[suprie]

Summary

The run queue no longer merges every queued phase into one LLM turn. prd-parallel-scheduling makes the phase the unit of execution: each run-plan phase runs in its own executor turn, in its own worktree and branch cut from the default branch, with its own provider session, token budget, and wall-clock watchdog. A new dependency-aware scheduler validates the plan graph, holds dependents in Waiting until every upstream lands green, then fans out ready phases concurrently — four by default, configurable per plan — while a non-green upstream fails only its dependents and leaves independent in-flight phases running. Restart reconciliation now handles multiple in-flight phases, auto-resuming only when the worktree and saved provider session both survive and parking the rest with a reason for the morning report.

What changed

• Add scheduler.rs: graph validation (cycles, self-edges, dangling edges), Waiting gate release/fail, ready-set selection, per-phase watchdog, restart classification, and generated-graph property tests.
• Rewrite the run executor: one PhaseTask and one turn per phase, a JoinSet bounded by RunPlan.max_concurrency, per-phase worktrees, per-phase token/wall-clock budgets, and serialized plan bookkeeping.
• Add RunPhaseStatus::Waiting, RunPlan.max_concurrency (default 4), and per-phase worktree/session fields to run-plan.yaml, with legacy plans loading unchanged.
• Extend restart reconciliation to auto-resume recoverable in-flight phases via a resume-aware session spawn.
• Seed upstream handoff artifacts into each dependent phase's worktree so dependents never read an unmerged sibling branch.
• Expose a concurrent-phases control in the plan-tonight wizard and surface waiting in the run surfaces.
• Add the six loop artifacts under .loopdeck/handoff-artifacts/.

PRD

docs/epics/role-based-orchestration/prd-parallel-scheduling.md

Decisions

• 2026-10-30 — Execute run-plan phases one turn each with a ready-set scheduler (accepted): each phase runs in its own worktree and turn under max_concurrency, with a Waiting gate and fail-dependents policy.
• 2026-10-10 — Use write-capable Codex policy for unattended runs (accepted): autonomous turns use workspaceWrite with on-request, subject to the destructive-command floor.
• 2026-10-09 — Correlate Codex turn failures across the run watchdog (accepted): lifecycle logs now distinguish tool hangs, protocol failures, watchdogs, and cleanup failures.

Test plan

• cargo test (from src-tauri/)
• cargo clippy -- -D warnings (from src-tauri/)
• npx tsc --noEmit
• npm run test:frontend
• Manual: queue a two-phase plan with disjoint dependencies and confirm two phase worktrees run at once, each opening its own draft PR.

Verify Verdict

PRD Verification — prd-parallel-scheduling.md

Verdict: PASS

#	Criterion	Status	Evidence
1	One executor turn per run-plan phase, replacing the combined batch (P0)	PASS	src-tauri/src/commands/run_queue.rs:2218 (execute_run) spawns one PhaseTask per ready phase; :1803 builds the prompt from a single-loop slice; next_queued_batch and the combined-batch turn are deleted
2	Concurrent execution of phases whose depends_on are satisfied and disjoint, each in its own worktree, under a bounded concurrency limit (P0)	PASS	src-tauri/src/scheduler.rs:304 (select_ready); run_queue.rs:2239 (JoinSet), :2293 (max_concurrency), :202 (each branch cut from the default branch); default 4 in limits.rs:114; picker PlanTonightWizard.tsx:659
3	Gated starts: dependent phases hold in Waiting until upstream completes, then release or fail per the stall policy (P0)	PASS	runplan.rs:54 adds Waiting; scheduler.rs:153 (mark_waiting), :186 (reconcile_gates), :272 (fail_dependents leaves Running phases alone); gate edges in transition.rs
4	Restart reconciliation covering multiple in-flight phases (P1)	PASS	run_executor.rs:582 iterates every Running phase; scheduler.rs:354 classifies resume vs park on worktree+session survival; the resumed turn passes the saved id through agent.rs:1947
5	Per-phase token and wall-clock budgets per in-flight phase, not per batch (P1)	PASS	run_queue.rs:1705 creates one TokenBudget per PhaseTask; scheduler.rs:333 (phase_watchdog); run_queue.rs:1991 kills only that phase and fails its dependents
6	Add scheduler property tests (diamond graphs, cycle rejection, mid-flight crash recovery) and worktree cleanup race checks (Phase 4 loop)	PASS	scheduler.rs:721/:773/:799 generated-DAG, cycle, and restart properties; run_executor.rs:1006/:1520 crash recovery; run_queue.rs:3216 disjoint-worktree cleanup check
7	Must reject cycles at plan creation (Design)	PASS	run_queue.rs:636 validates before persisting; scheduler.rs:69 rejects cycles, self-edges, and dangling edges; also checked at run start (run_queue.rs:2232)

Non-goals audit

• No scope creep detected. Cron/delays, cross-project scheduling, the one-run-per-project lock, and branch auto-merge are untouched.

Roll-up: PASS — 0 FAIL, 0 PARTIAL, 7 PASS.

Run metadata

• Phases:
  • parallel-scheduling/one-turn-per-phase — "Replace the combined queued-batch turn with one executor turn per run-plan phase" (role-based-orchestration / prd-parallel-scheduling / Phase 1 — Per-phase turn execution)
  • parallel-scheduling/parallel-batches — "Run phases with satisfied, disjoint depends_on concurrently in separate worktrees under a bounded concurrency limit" (role-based-orchestration / prd-parallel-scheduling / Phase 2 — Dependency-aware fan-out)
  • parallel-scheduling/gated-start — "Hold dependent phases in Waiting until upstream phases complete, releasing or failing per the stall policy" (role-based-orchestration / prd-parallel-scheduling / Phase 2 — Dependency-aware fan-out)
  • parallel-scheduling/restart-reconcile — "Extend restart reconciliation to multiple in-flight phases after a crash or restart" (role-based-orchestration / prd-parallel-scheduling / Phase 3 — Reconciliation + budgets)
  • parallel-scheduling/budget-split — "Apply per-phase token and wall-clock budgets per in-flight phase instead of per combined batch" (role-based-orchestration / prd-parallel-scheduling / Phase 3 — Reconciliation + budgets)
  • parallel-scheduling/tests — "Add scheduler property tests (diamond graphs, cycle rejection, mid-flight crash recovery) and worktree cleanup race checks" (role-based-orchestration / prd-parallel-scheduling / Phase 4 — Tests/verification)
• Budgets: 3000000 tokens total · 32400s wall-clock cap for this turn · 28800s total-run wall-clock cap