| name | octospec-workflow |
| description | Use when implementing a feature, fixing a bug, or making any non-trivial code change in this repository. Drives the octospec 4-phase engineering flow (Plan, Implement, Verify, Finish) so the change follows this repo's rules in .octospec/ and ships a PR with a linked spec. Triggers on requests like "add ...", "implement ...", "fix ...", "refactor ...", "change the ... API", 写功能, 修 bug, 加接口, 改逻辑. Skip for trivial edits (typo, docs, lint, pure config) — those do not need the flow. |
octospec workflow
This repository uses the octospec engineering standard. When you are asked to
make a non-trivial code change here, run the 4-phase flow instead of editing code
directly. Rules live in .octospec/ and are the source of truth for this repo's
conventions.
When to run this
Run the flow for: a new feature, a bug fix, a refactor, an API change, or any
change that touches load-bearing behavior.
Do NOT run the flow for trivial changes: a typo, a docs-only edit, a
lint-only fix, a pure config or dependency bump. Just make those directly.
The 4 phases
Run them in order. Each phase maps to a slash command the user can also invoke
manually; as a skill you perform the same steps.
1. Plan
- Choose a short kebab-case
<slug>.
- Read
.octospec/tasks/_brief.template.md.
- Inspect the relevant existing code, then write
.octospec/tasks/<slug>/brief.md with: Goal, Load-bearing list (use the same
tags as .octospec/rules/_index.yaml inject_when.touches where they apply),
Out of scope, Acceptance.
- Show the brief and get confirmation before writing code.
2. Implement
- Resolve applicable rules: read
.octospec/rules/_index.yaml and
.octospec/_global/ (if synced). A rule applies when its inject_when.paths
glob matches a file you will touch, OR its inject_when.touches tag is in the
brief's load-bearing list. A repo-tier rule overrides a global one with the
same id.
- Read the full text of each matching rule and follow it. Prioritize
load_bearing: true rules.
- Record what you used in
.octospec/tasks/<slug>/context.yaml.
- Write the code. Do not commit yet.
3. Verify
- Review the diff against each injected rule (trace load-bearing paths, not just
the happy path).
- Confirm the diff meets the brief's Acceptance and did not touch anything in
Out of scope.
- Run this repo's gates (lint / type-check / tests; see CLAUDE.md / AGENTS.md).
- Self-fix what you can.
4. Finish
- Run the final gate once more.
- Write
.octospec/journal/shared/<slug>.md (what was done + any learning).
Start it with OKF frontmatter (type: Journal + title/description/tags/
timestamp) and add a dated entry to .octospec/log.md.
- Promote any reusable learning in this same PR: edit the relevant
.octospec/rules/<rule>.md in place (or add a new rule + _index.yaml entry)
— the PR review is the gate. The helper .octospec/scripts/octospec-update-spec.sh
gives you the raw material (--kind=rule → a draft in
learnings/pending/<slug>-rule-draft.md + a promotion block on stdout;
--kind=task → a per-actor journal entry). It never auto-writes rules/, so
you copy the draft into rules/<id>.md and update _index.yaml here, in
this PR, then drop the scratch draft — never defer it to a separate PR or a
later move. Only use .octospec/learnings/pending/ for unresolved learnings
that still need human design before becoming a rule; finished learnings must
not be stranded waiting on a phantom separate PR.
- Open a PR. Fill the PR template's Linked Spec (→ the brief) and the
COMPREHENSION three questions to substance, for load-bearing /
architectural / P0 changes.
Notes
- This skill orchestrates the flow automatically. The user can still drive any
single phase manually with
/octospec-plan, /octospec-go, /octospec-check,
/octospec-finish.
- The flow is guidance, not a hard gate; the repo's PR/CI checks are the
enforcement layer.