fix(goal): resume blocked goal at before_agent_start to close stale-flag race

[code-yeongyu]

Defect

src/goal/lifecycle.ts set nextAgentStartWasUserTriggered = true in the before_agent_start handler and consumed it only in agent_start. In the host (agent-session.ts), before_agent_start is emitted before the final provider admission check (_enforceFinalProviderAdmission), which can reject the run — so no agent_start follows, the flag survives, and the next run that bypasses before_agent_start (e.g. a goal-continuation / queued custom-message turn that starts the agent directly via agent.prompt) consumes the stale flag and wrongly resumes a blocked goal.

Fix (option a)

Move the blocked -> active resume into the before_agent_start handler itself (it fires only for real user prompts) and remove the flag entirely. agent_start becomes pure accounting init: it reads the goal, sees status: active (if just resumed), and begins accounting for the resumed turn. A continuation-style agent_start now has no resume path of its own and cannot trigger a stale resume.

A rejected user prompt now resumes the goal at before_agent_start (the user's expressed intent to resume); no accounting begins until agent_start, preserving the invariant that the first resumed turn is accounted.

This is the cleaner of the two options: option (b)'s generation-token binding is not implementable against this host's available extension events — BeforeAgentStartEvent carries no generation token, and no extension event fires between a rejected before_agent_start and a later continuation agent_start, so a flag set in before_agent_start cannot be deterministically invalidated before the continuation consumes it. Option (a) eliminates the race mechanism entirely.

Tests (TDD)

• regression (resumes a blocked goal at before_agent_start even when admission rejects the run): a before_agent_start with no following agent_start must resume the blocked goal to active. Fails on the original flag-based code (which left it blocked waiting for agent_start), passes with this fix.
• does not resume a blocked goal on a continuation-style agent_start with no preceding before_agent_start: a continuation-style agent_start must not resume a blocked goal on its own.
• existing happy path (before_agent_start -> agent_start -> resume + first resumed turn accounted) still passes.

Evidence captured locally under .omo/evidence/pi-todo-goal-upgrade/ (task-f2a-red.txt, task-f2a-green.txt).

Gates

• npx vitest run — 87 passed (9 files)
• npm run check (tsgo + biome + no-excuse) — green (exit 0)

---

Summary by cubic

Fixes a race where a stale user-trigger flag could wrongly resume a blocked goal on continuation-only turns by moving the resume logic to before_agent_start and removing the flag. agent_start is now init-only, so rejected user runs no longer leak a resume into later turns.

• Bug Fixes
  • Resume blocked -> active in before_agent_start; removed nextAgentStartWasUserTriggered.
  • agent_start only starts accounting when the goal is already active.
  • Rejected user prompts still resume the goal; continuation-style agent_start no longer resumes.

^{Written for commit 66224ec. Summary will update on new commits.}