Briefs pointing a worktree-based reviewer at task state must pin the absolute main-checkout path (worktree plans/plan.md resolves to a stale version)

Merged task. This body absorbs task #2418 (filed 2026-08-20 06:00Z from incident #2329) into #2422 (filed 07:58Z from incident #823). Same defect, same target surfaces, two independent incidents two hours apart. #2422 is the canonical executing task; #2418 is archived pointing here. Nothing from #2418 is dropped — its mechanism write-up, its measured table, its three fix directions, and its acceptance criteria are folded in below and marked [from #2418].

Goal

Close a silent-wrong-review hazard in the plan-handoff convention: a brief that names task state (plan, manifest, body, events) by a path that RESOLVES INSIDE a worktree hands the subagent a version frozen at the worktree's base commit. The reviewer then grades the diff against the plan the diff SUPERSEDES, producing confidently-wrong plan-deviation and manifest-completeness blockers with no error, no warning, and no signal anywhere in the marker trail.

The mechanism [from #2418]

CLAUDE.md § Code Style pins the plan-handoff convention: "pass the PATH to the plan, never the body" — and orchestrator briefs accordingly name tasks/<status>/<N>/plans/plan.md, the symlink to the highest plan version. That path is CORRECT at the repo root and STALE inside any worktree cut before the plan was revised: the worktree holds a frozen checkout of the task folder at its base commit, so plan.md points at whatever version existed then, and later versions are absent from the tree entirely. The same holds for artifacts/planned_manifest.json, body.md, and events.jsonl.

Nothing in the pipeline currently catches this. The Step 5a spec-freshness sync (.claude/skills/issue/steps/09-step-5.md) exists for exactly this staleness class but is scoped to the WORKFLOW SURFACE (.claude/** agent specs and skills) — it does not touch the task folder. A reviewer or implementer whose cwd is the worktree reads a stale plan with no diagnostic, and the verdict looks normal.

The worktree's tasks/ tree being frozen at its branch point is CORRECT behavior. The defect is briefs that read it as if it were live.

Incident A — #2329, q35_ladder_decay Step-4 review round 1 [from #2418]

Worktree .claude/worktrees/issue-2329-q35-ladder-decay, cut from a base commit ~3h before plan v5 was written. Both symlink targets and byte counts re-verified with ls -la in each location before filing:

location	plans/plan.md →	versions present	manifest bytes
repo root (main)	v8.md	v1–v8	16,175
the round's worktree	v4.md	v1–v4	8,051

v4 is the PRE-CRITIQUE draft. v4 → v8 spans three post-approval critique rounds and changed exactly what reviewers grade against: the registered verdict lattice and its decision predicates (the negative branch tightened to require a paired scale-normalized companion), the judge-wave call-count declaration (a stale ~7.5k figure re-derived to ~4,720), the R2-M1 donor-geometry and donor-derivation prescriptions, and the manifest's own contents (it gained a 4-label Leg B lattice). A plan-adherence verdict graded against v4 measures adherence to a superseded document and is worthless; a statistics verdict graded against v4's looser predicate can PASS code that violates the registered one.

Every #2329 reviewer that needed the plan detected the stale symlink on its own and re-read v8 from the main checkout, so the round was not corrupted. That is agent diligence compensating for a briefing defect, not a control. The orchestrator learned of it only because one sub-reviewer volunteered a "Note for the orchestrator" line.

Incident B — #823, P-Gen v13 amendment round, caught pre-verdict

The orchestrator hand-wrote a code-reviewer brief naming the plan as tasks/followups_running/823/plans/plan.md and, separately, told the reviewer to read the diff in the worktree .claude/worktrees/issue-823-ladder. Verified resolution of that same relative path:

• worktree .claude/worktrees/issue-823-ladder/tasks/followups_running/823/plans/plan.md → v10.md (title line "Plan v10")
• main checkout /home/thomasjiralerspong/explore-persona-space/tasks/followups_running/823/plans/plan.md → v13.md (title line "Plan v13")

v10 is exactly the version the round's amendment supersedes. It registers the OLD generation caps (1024/2048) and contains NO generation-config fingerprint, NO metadata-free hash-basis pin, and NO P0 prompt-integrity gate. A v10-based review of the v13 diff reads the 4096/8192 caps as an unauthorized deviation and the fingerprint/P0 work as unplanned scope creep — four false blockers, all plan-lens, all requiring a fix round to refute.

The asymmetry that makes this mechanizable. The codex-code-reviewer composer caught this at compose time on its own and handled it (.claude/agents/codex-code-reviewer.md § Step 2-pre-b): it diffs the worktree plan against the canonical main copy, references a byte-identical /tmp copy as primary with the main-checkout ABSOLUTE path as fallback, BARS all worktree tasks/ reads, and supplies sed line-window anchors. The Claude-side brief, hand-written by the orchestrator with no equivalent discipline, did not. Same round, same plan, same worktree; one twin was structurally protected and the other was not.

Recovery was a SendMessage correction to the live reviewer naming both resolutions. That worked only because the Codex composer surfaced the hazard while the Claude reviewer was still running — pure timing, not a control.

The two identified defect sites (Claude side)

1. Brief-composition side — .claude/skills/issue/steps/09-step-5.md ("Both reviewers see the same brief"): the shared brief names "the approved plan (via the plans/plan.md symlink)" while the Claude reviewer's additional fields are worktree path + base ref. That composes exactly into the defect. The Codex twin's adjacent bullet already documents the absent/stale inlining; the Claude bullet has no counterpart.
2. Reviewer side — .claude/agents/code-reviewer.md § Context budget: "plans via Read on tasks/<status>/<N>/plans/v<K>.md (or the path in your brief), sliced" — a relative path resolved from the reviewer's cwd, which on every /issue round IS the worktree.

Why the existing rule does not cover it

CLAUDE.md already carries "Never form tasks/... paths relative to cwd or __file__ — from a worktree that path is stale", enforced by tests/test_no_direct_task_path_construction.py. That rule and its test are scoped to CODE constructing task paths (AST scan for bare-name imports and ROOT / "tasks" constructions). This channel is PROSE: a brief handing a relative task path to a subagent whose working directory is a worktree. The lint cannot see it and the rule's wording does not reach it.

Fix directions (pick after reading the surfaces; do not implement blind) [from #2418]

1. Make the brief-composition convention worktree-safe. Wherever the orchestrator names a plan / manifest / body / events path for a subagent, resolve it against the MAIN checkout absolutely (the existing worktree-safe idiom is already used for workflow helpers: "$REPO_ROOT"/scripts/..., never the worktree copy) — or reference a verified-identical copy outside both trees, the codex-code-reviewer § Step 2-pre-b shape. Resolve the intended version at compose time (readlink) and STATE the version number in the brief so a mismatch is visible to the reader. Explicitly bar reading tasks/ from inside the worktree. Candidate surfaces: CLAUDE.md § Code Style plan-handoff bullet, .claude/skills/issue/steps/09-step-5.md (the shared brief + Step 5a), .claude/agents/code-reviewer.md § Context budget, the implementer specs, and the equivalent v2 dispatch text in .claude/skills/issue-v2/SKILL.md Step 4.
2. Extend the Step 5a freshness sync to the task folder — surgically refresh tasks/<status>/<N>/plans/ + artifacts/ from fetched origin/main in the worktree, mirroring the existing branch-side-edit guard so a legitimately branch-local plan revision is not clobbered.
3. Add a cheap fail-loud version assertion available to any agent handed a plan path: compare the resolved plan.md target against the highest plans/v*.md on origin/main for that task and raise on mismatch. Cheapest of the three and catches the whole class rather than the surfaces enumerated in (1).

(3) plus (1) is probably the right combination — (1) prevents it, (3) catches it when a new brief site is added later and forgets. Decide and record the choice.

Do NOT change plan_patch.py, task.py, the task_workflow resolver, or any worktree mechanics: the worktree's frozen tasks/ tree is correct behavior.

Mechanical-guard idiom. The project's established pattern for pinning a prose contract across multiple surfaces is the region-anchored per-surface token ladder in scripts/workflow_lint.py (check_smoke_blind_spot_review_lens, check_pre_split_review_guard, check_null_gate_calibration_lens, check_two_tier_yield_floor, check_cvd_scoped_gpu_verdict_lens — #2165, #2158, #2144, #2242, #2120), bundled into the no-flags default run. Prefer that idiom over a free-text grep for tasks/<status>/<N>/ near a worktree path, which is false-positive-prone across brief templates.

Acceptance criteria

1. A subagent dispatched with a plan / manifest path from a session whose cwd is a stale worktree either reads the CURRENT version or fails loud — demonstrated with a reproduction (a worktree cut before a plan revision: exactly the #2329 and #823 shapes). [from #2418 #1]
2. The chosen fix covers the MANIFEST as well as the plan; both #2329 artifacts were stale, and a manifest-completeness verdict graded against a stale manifest fails identically. [from #2418 #2]
3. A brief-contract line exists on the Claude-side reviewer/implementer brief surface requiring absolute main-checkout task paths (or a verified-identical out-of-tree copy) + an explicit worktree-tasks/ read bar, with compose-time version resolution stated in the brief.
4. At least one MECHANICAL guard backs the prose — prose alone did not prevent either instance. If a lint check is added it has a test that fails before the fix. [from #2418 #4 + #3]
5. No change to worktree tasks/ semantics; no new tasks/ path construction in code.
6. uv run python scripts/workflow_lint.py (no flags) clean and the mapped select_step9c_tests.py --map-files selection passes.

Folded-in sibling: the implementer brief's missing WAIT MECHANISM

The same #823 round carried a second brief-composition omission of the same class. The implementer brief stated the prohibition ("do not end your turn with work staged-but-uncommitted") but not the WAIT MECHANISM the CLAUDE.md § "Teammate coordination" (g) duty prescribes (a bounded in-turn Monitor until-loop; foreground sleep chains are hook-blocked). The implementer then invented a background-watcher shape and orphaned its own landing — ~62 min of work left uncommitted with a lint gate running unobserved; the orchestrator recovered it.

#2422's original filing recorded this as evidence only and invited a merge ("If a reviewer of this task judges the two worth one combined brief-contract hardening, that is a reasonable merge"). Taking that invitation: it is IN SCOPE here as a second one-line brief-contract addition, because it is the same surface and the same defect class — a hand-written brief omitting a discipline that already exists elsewhere on the surface. Descope it if the plan shows the two additions interfering.

Provenance

workflow_fix_target: CLAUDE.md § Code Style (plan-handoff convention); .claude/skills/issue/steps/09-step-5.md (the shared reviewer brief + Step 5a spec-freshness sync scope); .claude/agents/code-reviewer.md § Context budget (plan-read path); .claude/skills/issue-v2/SKILL.md Step 4

Two independent incidents, 2026-08-20, ~2h apart:

• #2418 (06:00Z, incident #2329) — surfaced during the q35_ladder_decay Step-4 review round 1. Detected by the code-reviewer-lean g6 sub-review, which flagged it to the orchestrator; independently detected and worked around by the efficiency-critic and the g5 sub-review. The orchestrator re-verified the symlink targets and byte counts, then sent in-flight corrections to the seven reviewers still running.
• #2422 (07:58Z, incident #823, parent task) — surfaced during the P-Gen v13 amendment review dispatch. Caught pre-verdict because the codex-code-reviewer composer independently handled the same hazard, exposing the twin asymmetry.

Both filed under the workflow-fix-on-bug protocol: the defect is in the workflow surface (the plan-handoff convention, the Step 5a freshness sync's scope, and the Claude-side brief contract), not in any experiment payload. Merged into #2422 by the /issue 2422 clarifier context-gathering pass (2026-08-20), which found #2418 as a same-fingerprint duplicate on the same target files; #2422 was chosen as canonical because it carried the live session.