A hardening workflow for changes that need more than a happy-path pass.
contract → implement → verify → break/review → fix → re-verify → re-break/review → stop
Use it when a change touches risk: deploys, migrations, backups, auth, credentials, rollback, CLI/tools, filesystem/process handling, generated artifacts, or operational docs people may copy-paste.
Generated with image_gen / openai-codex / gpt-image-2-medium. A deterministic HTML/SVG fallback is in assets/anvil-loop-flowchart.zh.html.
In the graph, 必须处理? means a blocker or proven issue that must be handled before shipping. Non-blockers can be documented, accepted, or deferred.
Do not stop at “it works once.”
- State the contract.
- Implement the smallest useful change.
- Verify directly.
- Get independent break/review.
- Classify findings.
- Convert real findings into tests, repros, or exact doc evidence.
- Fix only proven problems.
- Re-run checks and break/review again.
- Stop only when there are no blockers.
If no independent reviewer is available, label it as self-review and do not claim the anvil loop is complete.
Canonical skill file:
skills/software-development/anvil-loop-development/SKILL.md
For a local Hermes install, use the raw skill URL:
hermes skills install \
https://raw.githubusercontent.com/getaskclaw/anvil-loop/main/skills/software-development/anvil-loop-development/SKILL.md \
--category software-developmentManual fallback, run from this repo root:
mkdir -p ~/.hermes/skills/software-development/anvil-loop-development
cp skills/software-development/anvil-loop-development/SKILL.md \
~/.hermes/skills/software-development/anvil-loop-development/SKILL.md
hermes skills list | grep anvil-loop-developmentThen start a new Hermes session. In an existing CLI session use /reset; in Telegram/Discord gateway use /restart if the tool/skill list does not refresh. Load it with /skill anvil-loop-development.
There is intentionally no root SKILL.md, to avoid duplicate drift.
- Contract stated
- Intended files only changed
- Tests/build or Markdown smoke checks passed
- Independent review completed
- Blockers fixed with evidence
- Checks re-run after fixes
- Re-review completed if blockers were found
- Final state and limitations reported plainly
MIT
