DONE [x] 81 — <title> DONE [~] 82 — <title> IN PROGRESS [ ] 83 — <title> READY ... ``` 5. If every task is DONE, set `Status: done` in `$AY/goal.md` and say the goal is complete. Stop. --- ## `/goal <objective>` — set up and run ### Step 1 — Record the goal Write `$AY/goal.md` (overwrite any prior goal only after confirming with the human if one is still `active`): ```markdown # Goal — <objective> - **Objective:** <objective, verbatim> - **Status:** active - **Started:** <ISO-8601 UTC> - **Updated:** <ISO-8601 UTC> ## Tasks  ``` Get timestamps with `date -u +%Y-%m-%dT%H:%M:%SZ`. ### Step 2 — Decompose into tasks Break the objective into **3–8 concrete, independently-shippable tasks**. Each task must be atomic (one clear deliverable, ideally < 5 files) and ordered so earlier tasks unblock later ones. Avoid vague tasks ("improve X") — every task needs a verifiable done-state. 1. Determine the next free numeric task IDs (never reuse an existing one): ```bash last=$(grep -oE '^\|\s*[0-9]+' "$AY/tracking/BOARD.md" | grep -oE '[0-9]+' \ | sort -n | tail -1) next=$((last + 1)); echo "first new id: $next" ``` 2. For each task, create a task file `$AY/tasks/{id}-{slug}.md` (use the task template if present at `$AY/tasks/_template.md`) with What / Why / Deliverables / Acceptance sections. 3. Append each task to `$AY/tracking/BOARD.md` as a **READY** row under a `## Goal: <objective>` section: ``` | {id} | {title} | goal-runner | READY | {blocked-by or -} | ``` 4. Record the task IDs back into the `## Tasks` block of `$AY/goal.md` as `- [ ] {id} — {title}` lines. Present the decomposition to the human as a numbered list before looping. This is the one lightweight gate — if they say adjust, revise the tasks; otherwise proceed immediately (do not wait for a second confirmation). ### Step 3 — Worker loop Now work the tasks back-to-back. Repeat this cycle until **all goal tasks are DONE** or the session must end: 1. **Pick** the next task from the goal whose BOARD status is READY and whose dependencies (Blocked By) are all DONE. If none are READY but some are BLOCKED, re-check whether a just-completed task unblocks them; if still none, stop and report. 2. **Claim** it: create `$AY/tracking/locks/task-{id}.lock` (agent `goal-runner`, ISO timestamp, session id) and set the BOARD row to IN PROGRESS. If a lock already exists, skip to the next task. 3. **Work** it. For non-trivial tasks, run the standard `/go` BUILD discipline: follow a short plan, verify as you go, and self-review. Honor the **Fix-Loop Cap** — after 3 consecutive failed checkpoints on one task, stop that task, write a BLOCKER, and move to the next task rather than looping. 4. **Finish**: verify the deliverable actually works (read files back, run build/tests), then set the BOARD row to DONE, delete the lock, tick the task in `$AY/goal.md` (`- [x] {id} — {title}`), update `Updated:`, and append a one-line entry to `$AY/tracking/AGENTS-LOG.md`. Write a HANDOFF entry if there are gotchas for later tasks. 5. **Continue** to the next task automatically — do not ask the human between tasks. ### Step 4 — Stop conditions Stop the loop and report when any of these is true: - **All goal tasks are DONE** → set `Status: done` in `$AY/goal.md`, print a summary (tasks completed, files touched), and stop. - **Context pressure ≥ 85%** (see the Context-Pressure check in `/go`) → `/pause` to snapshot state to `.ay/sessions/`, leave `Status: active`, and tell the human how to resume (`/goal` re-reads `goal.md` and continues the remaining READY tasks). - **A task escalates** (fix-loop cap hit / genuine blocker) → record the BLOCKER, keep working other READY tasks, and surface the blocker in the final report. - **No READY tasks remain but some are BLOCKED/DONE-incomplete** → report what is blocking and stop. ### Resuming Running `/goal` again with no new objective (or `/goal <same objective>`) while a goal is `active` resumes the loop: re-read `$AY/goal.md`, skip DONE tasks, and pick up the next READY one. Only start a fresh decomposition when the objective is new or the prior goal is `done`.

--- name: ayf:goal description: | Sustained all-day autonomous work: decompose one big objective into tasks and work them back-to-back until done. Usage: /goal | /goal status. allowed-tools: - Read - Write - Edit - Bash - Glob - Grep - Agent - AskUserQuestion --- # /goal -- Sustained Objective Runner `/goal` turns a single large objective into a self-driving queue of tasks and works them one after another, unattended, until they are all DONE or the session ends. Use it when you want Claude to make progress on a big goal all day without babysitting each task. For *recurring* checks on a fixed interval, use `/loop` instead (see **Sustained Work** in CLAUDE.md). Two forms: - `/goal ` -- set the objective, decompose it, and start the worker loop. - `/goal status` -- report progress on the active goal (X of Y tasks done). Resolve `.ay/` first (works from any worktree): ```bash AY="$(dirname "$(git rev-parse --git-common-dir 2>/dev/null)")/.ay" ``` --- ## `/goal status` — report progress If the argument is exactly `status`, do NOT start any work. Instead: 1. If `$AY/goal.md` does not exist, say "No active goal. Start one with `/goal `." and stop. 2. Read `$AY/goal.md`. Extract the objective, status, and the list of task IDs. 3. For each task ID, look up its current status in `$AY/tracking/BOARD.md`. 4. Print a short report: ``` Goal: (status: active) Progress: 3 of 6 tasks done [x] 80 — DONE [x] 81 — <title> DONE [~] 82 — <title> IN PROGRESS [ ] 83 — <title> READY ... ``` 5. If every task is DONE, set `Status: done` in `$AY/goal.md` and say the goal is complete. Stop. --- ## `/goal <objective>` — set up and run ### Step 1 — Record the goal Write `$AY/goal.md` (overwrite any prior goal only after confirming with the human if one is still `active`): ```markdown # Goal — <objective> - **Objective:** <objective, verbatim> - **Status:** active - **Started:** <ISO-8601 UTC> - **Updated:** <ISO-8601 UTC> ## Tasks  ``` Get timestamps with `date -u +%Y-%m-%dT%H:%M:%SZ`. ### Step 2 — Decompose into tasks Break the objective into **3–8 concrete, independently-shippable tasks**. Each task must be atomic (one clear deliverable, ideally < 5 files) and ordered so earlier tasks unblock later ones. Avoid vague tasks ("improve X") — every task needs a verifiable done-state. 1. Determine the next free numeric task IDs (never reuse an existing one): ```bash last=$(grep -oE '^\|\s*[0-9]+' "$AY/tracking/BOARD.md" | grep -oE '[0-9]+' \ | sort -n | tail -1) next=$((last + 1)); echo "first new id: $next" ``` 2. For each task, create a task file `$AY/tasks/{id}-{slug}.md` (use the task template if present at `$AY/tasks/_template.md`) with What / Why / Deliverables / Acceptance sections. 3. Append each task to `$AY/tracking/BOARD.md` as a **READY** row under a `## Goal: <objective>` section: ``` | {id} | {title} | goal-runner | READY | {blocked-by or -} | ``` 4. Record the task IDs back into the `## Tasks` block of `$AY/goal.md` as `- [ ] {id} — {title}` lines. Present the decomposition to the human as a numbered list before looping. This is the one lightweight gate — if they say adjust, revise the tasks; otherwise proceed immediately (do not wait for a second confirmation). ### Step 3 — Worker loop Now work the tasks back-to-back. Repeat this cycle until **all goal tasks are DONE** or the session must end: 1. **Pick** the next task from the goal whose BOARD status is READY and whose dependencies (Blocked By) are all DONE. If none are READY but some are BLOCKED, re-check whether a just-completed task unblocks them; if still none, stop and report. 2. **Claim** it: create `$AY/tracking/locks/task-{id}.lock` (agent `goal-runner`, ISO timestamp, session id) and set the BOARD row to IN PROGRESS. If a lock already exists, skip to the next task. 3. **Work** it. For non-trivial tasks, run the standard `/go` BUILD discipline: follow a short plan, verify as you go, and self-review. Honor the **Fix-Loop Cap** — after 3 consecutive failed checkpoints on one task, stop that task, write a BLOCKER, and move to the next task rather than looping. 4. **Finish**: verify the deliverable actually works (read files back, run build/tests), then set the BOARD row to DONE, delete the lock, tick the task in `$AY/goal.md` (`- [x] {id} — {title}`), update `Updated:`, and append a one-line entry to `$AY/tracking/AGENTS-LOG.md`. Write a HANDOFF entry if there are gotchas for later tasks. 5. **Continue** to the next task automatically — do not ask the human between tasks. ### Step 4 — Stop conditions Stop the loop and report when any of these is true: - **All goal tasks are DONE** → set `Status: done` in `$AY/goal.md`, print a summary (tasks completed, files touched), and stop. - **Context pressure ≥ 85%** (see the Context-Pressure check in `/go`) → `/pause` to snapshot state to `.ay/sessions/`, leave `Status: active`, and tell the human how to resume (`/goal` re-reads `goal.md` and continues the remaining READY tasks). - **A task escalates** (fix-loop cap hit / genuine blocker) → record the BLOCKER, keep working other READY tasks, and surface the blocker in the final report. - **No READY tasks remain but some are BLOCKED/DONE-incomplete** → report what is blocking and stop. ### Resuming Running `/goal` again with no new objective (or `/goal <same objective>`) while a goal is `active` resumes the loop: re-read `$AY/goal.md`, skip DONE tasks, and pick up the next READY one. Only start a fresh decomposition when the objective is new or the prior goal is `done`.