run -- harness-trace baseline --base-dir /path/to/claude-forge
| Command | Purpose | Output |
|---|---|---|
extract <session.jsonl> | Parse session into trace JSONL | Trace entries (stdout or file) |
count-tokens <path> | Count tokens per file | Token counts per file |
baseline --base-dir <dir> | Full harness token baseline | TSV with tier classification |
extract uses three signal sources, in order of precision:
Tool-use blocks (primary, high precision). Walks every tool_use block in the assistant messages, then resolves outcomes from the paired tool_result blocks (matched by tool_use_id) in the user messages:
| Tool / pattern | Emits |
|---|---|
Agent(subagent_type=*-reviewer) | ROUTE + REVIEW; findings parsed from the reviewer's returned report (section-style ### MAJOR + one bullet per finding, or explicit MAJOR: 1 counts) |
Agent(subagent_type=*) (non-reviewer) | ROUTE |
Bash matching pytest|make test|go test|npm test|cargo test|rspec | VERIFY, tests_pass resolved from the result |
Bash matching make check | VERIFY, lint_clean resolved from the result |
Edit / Write / MultiEdit | counted into IMPLEMENT aggregate |
WebFetch to arxiv|docs | RESEARCH |
AskUserQuestion | counted into safety_compliance.hitl_gates_hit |
Outcome resolution is fail-safe: a result that never arrives (truncated session, async agent) or says nothing about an axis leaves it unknown; failure markers in output override a clean exit (pytest || true chains); counts are never fabricated from prose.
Literal report lines (exact formats mandated by rules/orchestrator-protocol.md): LOCALIZE: planned=… proposed=… → LOCALIZE, REPRODUCE: script=… fails_before_fix=… → REPRODUCE, DRIFT: subtask=… verdict=… → DRIFT_CHECK, BLAST-RADIUS: clean|MAJOR=n MINOR=m|skipped (…) → BLAST_RADIUS, EXECUTOR: <executor> model=<id> subtask=<id> → EXECUTOR. Detection and extraction share one compiled pattern per step; lines are checked independently per message (the protocol co-locates them). BLAST_RADIUS is keyed ONLY on this line: ast-grep usage is not a signal (the always-use-sg rule saturated it). Fenced code blocks are stripped first so quoted lines cannot forge events.
Text regex (fallback prose heuristics). Applied only to messages whose tool stream did not already surface the step, and only for steps that are typically pure text (SCORE, FIX, LOOP, UAT, plus VERIFY/REVIEW when no tool signal exists). This avoids the v1 problem where prose like "tests pass" in a chat triggered fake VERIFY entries.
SUMMARY is always synthesized last from JSONL aggregates: duration_min (last_ts - first_ts), files_changed (Edit+Write count), and the 6-dimension metrics object (paper §5.2.1). No SUMMARY is emitted for sessions with zero other entries.
See references/schema.md.
file tokens words lines tier loaded
rules/orchestrator-protocol.md 1247 450 67 rule always
skills/golang/SKILL.md 2890 1050 145 skill on-demand
Tiers: rule (always-on), skill/skill-ref (on-demand), agent (on-demand), config (always-on).
See references/troubleshooting.md.
See references/schema.md.