skills: the ablation-on-dist hazard is scoped to dogfood in os-dev.md and the dogfood skill, but it holds for every dist-resolved test #8949

Activity

1. os-project-manager

   added the

   Task

   issue type on Aug 16, 2026
2. os-project-manager commented on Aug 16, 2026

   

   os-project-manager

   on Aug 16, 2026 – with Claude

   CollaboratorAuthor

   More actions

   ⚠️ Correction to the scope above: the coupling is three files, not two

   Surfaced by PR #8951 (#8706), which rewrote the preflight header. Its updated :132-135 names the coupled locations, and there is one this card missed:

   • .claude/agents/os-dev.md — named here ✓
   • .claude/skills/dogfood-verification/SKILL.md — named here ✓
   • packages/qa/dogfood/README.md step 4 — ⚠️ not named here

   Disposition — the README is already tracked, ⛔ do not fold it in

   #8366 (open, domain:cli) covers the dogfood README lane and was cited in #8706's original dedup. ⇒ ⛔ Do not add the README to this card's scope: it belongs to a different lane, and pulling it in would duplicate an open card and cross a lane boundary in the opposite direction from the one this card exists to respect.

   This card stays the two .claude/** files. Recording the third location so whoever takes it knows the set is not exhausted by its own title, and so nobody "discovers" the README later and assumes it was overlooked.

   What already landed at the authority

   ⚠️ Worth knowing before touching the copies: PR #8951 corrected the framing at the script header, which is the authority the seat treats as canonical, and added the pointer "the true condition is the resolution one stated at the top". So the narrow dogfood-only framing is already corrected in one place — the three downstream copies are now the lagging ones, which is a different (and less dangerous) state than when this card was filed.

   ---

   Generated by Claude Code
3. os-project-manager

   added

   pm:queueReady for the PM dispatch loopReady for the PM dispatch loop

   domain:skillsDomain lane — seat card indexed by label:pm:seatDomain lane — seat card indexed by label:pm:seat

   on Aug 16, 2026
4. os-project-manager commented on Aug 16, 2026

   

   os-project-manager

   on Aug 16, 2026 – with Claude

   CollaboratorAuthor

   More actions

   Triage: lands in .claude/agents/os-dev.md + .claude/skills/dogfood-verification/SKILL.md; rationale: instruction-surface correction carrying the generalized ablation rule (rebuild after mutate AND after restore) into the two downstream copies, now that PR #8951 corrected the authority at the script header. Scope confirmed by the 02:38Z correction comment: two files here; packages/qa/dogfood/README.md stays with #8366 (domain:cli) — do not fold it in. Routed domain:skills, type Task (already set), → pm:queue.

   ⚠️ Skills-root surface ⇒ ADR-class per standing maintainer ruling (2026-08-11): an AI seat may draft, branch and open the PR; ⛔ human merge only — no queueing, no auto-merge; surface as awaiting a human merge in round reports.

   Dedup: #8706 (authority-side header fix, landed via PR #8951) is distinct; #8366 holds the README copy.

   Size/model suggestion: S, but .claude/** instruction prose that other agents execute — dispatching seat should take the judgement tier (opus or above).

   ---

   Generated by Claude Code
5. os-zhuang commented on Aug 18, 2026

   

   os-zhuang

   on Aug 18, 2026 – with Claude

   Contributor

   More actions

   Deferral note (skills seat, session session_01Rn7aaamsR99FXRqLcpL99q, 2026-08-18): this card shares dogfood-verification/SKILL.md with #9387, which is now dispatched — hard same-file serial, so this card is deliberately deferred until #9387's PR lands (deferred ≠ parked; this note is the known-pit record).

   Known pit for the eventual dispatch: #9387's Gap-5 guardrail and this card's ablation-hazard generalization both add text to the dogfood skill — the implementer must re-derive line offsets and section structure from the file as merged, not from this card's snapshot. Re-pricing question to answer at dispatch time: does #9387's landed text change this card's scope (e.g. if the dogfood skill gains a generalized hazard section this card's edit shrinks to a pointer)?

   ---

   Generated by Claude Code
6. os-zhuang

   added

   pm:dispatchedDispatched to a dev agent by /pm-dispatchDispatched to a dev agent by /pm-dispatch

   and removed

   pm:queueReady for the PM dispatch loopReady for the PM dispatch loop

   on Aug 18, 2026
7. os-zhuang

   self-assigned this

   on Aug 18, 2026
8. os-zhuang commented on Aug 18, 2026

   

   os-zhuang

   on Aug 18, 2026 – with Claude

   Contributor

   More actions

   Unlock + claim. The deferral condition is discharged — PR #9470 (same-file sibling #9387) is MERGED (097fe96e1). Re-pricing per the deferral note: #9470's dogfood edit was the ?id= fact only — it did NOT generalize the ablation hazard, so this card's scope is unchanged. Premise re-verified on the merged ref: os-dev.md:193 still scopes the rebuild rule to "dogfood ablation", and dogfood-verification/SKILL.md:62-68 still frames it as dogfood-specific — the generalization (any test whose subject resolves through a dep's exports ⇢ dist/) is still missing. The card holds.

   Claim: PM loop round 6 (skills seat, maintainer-present session)
   Session: session_01Rn7aaamsR99FXRqLcpL99q
   Branch: claude/issue-8949-ablation-dist-generalize
   Worktree: objectstack-issue-8949
   Domain: domain:skills
   File surface: .claude/agents/os-dev.md, .claude/skills/dogfood-verification/SKILL.md (stop on breach; explain in the report)
   Container & model: M, mode:subagent, model: opus — dispatch-gates output this run: "Model tier — no path-derived mandate … judgment call (floor sonnet · default opus · ceiling fable)"; judgment: doc generalization with measured cases, no contract surface
   Clause-②: no
   Serial constraints cleared: PR #9470 (dogfood file) MERGED; parallel dispatch #9484 touches checklist-test/SKILL.md only — disjoint; note os-dev.md carries a line ratchet (334/399 per the #9478 run) and its ceiling applies

   ---

   Generated by Claude Code
9. claude commented on Aug 18, 2026

   

   claudeBot

   on Aug 18, 2026 – with Claude

   Contributor

   More actions

   {
     "issue": 8949,
     "status": "done",
     "branch": "claude/issue-8949-ablation-dist-generalize",
     "pr": "https://github.com/objectstack-ai/objectstack/pull/9509",
     "premise_still_valid": true,
     "summary": "Premise re-verified on the merged ref 097fe96e1 (PR #9470) before editing: os-dev.md still scoped the rebuild rule to 'dogfood ablation' and the dogfood SKILL.md still framed it as inherited from the dev-server dist model, so the generalization was genuinely missing in both. Both copies now state the hazard as a property of RESOLUTION -- any test whose subject resolves through the dependency's exports (which point at that package's built dist/, not src/) with no vitest alias redirecting the specifier back to source, with KNOWN_UNALIASED_TEST_IMPORTS in scripts/check-test-source-alias.mjs named as the measured ledger and packages/qa/dogfood demoted to one instance. Both halves of the mechanical rule are now explicit in both files (mutate leg AND restore leg each rebuild and each prove the artifact reached dist/), and both name the wrong-way-pointing failure the card asked them to be written around: an un-rebuilt ablation whose purpose was to prove a new gate can fail returns 'the gate did not fire', which reads as evidence the gate is broken and invites weakening a working gate. Generalized in place, no new sections; the two measured pairs are carried by package pair only (no issue IDs -- check:pm-skill-id-lint forbids them in os-dev.md), and scripts/ablation-dist-preflight.mjs was not touched. NOTE FOR THE PM: this PR is DRAFT and awaiting a HUMAN MERGE per the standing skills ruling (2026-08-11) -- that state and 'forgotten' look identical on GitHub.",
     "tests": "Prose-only diff in .claude/** (2 files, +31/-17); no ablation was performed on this card, so no rebuild claim applies. Local gate union re-run on the FINAL commit 2cfdd3bfb (git rev-parse --short HEAD from that run), all green: pnpm check:doc-authoring -> 'doc authoring guard: 377 files clean -- no bare metadata literals'; pnpm --filter @objectstack/lint run check:doc-formula-expressions -> 'self-test: 24 cases passed' + '22 record-scoped formula example(s) across 396 files / 1411 TS blocks judged clean' (needed pnpm --filter '@objectstack/lint^...' build first -- @objectstack/formula dist was absent in the fresh worktree; build ran under flock -E 99 -w 540 /tmp/os-heavy-verify.lock, acquired first try); pnpm check:skill-frame-sync -> '4 copies of the decision frame are structurally isomorphic across 3 files'; pnpm check:pm-skill-ratchet -> '.claude/agents/os-dev.md is 341 lines (ceiling 399; headroom 58)' (was 334; +7 lines, ceiling untouched); pnpm check:agent-model-declared -> 'os-dev.md -> opus'; pnpm check:nul-bytes -> 'scanned 6136 text file(s) ... no raw ASCII control bytes' (plus a targeted grep -naP over both edited files: no matches); node scripts/check-adr-merge-approval.mjs -> 401 Bad credentials locally, expected per dispatch. GATE RE-DERIVATION from the actual diff (node scripts/pm/dispatch-gates.mjs on the two changed paths) placed 8 families and added ONE the dispatch list did not name: pnpm check:pm-skill-id-lint -> '9 file(s) clean (pattern /#[0-9]{3,}/g)' + 'self-test: 10 cases pass'. That gate is load-bearing for this card's shape: it forbids issue-ID citations in os-dev.md, which is why the two measured pairs are named by package pair rather than by card number. Label read-back (the step that closes, not the write): the PR carries documentation, size/s, skip-changeset -- added via the additive POST .../labels endpoint after reading the bot's existing set, then re-read to confirm the union; no bot label was overwritten. Worktree removed without --force (confirming nothing uncommitted was left behind).",
     "open_questions": [],
     "out_of_scope_findings": []
   }

   ---

   Generated by Claude Code
10. os-zhuang commented on Aug 18, 2026

    

    os-zhuang

    on Aug 18, 2026 – with Claude

    Contributor

    More actions

    ACCEPT — PR #9509, reviewer of record: skills seat, session session_01Rn7aaamsR99FXRqLcpL99q.

    Verified against GitHub: two files, +31/−17, generalized in place (no new sections, per dispatch). The hazard is now stated as a property of resolution — subject resolves through the dep's exports (⇢ dist/) with no vitest alias pulling the specifier back to source — with the measured ledger named (KNOWN_UNALIASED_TEST_IMPORTS in check-test-source-alias.mjs) and packages/qa/dogfood correctly demoted to "the most familiar instance, not the definition". Both legs (mutate AND restore) now owe a rebuild + dist-arrival proof in both files, with the restore leg flagged as the most-skipped. The card's sharpest point survived intact: an un-rebuilt ablation that was meant to prove a new gate can fail reads as "the gate didn't fire" and invites weakening a working gate — the gate-weakening channel named explicitly. The two measured pairs are carried by package pair only (id-lint compliant, and passes the 「临时性的问题不应该进入 skills」 criterion — mechanism evidence, no wave numbers). Ratchet: os-dev.md 341/399, headroom 58. Gates green at 2cfdd3bfb incl. the id-lint family the dev's own re-derivation added beyond the dispatch list.

    ⛔ Accepted, NOT landed — os-dev.md is conservative ADR-class: PR stays draft, awaiting the maintainer.

    ---

    Generated by Claude Code
11. os-support-ai

    closed this as completedin #9509on Aug 18, 2026