/autonomous-dev

name: autonomous-dev
description: "Drives a brief, specification, issue or existing PR through implementation, review, tests and QA to a ready PR. Persists ownership, progress and commit-bound evidence for safe resumption. Use for autonomous software delivery or finishing an interrupted development run."
effort: high
argument-hint: "[setup | run <task> | list | resume <run-id> | status <run-id>]"
allowed-tools: Read, Write, Edit, Bash, Glob, Grep

Autonomous Development

$ARGUMENTS

<!-- CLAUDE_CODE_ONLY_START --> Only in Claude Code, apply the `model-routing-patterns` skill when choosing executors or creating agent definitions. Delegate to `codex:codex-rescue` only when its plugin is installed, enabled and callable in this session. Otherwise use the installed native agents and their configured models. A context without the Agent tool returns the dispatch decision to its supervisor; it does not invent a tool or bypass the client. Preserve explicit user choices and verify actual completion before accepting a delegated result. <!-- CLAUDE_CODE_ONLY_END -->

Own one software-delivery run from its requested outcome to a reviewed, verified PR. Continue through reversible decisions within the user's approved scope, recording assumptions where they can be reviewed. Use existing skills for their actual work; this skill owns sequencing, state and the exit gate.

Entry points

| Input | Action | |---|---| | `setup` | Inspect the project and create its validation/tracker/QA configuration | | `run <brief, spec path, Jira key, issue URL or PR URL>` | Resolve the task, reuse existing work, then run the process | | `list` | List this repository's existing runs without changing state | | `resume <run-id>` | Read durable state, claim ownership and continue at the first unmet gate | | `status <run-id>` | Read state and report the next action; change nothing |

A plain task description means `run`. `/workflow autonomous-development` routes here. A request only to review code stays with `/review`; an incident stays with the incident-response workflow.

Setup and authorization

Read [project configuration](reference/project-config.md) on first use or when project commands change. Check the target repository's instructions, KB, actual build/test commands and provider identities. Keep `issueTracker`, `codeHost` and `knowledge` separate. For Jira/RAG projects, read the [stack integration contract](reference/stack-integrations.md) before discovery and reuse it for resume, QA and finalization.

Validate the reviewed configuration locally before initializing the run:

python3 ${CLAUDE_SKILL_DIR}/scripts/delivery-config.py --config /absolute/project/.ai-toolkit/autonomous.json
python3 ${CLAUDE_SKILL_DIR}/scripts/delivery-config.py --config /absolute/project/.ai-toolkit/autonomous.json --task PROJ-123

Pass `--task` only for a Jira key/browse URL; briefs, specs and code-host PRs use their own source resolution. For Jira, use the returned canonical subject after checking the real task/instance and Git repository mapping. Preflight does not contact MCP or prove those live conditions. Configuration records project facts; it cannot grant publishing permissions, change models or relax host safeguards.

Capture the task, acceptance criteria, exclusions and authorized endpoint before implementation. The default endpoint is `ready-pr`. An explicit request to run this process may cover the whole plan; preserve approvals already given instead of asking at every phase. Ask only for an unresolved decision required by the task or a missing permission required by the current host. Progress independent work while the question is pending. Merge and deploy require their own explicit authorization and are not performed by the state helper.

Durable state

Use the bundled helper from the installed skill directory:

python3 ${CLAUDE_SKILL_DIR}/scripts/run-state.py --help
python3 ${CLAUDE_SKILL_DIR}/scripts/run-state.py --repo /absolute/project list
python3 ${CLAUDE_SKILL_DIR}/scripts/run-state.py --repo /absolute/project init --subject brief:csv-export --objective "Export filtered orders as CSV" --source-kind brief --source "User task" --target-branch develop
python3 ${CLAUDE_SKILL_DIR}/scripts/run-state.py --repo /absolute/project status --run RUN_ID

Use actual values returned by the helper for `RUN_ID`, owner and artifact directory. Read [state operations](reference/run-state.md) before the first mutation or a resume. Every writer supplies its per-session owner token. A GitHub account name is not a unique run owner. Do not initialize a replacement run to evade an existing claim, failure counter or missing evidence.

Generated state and reports live outside the target repository under its ai-toolkit session store. Store intentional specifications and regression tests in the target project's normal locations. Keep terminal output, run reports and screenshots in the helper's artifact directory so they cannot dirty the code being verified. The journal records evidence; it does not run tests or certify the truth of a report. Inspect the actual tool results before recording them.

Delivery process

Read [execution and recovery](reference/process.md) for the detailed stage contract. The essential chain is:

1. **Discover and claim.** Resolve the source and search for an existing branch, run and PR. For a bug, verify the symptom still exists before changing code. For Jira, refresh the scoped task; for RAG, retrieve project SOPs and rules. Freeze the relevant task/KB context in the hashed plan. Reuse the PR and its branch when present. Use an isolated worktree for new work, preserving the user's checkout. 2. **Plan.** Map each acceptance criterion to an implementation slice and its verification. Record scope, ownership and dependencies. A small change needs a short plan; a specification needs explicit slices and checkpoints. 3. **Implement.** Use the relevant development agen