Filed by the skills seat (session session_012Gg6rMAti8ZaueWb6BRsDn), 2026-08-12. This is design-round decision D of #7885's joint design round, approved by the maintainer (verbatim: 「同意」 to the four-point proposal whose point D was "os-dev.md follows in its own card, not mixed into #7885").
Why
.claude/agents/os-dev.md is 686 lines / 44,495 chars ≈ ~11k tokens, loaded by every dev agent session (subagent and cloud alike) — the dev-side twin of the PM-side cost #7885 attacks. The same 2026-08-12 rulings apply (「只需要说原则,不需要写细节」;「处理 issue 时犯的错应该总结成经验,保留 issue id没有意义」).
Scope — same three-way sorting rule as #7885
- Mechanize: details already enforced by hooks/gates (worktree guard, stash guard, duplicate-fix guard, changeset checks) lose their prose twins — one principle line each.
- Delete: procedural narration and incident storytelling; lessons distilled in place (failure mode + discipline + boundary), no issue-ID citations, no dereference expectation.
- Keep as data: operational lookups only (e.g., the report-contract JSON shape, gate-family names).
Target: ~200–300 lines (from 686). ID lint from #7885 covers this file too.
Constraints
Filed by the skills seat (session
session_012Gg6rMAti8ZaueWb6BRsDn), 2026-08-12. This is design-round decision D of #7885's joint design round, approved by the maintainer (verbatim: 「同意」 to the four-point proposal whose point D was "os-dev.md follows in its own card, not mixed into #7885").Why
.claude/agents/os-dev.mdis 686 lines / 44,495 chars ≈ ~11k tokens, loaded by every dev agent session (subagent and cloud alike) — the dev-side twin of the PM-side cost #7885 attacks. The same 2026-08-12 rulings apply (「只需要说原则,不需要写细节」;「处理 issue 时犯的错应该总结成经验,保留 issue id没有意义」).Scope — same three-way sorting rule as #7885
Target: ~200–300 lines (from 686). ID lint from #7885 covers this file too.
Constraints
.claude/agents/**is not literally inside the two ruled skill roots, but an agent definition governs every dev session's behavior — treated conservatively as ADR-class (draft PR, human merge) unless the maintainer rules otherwise.model: claude-fable-5(instruction-architecture card, same tier rationale as the pm-dispatch SKILL structural optimization: role-split loading, issue-ID removal (lessons distilled in place), probe-first startup (maintainer-directed) #7885 chain; not the mandatory-clause path, recorded as the seat's explicit tiering decision).