[bwachman]

Note on base

This branch is stacked on #1393 (cl-10189-show-required-input-for-primary-session-ask_operator-at-the, tip 8638639), which provides the worker-wait / parked-question machinery this policy integrates with. The diff in this PR is the CL-10432 change only.

Summary

Codify an outcome-aware escalation policy for blocked workers: a closed classification union classifies every ask_director outcome from seven pre-escalation facts, routine questions never park/pending/wake, director_resolvable is parent-only (not operator), operator_decision_required stays distinct from minimum authority, and verification_blocked is a terminal outcome that is never verified, is retained in session state, and has a store-level replace-only-on-explicit-report API (reportVerifiedOutcome, not yet called at runtime, see Deferred). Least-authority is preserved: no permission source change, grant_request_id attaches only the envelope, resolutions deliver text only, and there is no auto-retry.

Verification

After the rebase onto main and the review fixes, CI passes on fd6416e; the results below are from the original verification.

Gate	Result	Exit
bun test src/subagent/escalation-policy.test.ts src/subagent/ask-director.test.ts src/subagent/session-store.test.ts src/subagent/agent-fleet.test.ts src/subagent/lifecycle-tools.test.ts src/subagent/run-ask-director-grant.test.ts src/permission/worker-grant.test.ts src/permission/worker-grant-flow.test.ts src/subagent/fleet-report.ask-wake.test.ts src/tui/worker-wait.test.ts src/tui/agent-ask-wake.test.ts src/tui/runner/wiring.ask-wake.test.ts	355 pass / 0 fail, typecheck clean, dead-exports 0 violations, no regression	0

Both blockers + all 8 should-fix + 5 nits in CL-10432-review-v2.md resolved (opt-in bypass removed, terminalOutcome visible, wait/list exposure, questionId validation, declines record outcome, per-field renderer bounding, question in wake, message/resolution overlap, routine audit, integration seams, nits). Least-authority boundary preserved.

AC to code and test mapping

• seven pre-escalation facts + closed classification union: src/subagent/escalation-policy.ts + escalation-policy.test.ts
• routine no-park enforcement (never parks/pending/wake): src/subagent/ask-director.ts + ask-director.test.ts
• director_resolvable parent-only, not operator: src/subagent/ask-director.ts + fleet-report.ts + tests
• operator_decision_required vs minimum authority distinct: src/subagent/escalation-policy.ts + tests
• verification_blocked terminal outcome (never verified, retained; replace-only store API deferred, see Deferred): src/subagent/session-store.ts + session-store.test.ts (get()/listForStrip())
• least-authority: no src/permission source change; grant_request_id only attaches envelope; resolutions deliver text only; no auto-retry (worker-grant-flow.test.ts, run-ask-director-grant.test.ts)
• wake/panel dedup key sessionId+questionId (assessment compared by value), routine absent from wake: fleet-report.ask-wake.test.ts, worker-wait.test.ts, agent-ask-wake.test.ts, wiring.ask-wake.test.ts

Deferred / Notes

• reportVerifiedOutcome has no production caller: the runtime has no typed "worker verified" signal, so a verification_blocked outcome cannot yet be replaced at runtime. Tracked in CL-10568.
• Durable on-disk persistence of assessments is a separate concern (this ticket retains typed terminal results in session state through complete/list/wait/resume, not across process restart).

Refs CL-10432

[linear-code]

CL-10432

[TheGreatAxios]

Hey @bwachman

Based on reading the PR -- is it possible to trim out the other elements and just be the relevant commits here? If this relies on other code, suggest using gh stack instead.

[TheGreatAxios]

reportVerifiedOutcome has no production caller (interface + tests only); "replace-only on explicit report" for verification_blocked is an uncalled store API, so a verification-blocked outcome can never be replaced at runtime. Wire it up or mark deferred.
Plain: there's a function for replacing a dead-end outcome, but nothing ever calls it — so a dead-end can never actually be replaced. Wire it or say it's coming later.

[bwachman]

Marked deferred rather than wired. The interface method now carries a doc comment saying there is no production caller because the runtime has no typed "worker verified" signal (comment), and the design options for that signal are laid out in CL-10568. I left this thread open for you to confirm that is acceptable.

[TheGreatAxios]

Policy codification verified — the classification union is closed, routine never parks, and least-authority is preserved (no permission-source change, envelope-only grant_request_id, text-only resolutions). Nine items — the two that matter: comparableConfidence overrides classification in routing (the policy's teeth), and reportVerifiedOutcome has no production caller (verification_blocked can never be replaced at runtime). Plus: inconsistent safeDefault rendering, dedup-key vs body claim, "six facts" miscount, routine-ask cap bypasses, stale base note, and a stale comment.

[corbits-triage]

@bwachman, please address the following:

• required checks are failing: test (src), eslint, build-and-test, typecheck, prettier, test (3/4), build, test (4/4), static-analysis, test (1/4), test (2/4)

[bwachman]

On trimming to the relevant commits: this PR is now a single commit (CL-10432 only) stacked on #1393's branch, so its diff shows only its own change. The earlier stack commits are no longer part of this diff.