shiwenbin1617 Show me your work
Use one canonical log for decisions that a later reviewer or agent needs to reconstruct. Ordinary tasks and routine actions do not need a log.
Format
Use the TSV template. Each row records ts, phase, decision, why, evidence, and result. Evidence is a resolvable path, URL, or identifier. Keep cells single-line.
Resolve scripts/log.mjs relative to this skill. Run it with node and arguments <logfile> <phase> <decision> <why> <evidence> <result>. It timestamps rows, writes the header, strips tabs and newlines, and escapes spreadsheet formula prefixes.
Log consequential choices, pivots, verification results, and blockers. For an experimental loop, record each measured iteration. Do not copy illustrative rows into a real log.
Storage and review
Keep the log local at decisions.tsv or .audit/<task-slug>.tsv, unless another task location is specified. Commit only with authorization. Record a superseding row for a changed decision; do not silently rewrite the historical record.
Check claims against the evidence already gathered. Consult this run's transcript only for a specific gap; use ../recall/scripts/find-transcripts.mjs with --host claude and the exact --workspace, never a whole-store search.
An independent review may help for a consequential or explicitly audited run. It is not required solely because a log exists, and no particular model family is mandatory unless configured or requested.
Link the log at delivery and call out material unresolved risks. An empty attention section and a second review of unchanged evidence add no value.