fix(code): treat a missing plan.json at COMPLETE as a spurious completion

[mikeangstadt]

Summary

Generate Implementation Plan silently produces nothing. The run completes, exits 0, and the user is left with an implementation-plan artifact that looks done and is empty.

Reproduced 2026-08-10 (symphony-alpha PLN-1688). From the loop's own output, in order:

No plan-source.md found. Launching plan-draft-writer to create the plan.
Plan-draft-writer is running in the background. Waiting for completion.
...
Completion promise detected: COMPLETE
Loop complete after 1 iterations!
→ Post-Loop Code Review

The orchestrator launched plan-draft-writer as a background agent, said in plain text that it was waiting for it, and that same turn carried the completion promise. The loop ended, the writer was abandoned mid-flight, and the post-loop code review then passed vacuously over an empty diff — "No findings — but this is a vacuous pass... The base ref you passed equals HEAD, so nothing was examined" — and the process exited 0.

Corroborating state left in the worktree: state.json still reads {"phase":"Phase 1: Planning","status":"IN_PROGRESS"} while runs.log records plan_execute → review_approve. Pre-exploration did finish (code-map.json 17.6K, investigation-log.md 16.4K, requirements-extract.json 14.1K) — the run stopped exactly at the plan-writing step.

detect_spurious_complete had the hole backwards. Its checks validate pendingTasks inside an existing plan.json, and its first branch returned "not spurious" whenever the file was absent:

if [[ ! -f "$plan_file" ]]; then
  echo '{}'   # <- let COMPLETE through
  return
fi

So "no plan at all" — the case that actually happens — was the one case nothing could catch. A --prd run exists to produce plan.json; claiming COMPLETE without one is the strongest spurious-completion signal there is.

Change

• A missing plan.json at COMPLETE is now flagged as PLAN_MISSING_AT_COMPLETION, routed through the existing handle_spurious_complete path (telemetry, lock release, user-visible failure, exit 1) — so it bails before the post-loop code review rather than compounding misleading state, exactly as the sibling subcodes do.
• Scoped: only fires when the run was asked for a plan (--prd / PRD_FILE). A run with no PRD never owed a plan and is untouched.
• The AWAITING_USER hard stop still outranks it, so a plan drafted and parked at the Phase 1.1 review checkpoint is not turned into a failure.
• detect_spurious_complete takes the PRD path as an optional second argument defaulting to the global, so the existing call site is unchanged and tests can be explicit.
• plugins/code version bumped 1.14.7 → 1.14.8 per the Plugin Version Bump gate.

Test plan

New plugins/code/scripts/tests/test_spurious_complete.sh (6 cases). run-loop.sh guards main with [[ "${BASH_SOURCE[0]}" == "$0" ]], so the suite sources it and calls the function directly:

• COMPLETE with no plan.json on a --prd run → PLAN_MISSING_AT_COMPLETION (the reported defect)
• COMPLETE with no plan.json and no PRD → not flagged (scope guard)
• AWAITING_USER + no plan → not flagged (the documented hard stop still wins)
• a plan with no pending tasks → not flagged
• pending tasks → still PENDING_TASKS_AT_COMPLETION
• pending tasks + open questions → still PENDING_TASKS_BLOCKED_BY_QUESTIONS

passed: 6  failed: 0

Counterfactual: neutering only the new branch fails exactly the first case and leaves the other five green — so the new code is doing the work, and the four pre-existing behaviours are provably unchanged.

Known gap: CI runs ruff/pyright/pytest and the TS suite; there is no bash-test job, so this suite (like the existing plugins/code/hooks/tests/*.sh) is not gated. Run it with bash plugins/code/scripts/tests/test_spurious_complete.sh. Wiring the bash suites into CI is worth doing separately.

Related

The other half of this failure is in closedloop-ai/symphony-alpha (ISS-5872, URGENT): the desktop loop harness computes missingRequired via validateResultBundle, logs Missing required artifacts for PLAN: plan.json, then discards the result and finalizes the loop as COMPLETED with error: null. Either layer alone would have caught this; both failed open. This PR closes the upstream one.