# Cy Loop Tasks > Task-graph execution loop for Compozy specs — self-healing checkpoint driver that ships a techspec end to end by executing its authored task graph (_tasks.md + task_NN.md), one atomic commit per Phase B task, QA, then deep-review peer-review rounds until SHIP. Use for continuous codex-loop goal runs over a .compozy/tasks/ whose work is decomposed into task files, or with --frontend to delegate frontend tasks to herdr workers. Do not use for one-off tasks, PRD/TechSpec authoring, or a slug with spec documents but no task graph (use cy-implement-spec). - Skill: `compozy/cy-loop-tasks-3` (Agent Skill, multi-file: 17 files) - Install (CLI): `npx skillmds@latest add compozy/cy-loop-tasks-3` - Raw SKILL.md: https://api.skillmd.com/api/skills/compozy/cy-loop-tasks-3/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Product & Planning - Author: compozy (https://skillmd.com/u/compozy) - Updated: 2026-09-21 - Page: https://skillmd.com/skills/compozy/cy-loop-tasks-3 --- # Loop Tasks Driver Drive a Compozy techspec to completion as a **self-healing continue** loop: each iteration detects the current phase, runs exactly one phase action, writes memory, updates `state.yaml`, and prints the iteration summary — then **continues** at detect unless the outcome is an evidence-backed external blocker or Phase E. A failed command stays inside the current phase action: diagnose, repair, and rerun it before writing iteration state. Filesystem state still resumes cleanly if the session ends mid-loop. The loop is a five-phase state machine: | Phase | Action | Executor | |-------|--------|----------| | 0 | bootstrap | orchestrator | | B | one task or slice + verify + checkpoint commit | orchestrator, or herdr frontend worker | | C | `qa_report`, then `qa_execution` | Fable 5 herdr worker, then orchestrator | | D | `deep-review` rounds until SHIP + checkpoint commit per round | orchestrator | | E | done-signature | orchestrator | Compatible with `~/dev/ai/codex-loop-plugin` goal mode; the plugin itself is never modified. Prefer in-session **continue** over waiting for a Stop→restart — restarts are a resume safety net, not the driver. ## Inputs - `` — directory name under `.compozy/tasks/`. - `` — verbatim `[[CODEX_LOOP goal="..."]]` text or the manual reason for the run. Captured once at bootstrap into `state.yaml.goal_signature`. - `--frontend ` — optional. Selects the frontend worker agent and activates the herdr frontend lane for the whole loop. Captured once at bootstrap into `state.yaml.frontend_agent`; when absent, every task runs locally. Syntax examples in `references/goal-header-template.md`. - A pre-authored `.compozy/tasks//_techspec.md`. Without it, bootstrap stops with a blocker. ## Helper scripts Bundled under `.agents/skills/cy-loop-tasks/scripts/` — stdlib-only Python 3.11+, no network, no model calls. Invoke by the explicit repo-root paths shown in the workflow steps. | Script | Role | Phase | |--------|------|-------| | `_state_io.py` | private strict state codec (imported, not invoked) | all state helpers | | `init-state.py` | bootstrap (mutating once) | 0 | | `detect-phase.py` | read-only | every iteration | | `update-state.py` | mutating | every iteration | | `commit-checkpoint.py` | mutating (git commit) | B, D | | `test_scripts.py` | read-only self-test | skill maintenance only | ## Herdr delegation lanes Two lanes dispatch work to herdr worker TUIs. Before any dispatch, read `references/herdr-delegation.md` in full and activate the `herdr-orchestration` skill it builds on. - **Frontend lane (Phase B)** — active only when `state.frontend_agent` is set. While active, every frontend task or slice is dispatched to the selected worker (`claude` → Claude Code Opus at xhigh effort, `cursor` → `cursor-agent --yolo --model grok-4.5`); the orchestrator session stays in orchestration mode and never implements frontend work itself. - **QA-report lane (Phase C)** — always active. `qa_report` is produced by a Claude Fable 5 worker (`claude --permission-mode auto --model claude-fable-5`), launched direct — never plan-first. The orchestrator runs `qa_execution` itself. ## Workflow Each **iteration** is one detect → phase action → memory → state → summary cycle. After a completed (non-blocked) summary that is not Phase E, **continue** at Step 1 in the same turn — do not end the session between rounds. If any command, gate, worker, or artifact check fails during a phase action, read `references/recovery-loop.md` in full immediately and run its repair loop. Do not write final iteration state or print the summary for an intermediate failure. **Step 1 — Detect.** 1. Print `pwd` and confirm the working directory is the repo root (the directory containing `.compozy/tasks/`). On mismatch, locate that root, change into it, and confirm again; block only when the task tree is absent. 2. Activate `cy-workflow-memory` so its protocol is loaded for later use (once per session is enough; re-activate only if context was dropped). 3. Run `python3 .agents/skills/cy-loop-tasks/scripts/detect-phase.py `. The printed line decides the whole iteration; the output catalog and entry/exit conditions are in `references/phase-transitions.md`. Done when: detect-phase emitted exactly one supported phase/action line and the matching branch below is selected. **Step 2 — Run exactly one phase branch.** ### Phase 0 — Bootstrap 1. Confirm `.compozy/tasks//_techspec.md` exists. Missing → scaffold `memory/MEMORY.md` from `references/memory-protocol.md`, record the blocker under `## Open Risks`, print the iteration summary with `outcome=blocked`, and stop (`state.yaml` does not exist yet, so there is no update-state call). 2. Run `python3 .agents/skills/cy-loop-tasks/scripts/init-state.py --goal ""`, adding `--frontend ` when the invocation text carries the parameter. Mode auto-detects: `tasks` when `_tasks.md` plus at least one `task_*.md` exist, else `free`. 3. Activate `cy-spec-preflight` for the phase the next iteration enters (`tasks`, or `task-body` when a single concrete file is next). 4. Scaffold `.compozy/tasks//memory/MEMORY.md` with the canonical sections from `references/memory-protocol.md`. 5. Run `python3 .agents/skills/cy-loop-tasks/scripts/update-state.py --phase 0 --action "bootstrap (mode=)" --outcome completed --memory-written "memory/MEMORY.md"`. Done when: `state.yaml` exists and `memory/MEMORY.md` has the canonical sections. ### Phase B mode=tasks — execute one task 1. Take the task printed by detect-phase (`task=`). Read `.compozy/tasks//.md` and confirm frontmatter `status:` is `pending` or `in_progress`. Frontmatter wins — on drift, reconcile with `update-state.py --task-completed ` for already-finished tasks or `--reconcile-tasks` for a late-authored graph, then re-run detect-phase. 2. Activate `cy-spec-preflight` in `task-body` mode for the picked file. 3. Resolve the shared and current memory paths from `references/memory-protocol.md` and pass them into the lane that executes the work. 4. **Frontend lane** — when detect-phase printed `lane=frontend agent=`: dispatch the task to that worker per `references/herdr-delegation.md`. The worker owns implementation, memory updates, scoped validation, and `cy-final-verify` evidence. It never commits. Skip step 5. 5. **Local lane** — activate `cy-execute-task` on the picked file with auto-commit disabled. Run the task's scoped validation, then `cy-final-verify`. Skip any per-task peer-review step the task file requests — that review is Phase D (see Critical Rules). 6. Confirm memory is updated (written locally, or verified from the worker) and that `cy-final-verify` evidence is PASS before any state flip. For the frontend lane, verify the worker's evidence instead of re-running verify. 7. Run `python3 .agents/skills/cy-loop-tasks/scripts/update-state.py --phase B --task-completed --action "executed " --outcome completed --memory-written "memory/.md,memory/MEMORY.md" --verify-pass`. 8. Run `python3 .agents/skills/cy-loop-tasks/scripts/commit-checkpoint.py --task `. Stdout is a commit SHA or `SKIP: no changes`; copy it into the iteration summary. On exit 1, enter the repair loop and retry the normal checkpoint after its root cause is fixed. Never bypass the hook with `--no-verify`. Done when: task frontmatter, memory, `state.yaml`, and the checkpoint result all reflect the same completed task. ### Phase B mode=free — execute one slice 1. Re-read `_techspec.md` deliverables and acceptance in full; compare against `state.progress.checklist[]`. 2. Pick the smallest coherent slice (≤ ~4 hours) that advances at least one acceptance criterion; capture its text exactly. 3. Run `python3 .agents/skills/cy-loop-tasks/scripts/update-state.py --add-progress "" --action "slice picked" --outcome completed`. 4. Re-read `state.yaml`; the current memory file is `memory/free-iter-.md`, `` = the new checklist entry's `iteration`, zero-padded to three digits. 5. **Frontend lane** — when `state.frontend_agent` is set AND the slice's owned paths are exclusively frontend surfaces (classification in `references/herdr-delegation.md`): dispatch per that reference. The worker owns implementation, memory updates, scoped validation, and `cy-final-verify` evidence; it never commits. Skip step 6. 6. **Local lane** — implement the slice, record decisions and learnings in the current memory file, run scoped validation, then `cy-final-verify`. 7. Confirm memory is updated and `cy-final-verify` evidence is PASS. For the frontend lane, verify the worker's evidence instead of re-running verify. 8. Acceptance self-check: when every techspec criterion has a completed checklist entry, add `--deliverables-complete` to the step 9 call. 9. Run `python3 .agents/skills/cy-loop-tasks/scripts/update-state.py --phase B --complete-progress "" [--deliverables-complete] --action "slice " --outcome completed --memory-written "memory/free-iter-.md,memory/MEMORY.md" --verify-pass`. 10. Run `python3 .agents/skills/cy-loop-tasks/scripts/commit-checkpoint.py --slice ""` with the exact step 3 text — same SKIP / exit-1 semantics as mode=tasks step 8. Done when: the slice's checklist entry is `completed` and the checkpoint result is recorded. ### Phase C — QA Run only the printed action. `qa_report` — dispatched, never authored locally: 1. When release-grade runtime scope needs a lab and no active `bootstrap-manifest.json` exists, activate the project's QA bootstrap skill first (e.g. `agh-qa-bootstrap` in AGH) when installed. 2. Dispatch the Fable 5 worker per `references/herdr-delegation.md` (QA-report lane). The worker activates `qa-report` with `qa-docs-path=docs/qa` and updates journey flows, `docs/qa/scenarios/` files, and cycle charters. 3. Verify the worker evidence (each reported artifact exists, no worker commit), then run `python3 .agents/skills/cy-loop-tasks/scripts/update-state.py --phase C --qa-report-done --action "qa-report produced" --outcome completed --memory-written "memory/qa-report.md,memory/MEMORY.md"`. `qa_execution` — local: 1. Activate `qa-execution` with `qa-docs-path=docs/qa`; it writes the dated run report at `docs/qa/reports/-.md` and updates scenario-file verdicts. 2. When the report is "not ready" or a Blocks-Completion/Data-Loss bug is open, keep the Phase C action open: repair every in-scope bug, rerun the affected QA, and repeat `qa-execution` through the recovery loop. Do not set `--qa-execution-done` on an intermediate report. 3. Once the report is ready, run `python3 .agents/skills/cy-loop-tasks/scripts/update-state.py --phase C --qa-execution-done --action "qa-execution produced" --outcome completed --memory-written "memory/qa-execution.md,memory/MEMORY.md" --verify-pass`. mode=tasks addition: when the printed QA action corresponds to the pending QA task file, flip that task's frontmatter `status:` to `completed` and add `--task-completed ` to the same update-state call so `tasks.pending` drains. Done when: the printed QA artifact exists on disk and its flag is recorded in `state.yaml`. ### Phase D — peer-review rounds until SHIP One round per iteration; detect-phase re-emits `peer_review` until the verdict is SHIP on a verify-PASS tree. Enter this phase only after every Phase B task or slice is complete and both QA flags are true. 1. Activate `deep-review` for the round number printed by detect-phase, scoped to the loop's full diff: `--base` = the ref the loop started from when known (default `main`), `--spec .compozy/tasks/` (contract conformance), `--subagent codex` (cross-LLM reviewer lane — the implementing model never solely reviews its own work). Later rounds ride deep-review's incremental state; never pass `--full` mid-loop. 2. The loop is the deciding authority over the round: remediate **every confirmed finding and every nitpick** from the round's review.md in this same iteration, then re-run the project verification gate. The round's verdict is the SHIP/FIX_BEFORE_SHIP/REWORK value in review.md/state.json. 3. Update `memory/peer-review.md` (a `## Round ` section per round), then run `python3 .agents/skills/cy-loop-tasks/scripts/update-state.py --phase D --review-round-done --action "peer-review round ()" --outcome completed --memory-written "memory/peer-review.md,memory/MEMORY.md" --verify-pass`. The call uses `--verify-pass`: a failed post-remediation gate stays inside the repair loop, and a SHIP verdict on a failing tree is void. 4. Run `python3 .agents/skills/cy-loop-tasks/scripts/commit-checkpoint.py --review-round ` — same SKIP / exit-1 semantics as Phase B. Done when: the round's review.md exists with a verdict, every confirmed finding and nitpick from it is remediated (or the verdict was SHIP), and `state.yaml` records the round. ### Phase E — done 1. Run a final `cy-final-verify` and confirm `state.verify.last_status=PASS`. A regression enters the repair loop and Phase E remains open until the fresh gate passes; skip the done-signature while repairing. 2. Walk the Phase E section of `references/checklist.md`; every box must pass. 3. Print the iteration summary block from `assets/iteration-summary.template.md` with `phase_out=E` and checkpoint field `n/a (phase != B/D)`. 4. Print the literal contents of `assets/done-signature.txt` on its own line — the codex-loop goal-check confirmation scans for it. 5. Stop — Phase E is the only successful terminal. Done when: the Phase E checklist passes, the iteration summary is printed, and the done-signature is the final output line. **Step 3 — Self-audit, summarize, then continue.** 1. Walk `references/checklist.md` for the phase just executed; every box must pass before summarizing. 2. Print the iteration summary block from `assets/iteration-summary.template.md` (Phase E already printed it and adds only the done-signature line). 3. **Continue gate:** stop only when `phase_out=E` or the external-blocker criteria in `references/recovery-loop.md` are proven and `outcome=blocked`. Otherwise re-enter Step 1 immediately — the summary marks the round; it does not end the session. Done when: the phase checklist passes, its summary is printed, and control either returned to Step 1 or stopped at a permitted terminal. ## Memory protocol Memory goes through the `cy-workflow-memory` skill — the exact paths per phase are in `references/memory-protocol.md`. Update memory **before** flipping any tracking field. ## Goal-mode integration The canonical `[[CODEX_LOOP ...]]` header, the manual invocation text, and `--frontend` syntax live in `references/goal-header-template.md`. ## Critical Rules - One phase action per iteration; repair failures inside that action, then **continue** at detect until Phase E or a proven external blocker — never idle between rounds waiting for a restart or re-invocation. - `state.yaml` mutates only through `init-state.py` and `update-state.py`; hand-edits void resume guarantees. There is no top-level `current_phase` — `detect-phase.py` derives it from durable state and filesystem truth every run. - Frontmatter `status:` on `task_NN.md` is the source of truth; reconcile `state.yaml` when they disagree. - Memory updates precede status flips. Always. - Frontend lane: `state.frontend_agent` set → herdr dispatch is the only way frontend work gets implemented; null → every task runs locally. Workers are interactive TUIs launched via `rtk herdr agent start` — a pane streaming raw JSON is a broken headless delegation: interrupt and relaunch per `references/herdr-delegation.md`. - `qa_report` is always produced by the Fable 5 worker; `qa_execution` always runs locally. - Every Phase B task or slice runs scoped validation then `cy-final-verify` before its checkpoint commit. A FAIL opens the repair loop; only the final PASS closes the phase action. - Peer review (`deep-review`) runs only in Phase D. Per-task peer-review instructions inside task files or specs are superseded by this loop's phase machine — note "deferred to Phase D" in the task memory and move on. - Phase D repeats in consecutive rounds until SHIP; every non-SHIP round remediates all blockers and nits before the next round starts. - Checkpoint commits (Phases B and D) belong to the orchestrator: `cy-execute-task` runs with auto-commit disabled, and every worker packet forbids committing. The checkpoint captures code, memory, task frontmatter, the master tasks file, and the advanced `state.yaml` in one atomic, restorable snapshot. - Phase E requires `qa.report_done=true`, `qa.execution_done=true`, `review.ship=true`, and `verify.last_status=PASS`. - Do not regenerate the loop's input graph with `cy-create-tasks`, `cy-create-techspec`, `cy-tasks-tail-qa-pair`, or `cy-web-docs-impact`. The only exception is a repository-mandated two-touch corrective TechSpec: activate its required spec skills, let the loop decide choices already bounded by the current goal/contract, persist the corrective design, and continue without replacing the original task graph. ## Error Handling - **Any failure** — read `references/recovery-loop.md` in full and execute it before mutating iteration state. Failed commands are repair work, not blockers. Use `outcome=blocked` only after its external-blocker test passes. - **`_techspec.md` missing at bootstrap** — record the blocker in `memory/MEMORY.md` `## Open Risks`, print the iteration summary with `outcome=blocked`, stop. No update-state call: `state.yaml` does not exist yet. - **Mode disagreement** — `init-state.py` exits 4 when `--mode` contradicts the filesystem. Reconcile by adding/removing `_tasks.md` before bootstrap, or run `update-state.py --reconcile-tasks` when the task graph was authored after a free-mode bootstrap. - **`state.yaml` parse failure** — `detect-phase.py` exits 1 with the parse error on stderr. Diagnose the malformed writer or interrupted write from evidence and repair it without discarding unrelated worktree changes. - **`commit-checkpoint.py` exit 1** — repair the hook or commit failure and retry normally. If the repair changes tracked source after the last PASS, rerun `cy-final-verify` before retrying. `SKIP: no changes` is success. - **Worker launch or delegation failure** — the pane shows raw JSON instead of a TUI banner, `rtk herdr agent list` stays `unknown`, or the status wait times out with no progress: interrupt (`rtk herdr pane send-keys ctrl+c`) and relaunch once via `rtk herdr agent start`; if it fails again, diagnose and repair the worker environment through the recovery loop. - **Delegated run lacks PASS evidence or artifacts, or committed anyway** — keep the phase open, recover the missing evidence or rerun the lane, and do not advance. A worker commit is a contract breach that requires preserving the worker's work and repairing checkpoint ownership before continuing. - **Invalid peer-review round** (missing or malformed review artifacts, or no verdict) — the round does not count; follow `deep-review` error handling and re-run it. - **Two-touch rule** — on the third corrective touch, replace patching with the structural redesign required by the repository, validate it, and continue. It becomes a blocker only when that redesign needs an external product decision or authority unavailable to the loop. - **External blocker proven** — record the evidence and exhausted alternatives in memory, call `update-state.py` with `--verify-fail --blocker --outcome blocked`, print the summary, and stop without the done-signature.