Filed by the skills seat (session session_012Gg6rMAti8ZaueWb6BRsDn), 2026-08-12, on the maintainer's approval in the seat 专题 chat (verbatim: 「同意你的全部建议」, in response to the token-waste inventory that ranked AGENTS.md as the largest remaining lever after os-dev.md).
Measured baseline
AGENTS.md: 1,050 lines / 84,665 chars ≈ ~20k tokens — read in full by dev agents directed to it, per session. The same 2026-08-12 rulings apply (「只需要说原则,不需要写细节」;「保留 issue id没有意义」).
Scope — same three-way sorting rule as the SKILL.md / os-dev.md rewrites
- Mechanize ⇒ one principle line: everything the hooks/gates already enforce.
- Delete: incident narratives and procedural walkthroughs; lessons distilled in place (failure mode + discipline + boundary), no issue-ID citations; maintainer rulings keep date + verbatim quote.
- Keep as data: operational lookups (dev-stack commands, PR-flow steps that are commands rather than prose).
No-silent-semantic-loss red line applies: Prime Directives survive whole; a dropped-with-rationale list in the PR body for line-by-line review.
Cautions
- AGENTS.md is a human-facing document too (contributors read it, not only agents) — the rewrite must stay readable as the repo's source of truth, not become agent-telegraphese; where human onboarding narrative genuinely helps, it may survive the sort.
- Conflict rule to preserve verbatim: pm-dispatch SKILL defers to AGENTS.md (「AGENTS.md wins」) — the rewrite must not invert or blur that.
- CLAUDE.md's four inlined rules mirror AGENTS.md sections — keep the mirror consistent in the same PR.
- Treated as ADR-class conservatively (repo-wide agent governance): draft PR, human merge.
- Dispatch
model: claude-fable-5 (instruction-architecture card; seat tiering decision).
- Extend the
check:pm-skill-id-lint scan set to AGENTS.md in the same PR (with a legacy exact-count waiver if the two PRs need order independence — the pattern from the os-dev rewrite).
Filed by the skills seat (session
session_012Gg6rMAti8ZaueWb6BRsDn), 2026-08-12, on the maintainer's approval in the seat 专题 chat (verbatim: 「同意你的全部建议」, in response to the token-waste inventory that ranked AGENTS.md as the largest remaining lever after os-dev.md).Measured baseline
AGENTS.md: 1,050 lines / 84,665 chars ≈ ~20k tokens — read in full by dev agents directed to it, per session. The same 2026-08-12 rulings apply (「只需要说原则,不需要写细节」;「保留 issue id没有意义」).Scope — same three-way sorting rule as the SKILL.md / os-dev.md rewrites
No-silent-semantic-loss red line applies: Prime Directives survive whole; a dropped-with-rationale list in the PR body for line-by-line review.
Cautions
model: claude-fable-5(instruction-architecture card; seat tiering decision).check:pm-skill-id-lintscan set to AGENTS.md in the same PR (with a legacy exact-count waiver if the two PRs need order independence — the pattern from the os-dev rewrite).