Director prompt hygiene: impossible tool refs, triple-stated report contract, dead no-ship branch, persona coupling

[TheGreatAxios]

Closes CL-6807.

This is a hygiene-only pass on director prompts: removes what is wrong or duplicated, never anything load-bearing-but-advisory. Does not touch the skywalker coordination prose (4-worker cap, lane disjointness, re-fix caps, anti-cascade block — that removal is CL-6953, explicitly last in the chain). Does not touch worker style/philosophy guidance (CL-6803, human-gated).

Diff is prompt text plus a small amount of structural plumbing to remove a persona-string check — please read the wording, not just the mechanics.

Defects and status

1. Impossible/stale tool references — all present, all fixed

• tester/testsmith mount READ_TOOLS only (no write tools), but were told "Do not write_file/edit_file product code" / could write "optional notes under tmp/". Removed both — replaced with plain statements that no product-mutation tools exist.
  • tester, before: Do not write_file/edit_file product code. → after: you have no product-mutation tools, so there is no way to patch source to make green.
  • testsmith, before: Design in the report (and optional notes under tmp/ only if the brief allows). → after: Design in the report.
• critique mounts REVIEW_TOOLS which always includes run_shell, but hedged if shell is allowed.
  • before: Prefer reading tests/callers; if shell is allowed, a tiny sync call that would hang on a Promise is evidence. → after: Prefer reading tests/callers; a tiny sync call via run_shell that would hang on a Promise is evidence.
• emil/draper/gaasbot explained at length that write tools are denied, when the tool set already makes that true by omission.
  • emil, before: Verify — inspect code, run existing tests/linters when practical. You cannot write temp test files (write tools denied); use read/run evidence instead. → after: Verify — inspect code, run existing tests/linters when practical; use read/run evidence, not temp test files.
  • draper, before: Never write/edit/delete product files. Never spawn. Never commit. → after: Never commit.
  • gaasbot, before: Load philosophy when judgment trade-offs matter. Stay advice-only — no write_file/edit_file/delete_file. → after: Load philosophy when judgment trade-offs matter. Stay advice-only.
  • neckbeard, before: Do not apply fixes. Do not write, edit, or delete product files. Do not spawn agents. → after: Do not apply fixes. (spawn ban covered below)
• plan mounts REVIEW_TOOLS (no ask_operator, ever) but hedged the tool as "available" sometimes.
  • before: Prefer interview skill when requirements are fuzzy (ask_operator / structured questions when available). → after: When requirements are fuzzy, note the open questions under Blockers instead of guessing — you cannot ask the operator mid-run.
• Added src/agent/directors/registry.test.ts test "no director prompt references a tool its tool set does not mount" that scans every director's prompt text for known tool-name tokens and fails if a mentioned tool isn't in that director's mounted set (or the small always-mounted/spawn-conditional allowances). Skywalker is excluded — it's the primary session's role text, not a filtered sub-agent.

2. Triple-stated report contract — present, fixed
Every leaf's hand-written trailing line Report: Summary, Findings (...), Blockers, Paths. was a third restatement on top of (a) report.requiredSections (structural) and (b) the shared buildSubAgentReportContract() envelope appended to every sub-agent prompt. Removed the hand-written line from build, bruckheimer, explore, gaasbot, plan, neckbeard, greybeard, intern, shakespeare, tester, testsmith, critique, and emil — kept any domain-specific hint (e.g. "risk/sequencing advice") as a short standalone Findings: ... sentence where it added real information.

Also removed the redundant Never spawn. / Do not spawn agents. / Do not spawn sub-agents. lines from build, neckbeard, brand-reviewer, and draper — each already declares spawn: { maySpawn: false }, which structurally means the task tool is never mounted, and the shared sub-agent appendix already states the no-recursion rule in prose. Never commit. is kept where present — it isn't structurally enforced (these leaves keep run_shell), so it's a real instruction, not a duplicate.

Directors with real per-field report guidance (draper, brand-reviewer, emil's original "# Report" section) were left alone — that's domain content, not a restatement of the headings.

3. Dead no-ship branch — present, fixed
SubAgentDirector.decide in src/subagent/nudge-director.ts branched on stop === "no-ship", but evaluateSubAgentStop (src/subagent/stop-policy.ts) never returns that value — only evaluateThrashStop (which only returns "report-forced" or null) and the explicit never-acted/never-edited/incomplete-report/turn-budget/no-progress paths run. The branch was unreachable. Removed "no-ship" from SubAgentStopReason and from the nudge director's switch/checkpoint mapping. Left ForcedStopReason (a separate type used by forcedStopReport and brief-dispatch.ts's salvage classification, which genuinely does produce/consume "no-ship" reports) untouched — those tests still pass unchanged.

4. Persona-string coupling — present, fixed
shouldRequireEvidence in src/subagent/run.ts armed the empty-readCounts gate via systemPromptRole.includes("CritiqueDirector") — a rename of the persona or its identity-header text would silently change enforcement. Threaded the already-resolved closed-director id (resolvedDirectorId in task-tool.ts) through RunSubAgentParams.directorId and switched the gate to directorId === "critique".

Tests updated

• src/agent/directors/registry.test.ts — new tool-reference cross-check test.
• src/subagent/index.test.ts — shouldRequireEvidence tests rewritten to assert on directorId instead of constructing/matching systemPromptRole text; added a case for "no directorId resolved".
• No director package.test.ts needed changes — they already assert against report.requiredSections and tools.allow, not the removed prose.

Gate

bun run check (prettier --check, eslint, tsc --noEmit, build, full test suite) is green: 0 lint errors (1319 pre-existing warnings, untouched by this diff), typecheck clean, build succeeds, 5331 tests pass across 383 files.

[linear-code]

CL-6807

[TheGreatAxios]

Reviewed clean — holding the merge until immediately after the 0.2.106 tag, then this lands first in 0.2.107.

Reason is sequencing, not quality. This rewrites prompt text across 13 directors, which materially affects the inference loop. The new four-tier eval suite (#577) has no baseline yet — the 0.2.106 run establishes it. If this merges first, the baseline silently includes the prompt churn and we lose the ability to measure whether it helped or hurt. Landing it into 0.2.107 means it gets measured against a clean baseline, which is the whole point of resetting the suite.

Nothing to do on your side; I will merge it as soon as the tag is cut.