[coder-xieshijie]

What changes

• feedback (new, manual): stage D. After a delivery, or over a span of time:
  1. gathers the sessions, MR, CI and delivery records;
  2. it and a model from another family each mine them with one brief (seven lenses, including decision points by root cause and rules that were written but not followed);
  3. a fresh subagent recomputes every number;
  4. each gap is routed, and recorded in the user's ledger (a recurrence reopens its entry; closed entries are checked when this requirement can show them);
  5. the user confirms the table once;
  6. approved entries land, one Skill change per PR.
• scripts/sessions.mjs: read-only extract of the user's side of Claude Code, Codex (app database and CLI rollouts) and MCode sessions, and the collaboration numbers a time-window run compares. Tested in CI.
• deliver: the report ends by pointing to feedback.
• README (both), basis, glossary: stage D names feedback; "feedback ledger" is a glossary term.

Basis

• The stage had no Skill: README's D row was "—", and the gaps deliver lists in plan.md were mostly not followed up.
• The design follows the feedback-stage study (five delivered features) and its draft plan. The changes from that draft come from a two-month analysis of 1,442 sessions:
  • the number check: the analysis's own check found 11 wrong numbers and one cause credited to a rule added two days later;
  • the window mode;
  • the "rules written but not followed" lens with its five causes;
  • the rule that a second occurrence a command or hook can decide becomes a mechanism.
• Rules it relies on:
  • agent-prompt-rules: the author is not the only judge (independent mining by another family); change one component at a time, compared before and after (Skill entries).
  • basis §2: enforce by mechanism.

Verification

• node --test skills/feedback/scripts/sessions.test.mjs: 5 tests on fixture stores for all four sources (fork copies, subagents, codex exec, MCode child sessions and machine turns skipped; window half-open; stats).
• Real window: extract + stats over 2026-10-12 to 2026-10-25 on the author's machine gave a median of 4 sessions messaged in the past 30 minutes and a 74.6% switch rate. The earlier hand-run analysis of the same window found 4 and 74.4%.
• Dry run on a delivered requirement (a quick-path Bunny feedback change, merged), in a fresh agent, steps 1–4, stopping before confirmation:
  • Gaps: 15, of which 2 were confirmed recurrences of ledger entries, 2 possible recurrences, 9 new and 2 no-change. One closed entry was confirmed closed.
  • Cost: 26 minutes.
  • Number check: corrected 5 of 22 numbers, including one wrong causal claim that, once corrected, revealed another gap.
  • The run's notes on the Skill text were applied before this PR:
    • read session ids from the ledger first, with --grep only to fill gaps;
    • the owner's side is in the transcripts;
    • --skip-git-repo-check for the miner;
    • the ledger's own archiving duties;
    • same-cause matching, with possible recurrences shown to the user;
    • closed-entry checks only where observable;
    • what impact means;
    • one-offs left out;
    • mechanisms may live in Skill scripts;
    • the backlog entry is optional.
• Not yet verified: steps 5 and 6 (confirmation and landing) and a full time-window run end to end; the deliver pointer sentence was not compared in runs.
• Link checks and all existing script tests pass.