sprint-loops

name	sprint-loops
description	Structured five-phase workflow (Research → Plan → Build → Test → Loop) for long-horizon coding tasks. Trigger when the user mentions sprint loops, starting a sprint, continuing a sprint, or when the project root contains a `sprints/` directory.

Sprint Loops

Five-phase workflow for long-horizon agentic coding work.

Routing

1. Run bash scripts/current-phase.sh from the project root.
2. Read the matching phases/ file:
   • uninitialized or ready-for-next-sprint → phases/01-init-sprint.md
   • research → phases/02-research-phase.md
   • plan → phases/03-plan-phase.md (also invoke /plan slash command)
   • build → phases/04-build-phase.md
   • test → phases/05-test-phase.md
   • loop → phases/06-loop-phase.md
3. Execute. When the phase exit condition is met, re-route.

Plan mode integration

When entering the Plan Phase, suggest the user run /plan for native Codex plan mode. The plan-phase file's job is to produce build-plan.md and test-plan.md — these are the planning artifacts that survive across turns.

Approval mode guidance

• Research, Plan, Loop phases work fine with --ask-for-approval on-request (default).
• Build Phase benefits from workspace-write sandbox + --ask-for-approval never for batch task execution, since every task ends in a git commit that provides rollback.
• Test Phase can run with workspace-write + on-request for safety.

Never use --yolo or --dangerously-bypass-approvals-and-sandbox for sprint loops — the whole point of the system is auditable, recoverable iteration.

State on disk

The filesystem IS the state machine. Persistent state lives in sprints/ and agent-tasks/. Trust the disk; do not re-derive state from chat history.

Subagent opportunity

When the Build Phase has multiple independent tasks (no shared dependencies in the build-plan's execution sequence), consider spawning subagents for parallelization. Each subagent handles one task and commits its own diff. The parent agent merges results and proceeds.

Adversarial critics for Plan and Test phases: Plan Phase (after writing both plans) and Test Phase (after writing test artifacts) spawn a critic subagent with the matching prompt from prompts/:

• prompts/plan-critic.md — screens the build/test plans for vague EARS clauses, plan-test mismatch, missing risk coverage, hidden deps, ignored ADRs, granularity violations, E2E drift.
• prompts/test-critic.md — screens the test artifacts for EARS-clause coverage gaps, weak assertions, stub leakage, integration scope drift, E2E cop-out, missing negative-paths, flake risk.

The critic returns a structured critique (## Concerns + ## Confidence); the primary agent records it in critique.md and addresses each concern before locking. See phases/03-plan-phase.md and phases/05-test-phase.md for the full protocol.

Autonomous operation

When invoked for a multi-turn loop (e.g. via codex exec or when the user signals they're stepping away), default to working independently for the entire sprint:

• Commit, push, and merge your own PRs without asking per step. The per-task commit boundary already provides rollback; the sprint structure already provides review surfaces (research-report, plans, test-report, decisions ADR). Don't pause for confirmation on routine work.
• Defer rather than block. When a feature has a non-trivial dependency the plan didn't anticipate, ship the scoped piece, note the deferral in sprint-meta.md blockages or the PR body, and continue with the next executable task. Use scripts/abort-sprint.sh only for truly unrecoverable blockages.
• Use /plan for the Plan Phase, then drop back to standard mode for Build/Test/Loop. Plan Mode produces build-plan.md and test-plan.md without touching source files.
• One PR per concept, numbered sequentially (e.g. v117, v118) when the sprint is wrapped as a PR.

Safety floor

Autonomy stops at the safety floor:

• Don't weaken permission or security controls to keep the loop moving. If a tool call is denied, surface the decline clearly and continue with what is permitted — never invoke --yolo or --dangerously-bypass-approvals-and-sandbox.
• Don't skip pre-flight checks documented in phases/04-build-phase.md even when running unattended. The project's sanity gate blocks the per-task commit — that's the design.
• Never --no-verify or bypass hooks unless the user explicitly opted in. Hook failures are signal; investigate, don't suppress.
• Hard-to-reverse actions warrant a pause even in autonomous mode: force-push to a base branch, dropping a DB table, deleting infra. Surface the intent and wait for confirmation.