execute-issue

Files (1)

• SKILL.md

SKILL.md (first lines)

<!-- GENERATED by scripts/gen_adapters.py from skills/execute-issue/SKILL.md.
Do not edit: run the generator instead. The canonical workflow is
reference/execute-issue.md; this file only routes an agent to it. -->
**In Claude Code, check the model first, and again on every resume after a stop** (Codex: skip). Read `model_pins["execute-issue"]` from `python3 .cadence/tools/cadence_config.py --json`; if it is set and you are not running on it, say so and stop — name both models and how to switch (`/model <name>`). Continue only if the author says to. Unset: no check.

Follow the canonical procedure in `.cadence/reference/execute-issue.md` exactly.

Branch off fresh main, named per `[git].branch` (default: the lowercased issue id). Implement only the issue's scope.

Classify every decision before asking or guessing — reversal cost and blast radius
decide the branch, not how hard the question feels
(`.cadence/reference/execute-issue.md` § *Decisions*):

- One defensible answer, or wrong is cheap to undo → decide it and keep going.
- A one-way door whose effects stay inside this diff → with `[fanout].enabled = true`,
  **fan out**, per `.cadence/reference/decision-fanout.md`: build each
  option in its own worktree and run review in every leaf. With fan-out off (the
  default), ask, as below.
- `must-stop`, or a `fork` refused for a cap → a Shape A decision brief
  (`.cadence/reference/human-brief.md`) sent with
  `python3 .cadence/tools/ask.py ask` and issue context; mirror Q+A onto the issue.

Work the `must-stop` boundary from `[[must_stop]]` in `cadence.toml`
(`python3 .cadence/tools/cadence_config.py must-stop <path>`), never from a
prose summary of it.

Exit 3 alone is not the ask signal: it is every `fork` refusal. Read the printed
reason. A cap or `must-stop boundary:` is the third branch; `only N surviving
option(s)` is the first — decide it and `record`. See `execute-issue.md` § *Decisions*.

Green lint + tests, then run the code-review skill to LGTM or its circuit breaker, then `gh pr create --title "<[git].pr_title, default {issue}: {summary}>" --body-file <brief>` with its Shape B brief — never `--fill`, never `--body-file` alone (prompts, fails headless).

Post the PR URL on the issue and notify the ask transport, then **keep going** — CI must pass on the head commit — `gh pr create` is the middle of the
procedure, not the end. Follow `.cadence/reference/automated-review.md`
(when `[review].provider` is a PR bot) and `execute-issue.md` § *When this skill is done*
rather than restating them.

Never merge, never touch main, never force-push.

From arbor-ai-inc/cadence, Apache-2.0.

More from arbor-ai-inc/cadence