Axıs
技能目录

review

Review

Router with three jobs kept separate: Audit always, Polish always, Distill on signal only.

1. Audit (Always)

Check the final Blueprint (including v1.x) + §6 line by line. Open by rendering the visible trail — Done ([x] IDs), Now (first [ ] ID), Drifted (last §8 rows) — via forge/scripts/trail.sh when the full triad is installed, else read §5/§8 directly. Verdict is ternary, no middle prose:

  • Pass -> Polish.
  • Conditional pass -> Polish with ID-tagged leftovers for next Shape.
  • Fail -> freeze, route rework (gate C3): exec error returns to Forge (same v1.x); blueprint error returns to Shape Design/Validate; requirement change returns to Define as v2.0. Same issue reworks once; 2nd miss goes to human.

Finding table (from bugbot/security thinking, tool-agnostic):

Severity Location (file:line) Finding

Sorted severity-first. Validate Bot-style reports before acting; retry a failed inspection once only; never auto-fix unless asked.

Example verdicts: Pass — M1 done, no leftovers. / Conditional pass — leftovers: S2-T3 (retry backoff), next Shape first. / Fail — return to Forge: S2-T1 (exec error, same v1.x).

2. Polish (Always, Light)

  • Delete temp code, debug output, unused deps/files; tidy structure and names.
  • Complete README/usage/API notes; pack the deliverable.
  • Append 3-7 point project retro: completion %, key drifts + causes, leftovers/tech debt, next 1-2 actions. No essay, no cross-project rules here.

3. Distill (On Signal Only, Default Off)

Distill answers "what transfers to other projects", never "what happened here". Triggers: explicit user ask, or signal (major rework, >=2 blueprint revisions, recurring issue class, novel working trick, repeated user correction). Misfire control: skip when tasks <5 unless asked.

Output is structured entries for the user's own experience store (see reference), confirmed by the user before filing. Never a second retro essay.

Additional Resources