🤖 refactor: extract a pure goal continuation policy from workspaceGoalService

[ibetitsmike]

Summary

Extracts the goal-continuation decision logic out of WorkspaceGoalService into a pure, synchronous policy module (goalContinuationPolicy.ts). The service becomes the I/O shell (file locks, goal-file reads/writes, chat-tail sync, analytics/timeline emission) around a single decision point that returns a discriminated union: continue | defer(untilMs) | stop(reason, dropCandidate).

Background

workspaceGoalService.ts (5,401 lines) interleaved continuation decisions with I/O: checkGoalContinuationEligibility mixed in-memory candidate maps, runtime-bridge probes, gated goal-file reads, normalizeGoalLimits (a locked read-modify-write with chat-tail reconciliation), one-shot budget wrap-up gating, and cooldown math. The decision had no interface of its own, so testing one cooldown branch required a real session directory, file locks, and mocked services. Refactor #9 of the 2026-08-29 architecture review, based on main @ f04e0f8. Behavior-preserving.

Implementation

• evaluateGoalContinuation(state) mirrors the previous branch order and candidate-disposition semantics exactly; every skip reason keeps its value (they feed logs and CLI eligibility hints).
• A staged probe (evaluateGoalContinuationBeforeGoal) preserves the pre-existing I/O gating: undefined probe fields mean "not gathered yet", so the shell still performs no goal-file read while the workspace is busy/streaming and runs normalizeGoalLimits only after all runtime gates pass.
• Pure helpers moved with it: budget/turn-cap limit checks, budget-driven status transitions (applyBudgetDrivenStatus, now takes nowMs), wrap-up origin eligibility, and the cost micro-cent helpers.
• Per ADR-0004, interactive and CLI goal runs drive the same policy with different inputs: continuationCooldownMs (CLI: 0) and allowUserOriginBudgetWrapup (CLI: true) are plain policy-state fields; no CLI-vs-interactive conditionals in the policy. checkGoalContinuationEligibility keeps its public shape, so src/cli/goalRunDriver.ts is untouched.
• Stateful race guards stay in the shell unchanged: admissionStale generation probes, dispatch closures, pause/terminal/identity generations, chat-tail reconciliation.

Net LOC (vs f04e0f8)

• Production: +126 (+342/-216)
• Tests: -151 (+350/-501)
• Combined: -25

The production increase is the cost of giving the decision an explicit interface: the policy-state/probe/decision types (~70 lines) had no prior representation (state was implicit in service fields and closure context), and the staged-probe evaluation preserving I/O gating adds ~15 lines. All decision branches were moved, not duplicated; the service shed the same logic plus its duplicated cost helpers. On the test side, 42 table-driven pure cases (zero filesystem, zero mocks) replace 18 disk-harness permutation tests; all stateful race-invariant and shell-wiring tests were retained.

Validation

• Remote dogfood UAT (Coder Agents, real UI + real model, unmodified 60s cooldown) passed all five live goal scenarios on the pushed SHA: continuation fire + cooldown defer, turn-cap flip to budget_limited with exactly one wrap-up, re-arm on cap raise/removal, pause/manual-message suppression of pending continuations, and the user Stop acknowledgment gate.
• make static-check and whole-file runs of the five goal-related test suites (290 tests) are green.

Risks

Behavior-preserving refactor of race-sensitive logic; the main risk is decision/branch-order fidelity. Mitigated by mirroring the branch order exactly (reviewable side by side in the diff), leaving every stateful race guard in the shell untouched, retaining the race-invariant test suite, and live UAT of the continuation lifecycle.

---

Xum acted on Mike's behalf. Generated with xum • Model: anthropic:claude-fable-5 • Thinking: xhigh • Cost: $37.87

[ibetitsmike]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. More of your lovely PRs please.

Reviewed commit: 43783b0ddb

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".