autodev
The `/autodev` skill manages the operational contract defined in `PROGRAM.md` or `AUTODEV.md` that governs autonomous development loops. Use it to create, validate, repair, and inspect the contract parameters including mutable scope, immutable scope, validation commands, escalation rules, and stop conditions that the `evolve` and Factory drivers consume during each tick of the operating loop.
git clone --depth 1 https://github.com/boshu2/agentops /tmp/autodev && cp -r /tmp/autodev/images/gemini/skills/autodev ~/.claude/skills/autodevSKILL.md
# /autodev
`/autodev` manages the repo-local operational contract for autonomous development.
It does not replace `evolve` or `rpi`.
## Loop position
The config/intent layer the [operating loop](../../docs/architecture/operating-loop.md) reads each tick — NOT a loop itself. `/autodev` defines and validates the contract declared in `PROGRAM.md`/`AUTODEV.md` — mutable scope, immutable scope, validation commands, escalation rules, stop conditions. The drivers ([`evolve`](../evolve/SKILL.md) and Factory, the daemon) consume that contract and run the loop; autodev does not run it. Loop discipline still applies under autonomy: no parallel wave without the wave-validity check; no slice closes without a passing test mapped to a Given/When/Then; capture goes through the promotion ratchet, not into a landfill.
- `PROGRAM.md` or `AUTODEV.md` defines the contract: mutable scope, immutable
scope, experiment unit, validation commands, decision policy, escalation rules,
and stop conditions.
- `ao autodev` creates, inspects, and validates that contract.
- `evolve` runs the v2 autonomous improvement loop.
- `ao rpi` runs one research -> plan -> implement -> validate lifecycle.
## Routing
Use this split when the user asks whether the old evolve flow should become a
new command or skill:
| Intent | Action |
|--------|--------|
| define or repair the repo-local autonomous policy | use `/autodev` and `ao autodev` |
| run the autonomous improvement loop | use `evolve` or `evolve` |
| run one bounded lifecycle | use `rpi` or `ao rpi` |
`PROGRAM.md` takes precedence over `AUTODEV.md`. Treat `AUTODEV.md` as the
compatibility alias.
## Execution Steps
### Step 1: Detect the contract
```bash
if [ -f PROGRAM.md ]; then
PROGRAM_PATH=PROGRAM.md
elif [ -f AUTODEV.md ]; then
PROGRAM_PATH=AUTODEV.md
else
PROGRAM_PATH=
fi
```
If a contract exists, validate before using it:
```bash
ao autodev validate --json ${PROGRAM_PATH:+--file "$PROGRAM_PATH"}
```
If no contract exists and the user asked to initialize or define the loop, create
one:
```bash
ao autodev init "<objective>"
```
Infer the objective from the user request when it is clear. Ask only when the
objective cannot be discovered from repo context and inventing it would make the
contract misleading.
### Step 2: Repair or explain validation failures
When validation fails, inspect the missing fields and patch the program file if
the user asked to create or fix the contract. Required sections:
- `Objective`
- `Mutable Scope`
- `Immutable Scope`
- `Experiment Unit`
- `Validation Commands`
- `Decision Policy`
- `Escalation Rules`
- `Stop Conditions`
Prefer narrow mutable scope and concrete validation commands. If the needed work
crosses immutable scope, create or update a bead instead of silently widening the
contract.
### Step 3: Hand off to the loop
After `ao autodev validate` passes:
- For one lifecycle, run `/rpi "<goal>"` or `ao rpi ...`.
- For the repeated autonomous loop, run `evolve` (skill-driven) or dispatch it via NTM for out-of-session runs.
- If both `PROGRAM.md` and `GOALS.md` exist, `GOALS.md` is strategic fitness and
`PROGRAM.md` is the operational execution layer.
Do not mark an autonomous cycle successful only because the main tests pass. The
program validation bundle and stop conditions must also be satisfied.
## Examples
```text
User: turn this postmortem/analyze/plan/pre-mortem/implement/validate loop into
a v2 command.
Agent: Explain that `evolve` runs the loop, then create or validate
`PROGRAM.md` with `/autodev` so the loop has explicit scope and gates.
```
```bash
ao autodev init "Continuously improve AgentOps skills within explicit scope."
ao autodev validate
# then run the repeated loop: /evolve (skill-driven; /evolve CLI retired, ag-llni)
```
## Troubleshooting
| Problem | Response |
|---------|----------|
| `PROGRAM.md not found` | Run `ao autodev init "<objective>"` when setup is requested. |
| validation reports missing sections | Patch the missing required sections, then rerun `ao autodev validate --json`. |
| requested work is outside immutable scope | Stop direct edits and create a bead or ask for an explicit contract change. |
| user asks "is this evolve?" | Answer: `autodev` defines the loop contract; `evolve` runs the loop. |
## Reference Documents
- [references/autodev.feature](references/autodev.feature) — Executable spec: contract-bounded unattended loop, manages-not-replaces evolve/rpi, loop-discipline-under-autonomy (soc-qk4b)
- [references/autodev-cli.feature](references/autodev-cli.feature) — Executable spec: `ao autodev` CLI command behavior, linked to cmd tests (soc-jnfgi)Use Agent Mail from Codex for file leases, notifications, inboxes, and conflict prevention.
>-
>-
Use when converting markdown plans into br beads with dependencies for implementation or swarm execution.
Use when switching AI coding CLI accounts quickly to recover from subscription rate limits or OAuth friction.
>-
Use when starting non-trivial work, mining lessons, or preventing repeated mistakes with cm procedural memory.
Mine past agent sessions for working prompts, decisions, and patterns. Use when "what did I ask?", "find that prompt", session archaeology, or agent history.