Research: structured investigation, fact-checking, explanation traces.
The English summary is being prepared. Showing the original description.
See how to installReview the instructions the agent will receive before you install.
Three modes. Select by request signal:
| Signal | Mode |
|---|---|
| Formal research, investigation, sourced report, gather evidence | Research Pipeline |
| Fact check, verify claims, check facts, is this accurate, verify quote | Fact-Check |
| Why did you, explain routing, show trace, decision log, why that agent | Explanation Traces |
Default: Research Pipeline.
Final report. When the final report synthesizes two or more dispatched agents' findings, run one self-review pass on the draft before delivery: skills/shared-patterns/report-self-review.md (separate dispatch, model: opus). It applies to Mode A and Mode B.
Verify every factual claim in a draft before publish. Burden of proof sits on the claim, not the checker. Works standalone or as a pre-publish gate. Non-blocking: the report warns; the caller decides whether to publish.
List every checkable claim: statistics, prices, dates, quotes, attributions, titles, event facts, rankings, causal assertions. Opinions and speculation stay out.
For each claim, record: ID, verbatim text, type (stat/quote/attribution/event/title/price/causal), location. Extract quotes verbatim for exact-words comparison.
Gate: Every checkable assertion has a claim ID. Sweep the document twice.
Work claim by claim.
Gate: Every claim has an evidence record.
Assign each claim one label:
| Label | Assign when |
|---|---|
| Verified | Evidence supports; current within staleness window; sufficient source tier |
| Disputed | Evidence contradicts; newer source supersedes; quote fails any axis |
| Unverifiable | Sources engage the claim but settle nothing |
| Missing-source | No available source addresses it |
Rules: contradiction beats support. Partial verification gets the weakest label. Stale figure superseded by newer = Disputed; merely old with no newer figure = Unverifiable.
Gate: Every claim carries one label and a one-line justification.
# Fact-Check Report: [document]
## Summary
Claims: N | Verified: n | Disputed: n | Unverifiable: n | Missing-source: n
Unchecked: n (reason)
## Per-Claim Findings
### C1 -- [label]
Claim: [text] | Evidence: [source + passage] | Reasoning: [why this label]
## Warnings
[Every Disputed/Missing-source claim with correction]
## Publish Recommendation
[Hold / fix-then-publish / clear]
Every time-sensitive Verified claim carries its as-of date.
Gate: Report covers every claim ID. Warnings lists every Disputed and Missing-source finding.
Read the per-dispatch route event log and present routing decisions as a human-readable timeline. Answer "why did I get routed here?" from recorded events only -- never from reconstruction or rationalization.
Log path: ${CLAUDE_LEARNING_DIR:-$HOME/.claude/learning}/route-events.jsonl (append-only JSONL).
LOG="${CLAUDE_LEARNING_DIR:-$HOME/.claude/learning}/route-events.jsonl"
wc -l "$LOG"
If absent or empty, report the path and the producing hook (hooks/routing-decision-recorder.py). Do not reconstruct from memory.
Parse each JSON line. Two event types: DECISION (one per /do-routed dispatch) and OUTCOME (one per finalized dispatch). See references/trace-schema.md for full field semantics.
Filter to the user's query:
| User signal | Filter |
|---|---|
| Names an agent or skill | DECISION/OUTCOME events matching that name |
| "Why routed here" / latest | Most recent DECISION, current session first |
| Outcome question | OUTCOME events, joined to decisions |
| No specific target | Chronological timeline, most recent session |
Join OUTCOME to DECISION on same session AND key == "{agent}:{skill}". File adjacency is unreliable.
Sort by ts. For each decision, show: time, agent+skill, complexity, request snippet, health at decision (three states: numeric, no-weight-row, legacy), alternates, outcome.
Lead with the answer to the user's specific question, then offer surrounding context. Flag gaps honestly: pre-instrumentation entries, unmatched outcomes.
request_snippet is private session data: show to the session's user, keep out of PR bodies, issues, exports.
| Signal | Reference | Content |
|---|---|---|
| Event field schema, health states, join rules | references/trace-schema.md | DECISION/OUTCOME field semantics |
| Diagnosing thin trace data, consumer mistakes | references/preferred-patterns.md | Failure mode catalog for log reading |
| Parse/read errors, missing log, unmatched outcomes | references/error-handling.md | Error-fix mappings for trace reading |
No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Manage multiple Claude Code accounts: add, list, check, launch, and install shell aliases for 10+ isolated CLAUDE_CONFIG_DIR profiles.
The English summary is being prepared. Showing the original description.
Improve architecture across modules by deepening interfaces.
The English summary is being prepared. Showing the original description.
Assessment: read-only inspection, codebase overview, value analysis, health checks, ADR consultation, decision analysis, multi-perspective critique.
The English summary is being prepared. Showing the original description.
Background memory consolidation — overnight review, merge, and injection payload for memory files.
The English summary is being prepared. Showing the original description.
Jev-driven browser automation: Jev picks operations, programs execute, a text model writes field values only when Jev cannot pick one from the goal.
The English summary is being prepared. Showing the original description.
Write, compose, integrate, and improve programs that call Jev, TypeSafe's System One judgment model.
The English summary is being prepared. Showing the original description.