| name | autopilot |
| description | PR orchestration planner. Reads a task file, discusses PR splitting strategy
with the user, produces plan + spec files, and either dispatches a worktree
subagent in the current session (single PR, run now) or schedules execution
via cron (multi-PR, or single-PR scheduled). Use when the user wants a task
implemented, reviewed, and merged automatically — whether the scope fits in
one PR or needs to be split. Triggers: "autopilot", "orchestrate", "batch
implement", "split into PRs and run", or when the user provides a todo/spec
and wants unattended PR execution.
|
Autopilot
Planning phase for automated PR orchestration. Reads a task file, discusses splitting
strategy with the user, produces plan + spec files, and then either schedules execution
or dispatches inline (single-PR, run now).
This skill handles planning plus the single-PR + run-now dispatch. The
scheduled path (single-PR or multi-PR) hands off to the Orchestrator prompt
triggered via cron (see references/orchestrator.md).
Flow Overview
User provides todo file
↓
Step 0: Verify dependency skills are available
Step 1: Understand the task
Step 2: Plan PR(s) — natural split is 1 PR (single-PR branch) or N PRs (multi-PR branch)
Step 3: Write artifacts (varies by mode — see Step 3 table)
Step 4: User confirmation (single-PR adds a run-now / schedule choice)
Step 5: Execute
├─ Single-PR + run now → inline-dispatch a worktree subagent in this session
├─ Single-PR + schedule → CronCreate → orchestrator (loop with N=1)
└─ Multi-PR → CronCreate → orchestrator
Step 0: Verify Dependency Skills
Before proceeding, confirm that the execution-phase skills are available.
Co-packaged in solopreneur (always present alongside this skill —
no runtime check needed):
/greenlight — Automated PR review loop
/tech-vetting — Pre-implementation technical vetting (called by PR
subagent after planning)
External dependency (must be verified at runtime):
/merge-pr — PR merge workflow
Verification: check if /merge-pr appears in the current available skills
list (system-reminder). If it's missing, stop and tell the user what needs
to be installed.
Step 1: Understand the Task
Input: File path provided by the user (typically in todos/backlog/ or todos/doing/)
- Read the todo file
- Read related codebase context (architecture docs, specs, existing code)
- Summarize in one paragraph: "What this task does, which modules are affected,
what the expected output is"
- Confirm understanding is correct before proceeding to Step 2
If the todo lacks specificity (missing technical details, no clear acceptance criteria),
work with the user to fill in gaps before continuing.
Step 2: Plan PR(s)
Based on the task content, propose a PR splitting strategy. The same principles
apply whether the result is 1 PR or N PRs:
-
Splitting principles:
- Each PR should touch non-overlapping files where possible (enables parallelism)
- Import dependencies require sequential ordering
- Tests go in the same PR as their implementation
- Documentation updates go in the final PR
-
The resulting PR count branches the rest of the flow:
- Natural split is 1 PR → single-PR sub-flow (descriptor below; no
dependency graph, no parallel-safety check, no batch loop)
- Natural split is ≥2 PRs → multi-PR sub-flow (full dependency graph
- parallel-safety validation)
The user can flip between branches at any time. If the natural split is 1 PR
but the user pushes back ("split it into 2"), restart Step 2 in multi-PR mode.
The reverse is also true.
Single-PR descriptor (split = 1 PR)
Output a single PR descriptor instead of a graph:
PR
title: feat(scope): <summary>
branch: feature/<short-name>
files:
- path/a.py (new)
- path/b.py (modified)
subagent: ai-engineer
type: code # code | docs
size: m # optional — S/M/L review hint for /greenlight (upward-only); omit to auto-classify
Reason: scope is contained to a single module / single purpose, no split needed.
subagent accepts the same values as multi-PR (see the multi-PR list below:
ios-dev / android-dev / ai-engineer / neo4j-dev / marketer / designer).
What is dropped in single-PR mode: dependency graph rendering, parallel
safety check, batch loop planning. Proceed to Step 3.
Multi-PR descriptor (split ≥ 2 PRs)
For each PR, list:
- Short title
- Files to create/modify (paths)
- Dependencies on other PRs
- Required subagent type (ios-dev / android-dev / ai-engineer / neo4j-dev / marketer / designer)
- Optional review size (
s / m / l) — the S/M/L hint passed to /greenlight; omit to let it auto-classify from the diff
Validate the dependency graph:
- For PRs declared parallel: confirm file paths don't overlap
- For dependent PRs: confirm the dependency actually creates the required modules
- Check for circular dependencies
Present the dependency graph for user confirmation:
PR1 (models) ──→ PR3 (router, depends on PR1)
PR2 (worker) ──→ PR4 (docs, depends on all)
↑
Can run parallel with PR1
Proceed to Step 3 after user confirms.
Step 3: Write Artifacts
When to write — read this first:
- Multi-PR: write artifacts now, then proceed to Step 4.
- Single-PR (either run now or schedule): do not write any artifacts in
this step. Defer all writes until after the user confirms execution mode in
Step 4. This avoids leaving half-written
plan.yaml / state.json / spec
files on disk if the user cancels at the confirmation gate.
Output location is docs/loops/<YYYY-MM-DD>_<short-name>/ — same convention regardless
of mode. What gets written depends on mode:
| Mode | Files written |
|---|
| Single-PR + run now | pr1-<short>.md only |
| Single-PR + schedule | plan.yaml (1 PR entry) + state.json + pr1-<short>.md |
| Multi-PR | plan.yaml + state.json + pr1-<short>.md, pr2-<short>.md, … |
Spec naming: always pr1-<short>.md, pr2-<short>.md, … — the same convention
regardless of mode. Single-PR uses pr1-<short>.md (not spec.md) so that
references/schemas.md needs no exception, and a mid-flow pivot ("actually let's
add a PR2") doesn't require renaming the existing file.
Example tree (multi-PR):
docs/loops/2026-03-29_mining-queries/
├── plan.yaml
├── state.json
├── pr1-models.md
├── pr2-worker.md
└── pr3-router.md
plan.yaml
Read references/schemas.md for the full plan.yaml schema. In single-PR + schedule
mode, prs: legitimately has one entry — no special handling needed. Key fields:
name: "Short name"
source_todo: "todos/doing/xxx.md"
prs:
- id: pr1
branch: feature/xxx-pr1
title: "feat(scope): description"
type: code
size: m
subagent: ai-engineer
depends_on: []
spec: pr1-models.md
files:
- path/to/new_file.py
- path/to/existing_file.py
Spec Files
One .md file per PR. Format:
# PR Title
## Requirements
- What to do (functional description, not pseudo code)
- Constraints (e.g., no dependency on FooService, all methods sync)
## Files to Read
- path/to/reference1.py (understand existing structure)
- path/to/reference2.py (reference this pattern)
## Files to Create/Modify
- path/to/new_file.py — description
- path/to/existing_file.py — what to modify
## Acceptance Criteria
- [ ] Test command: `cd xxx && uv run pytest tests/test_xxx.py -v`
- [ ] Specific verifiable condition 1
- [ ] Specific verifiable condition 2
## Notes
- Technical decision reminders (if any)
- Known pitfalls (if any)
Specs describe what to do + how to verify completion, not pseudo code.
Let the implementation subagent decide how to write the code.
Spec quality gate (acceptance criteria)
Step 3 is where a fuzzy goal gets compiled into a loop contract, so gate the
acceptance criteria before writing the spec. This is a lightweight checklist
prompt — no subagent; the real enforcement is greenlight's verifier loop (see
../greenlight/SKILL.md "Inner verify loop"). Run it inline:
- Every acceptance criterion must be an executable command or a verifiable
assertion. An executable command is something a shell can run and pass/fail on
(
cd x && uv run pytest tests/test_x.py, grep -n "reason_class" file.md); a
verifiable assertion names an observable state or behavior ("stops within 60s of the
shutdown event"). Reject vague criteria — "works correctly", "handles errors
properly", "the UI looks right" have no verifier — and rewrite them into a command or
an observable condition before proceeding.
type: docs PRs are exempt (prose criteria are checklist assertions a reviewer
judges, not runnable commands), BUT cross-check the exemption against the mechanical
S whitelist from greenlight's sizing cascade
(../greenlight/SKILL.md → "Mechanical cascade"): the pure-prose whitelist is
docs/** (excluding docs/loops/**), todos/**, the repo-root README.md,
LICENSE, .gitignore. If any path in the PR's files list falls outside that
whitelist, the self-declared type: docs is not trustworthy — override it to
code (or at minimum flag it) and apply the gate in full. A SKILL.md edit is not
docs. This closes the gaming path where a mislabeled docs type buys both this
exemption and the sizing PR's S light-review.
state.json
Initial state (multi-PR or single-PR + schedule; not written for single-PR
- run now, since there is no orchestrator loop to resume):
{
"status": "pending",
"plan_dir": "docs/loops/2026-03-29_mining-queries",
"prs": {
"pr1": { "number": null, "status": "pending", "worktree": null },
"pr2": { "number": null, "status": "pending", "worktree": null }
}
}
Step 4: User Confirmation
Multi-PR
Present the final plan + Tech Vetting results for user confirmation:
Plan: 4 PRs
PR1 (models) + PR2 (worker) → parallel
PR3 (router) → waits for PR1
PR4 (docs) → waits for all
Tech Vetting: all passed ✅
Schedule for automatic execution? Tell me when you'd like it to run.
Single-PR
Present the plan summary, then directly ask「要現在跑嗎?」:
Plan: 1 PR
feat(scope): <summary> (subagent: ai-engineer, branch: feature/<short-name>)
Files: path/a.py, path/b.py
Spec target: docs/loops/<YYYY-MM-DD>_<short-name>/pr1-<short>.md
要現在跑嗎?(約 10-30 分鐘,session 會被佔用;不跑的話可以排程)
After the user picks, write the artifacts per Step 3's table, then proceed to Step 5.
Step 5: Execute
| Mode | Path |
|---|
| Single-PR + run now | New inline-dispatch path (below) |
| Single-PR + schedule | Existing CronCreate → orchestrator path. Orchestrator reads plan.yaml with N=1 and degenerates to a single iteration of the execution loop. |
| Multi-PR | Existing CronCreate → orchestrator path. |
Single-PR + run now (inline dispatch)
-
Read references/pr-subagent-template.md (template is unchanged).
-
Read the spec just written (pr1-<short>.md).
-
Resolve dispatch-time template variables:
{BRANCH} = feature/<short-name>
{TITLE} = feat(scope): <summary>
{PR_ID} = pr1
{PLAN_DIR} = docs/loops/<YYYY-MM-DD>_<short-name>
{SPEC_FILE} = pr1-<short>.md
{SIZE} = the PR's size (s/m/l) when the descriptor set one — Step 5
of the suffix passes it to /greenlight as size={SIZE}. When unset, drop the
size={SIZE} clause so no token is passed (greenlight then auto-classifies).
Leave {PR_NUMBER}, {REPO_ROOT}, {WORKTREE_PATH} as {...} literals — the
subagent fills them in at runtime.
-
Assemble the final prompt by concatenating, in order: the inner contents of
the Standard Prefix fenced code block, the spec content from step 2, and
the inner contents of the Standard Suffix fenced code block (do not
include the ## Standard Prefix / ## Standard Suffix markdown headers
or the surrounding triple-backtick fences). Substitute the variables from
step 3 into the result.
-
Dispatch — branch on Workflow tool availability (check whether a
Workflow tool is present in this session's available tools):
Workflow tool available: dispatch via the wave-workflow template
(references/wave-workflow.md) as a single-PR batch:
args.prs = one entry: { id: "pr1", branch, title, subagent, prompt, files },
where prompt is the assembled prompt and files is this PR's create/modify
list from the Step 2 descriptor. With one PR the overlap check trivially passes.
args.max_retries = 2.
- Invoke the
Workflow tool with the wave-workflow script and these args. The
call is async (see orchestrator.md Step 2a): it returns a launcher
response ({ status, taskId, error? }); if error is set the script failed
its syntax check — fall back to the Agent path. Otherwise wait for the
matching task_notification completion event and read the workflow result
from its output_file.
results[0] from that payload is this PR's result: the schema-validated
agent output plus an attempts count (in-script retries already handled).
Workflow tool unavailable (fallback — unchanged): dispatch the Agent tool:
subagent_type = <subagent declared in Step 2>
isolation = "worktree"
prompt = the assembled prompt
-
Get the result JSON: results[0] from the workflow (Workflow branch), or the
subagent's returned result JSON (Agent branch).
-
Print a completion report inline (see "Completion Report" below).
The full lifecycle still runs: Plan Mode → /tech-vetting → implement + test →
commit + push + gh pr create → /greenlight → CI poll → /merge-pr →
cleanup → result JSON. All defined by the unchanged pr-subagent-template.md.
Failure / interruption (run-now mode): there is no state.json to resume
from. If the session crashes mid-execution, manual cleanup is required: run
git worktree list and gh pr list to see what was left behind, then either
finish or close it before re-running /autopilot for a fresh attempt. If the
user wants crash recovery, they should pick "schedule" instead.
Single-PR + schedule and Multi-PR (CronCreate path)
- Ask the user for the desired execution time
- Read
references/orchestrator.md for the CronCreate prompt template
- Fill the plan path into the template
- Create a one-time schedule with CronCreate
CronCreate(
cron="<user-specified time>",
recurring=False,
prompt="<orchestrator.md template, with plan_dir path filled in>"
)
- Remind the user:
- The schedule is session-only — Claude Code must remain open
- If the session is interrupted, the orchestrator prompt can be pasted manually to resume (state.json tracks progress)
- No manual intervention needed during execution (unless blocked)
Completion Report (single-PR + run now)
On success:
Autopilot Single-PR done
═════════════════════════
✅ #91 feat(scope): <summary> — 2 rounds, fixed 1, pushed back 0
merged into main
spec: docs/loops/<YYYY-MM-DD>_<short-name>/pr1-<short>.md
On block:
❌ blocked: feat(scope): <summary>
reason: <error from subagent>
spec: docs/loops/<YYYY-MM-DD>_<short-name>/pr1-<short>.md
recovery: fix manually and resume, or re-run /autopilot
Important Notes
- This skill is interactive — each step requires user confirmation before proceeding
- Todos have typically already been through eng-review; no need to re-evaluate architecture direction, only confirm the PR split is reasonable
- Spec file quality directly determines execution success rate — better to spend extra time writing clearly