AGENT SKILL

day

暂无简介

License 许可
MIT
所属仓库
paranoica/bedrock
最近更新
2026-11-08
Stars
★ 3
Forks
1

02 资源文件 6

03 SKILL.md 正文

The working loop

One block = one task from "taken" to "done and merged" (D031). Each step ends on a check. The only hook that refuses is the pull-request review gate; everything else here is the session's own discipline.

Triggers

  • RU: «сделай X», «новая задача», «начинаем день», «закрываем день»
  • EN: "build X", "fix X", "new task", "start the day", "close the day"

0. Start of day (once per work day)

Follow references/start-of-day.md. Done when the owner has the three-line summary "вчера / входящие / сегодня".

1. Intake

  • Size the request: R1 (no logic), R2 (ordinary code), R3 (risky or new), and P0–P3 (D032).
  • P0 goes to the urgent path in references/urgent.md and skips the plan.
  • Look facts up instead of asking. Questions depend on the risk (D033):
    • R1: none.
    • R2: only when the scope is genuinely unclear. One message, at most five questions, each with a default and the cost of a wrong default.
    • R3 or a new feature, when the owner says «грилл» or the task is murky: an interview, one question at a time, each answer saved to the plan file at once. Otherwise as R2.
  • Log it: bedrock log taken --block <id> --detail "<R/P> <title>".

Done when: the taken line carries R and P, and the questions (or "no questions") are in the plan file. R1 has no plan file: its taken line is the record.

2. Plan

  • R1: no plan. R2: a short plan. R3: a full plan with a design section (D035).
  • Write it from references/plan.md to ~/projects-days/<project>/<date>/<block>/plan.md.
  • Every plan has: done-when (checkable), assumptions, files, tests named with what they assert, tasks with who and model, the agents it uses, estimate, out of scope.
  • The agents list says which roles run and how often, so bedrock agents can be read against it afterwards. Each agent is the bedrock role that fits (inventory, researcher, implementer, reviewer, critic); general-purpose only where no role fits, since it runs on the session's own model.
  • Research per references/research.md when the plan rests on library behaviour (D034).
  • The project's packs feed the plan: their test checklists go into Tests, their rules into the implementer brief's "Keep true".

Done when: the plan file exists and its done-when is a test, a command or a number.

3. Critic

  • R1: none. R2: one critic agent, one round. R3: two critics on different models, at most two rounds (D036). A second round with no fewer blockers goes to the owner.
  • Brief from the plugin's templates/briefs/brief.md; give it the plan, done-when, decisions and code rules.
  • Take each finding with evidence or drop it; a blocker is re-read in the code first.
  • Tell the owner only the delta: "критик: 5 найдено, 3 принято, +20 мин".

Done when: the plan file has a "Critic" section with taken and dropped findings.

4. Owner's go

  • R1: the request itself is the go.
  • R2, R3: show the plan in five lines and wait for «го» (one go covers the whole plan, D016).
  • Then bedrock log plan-approved --block <id>, and:
    bedrock session set task="<title>" plan='[0,<n>]' estimate_min=<m> state=work
    

Done when: the log line exists and the status line shows the task.

5. Cut into tasks

  • One task = one commit with its own test, about 30–40 minutes of human work, its own files (D037). Add each with bedrock backlog add '<json>'.
  • Set auto-continue for the first task: bedrock session set autocontinue.next='{"id": "T1", "title": "...", "safe": true}'. A task with an outward or destructive step is never safe: set null instead, so the session stops there and asks.

Done when: every task is in the backlog with files, test, who and estimate.

6. Implement each task

  1. bedrock backlog claim <id> --session "$CLAUDE_CODE_SESSION_ID" (it sets doing).
  2. Follow references/implement.md (D039): a failing test first, minimal code, verify after every few edits, refactor only while green, read the diff yourself, commit.
  3. After the commit: bedrock backlog status <id> done, bedrock log task-done --block <block> --ref <commit>, bedrock session set plan='[<k>,<n>]', and point autocontinue.next at the next task in order. It is null when none is left or the next one is not safe. Never skip over an unsafe task to a later one.

Done when: the task's test failed before and passes now, the commit exists, and the backlog shows the task done.

7. Review

  • Use the review skill for every block. It scales itself: R1 gets the checks and a quick read, R2 and R3 get both reviews (D050).
  • The full suite runs once at the end of the block and before a PR: bedrock cmd test (D040).
  • bedrock log review --block <id> --detail "<the summary line>".

Done when: the review summary line is logged and bedrock review pass <branch> --agent <id> is recorded with the reviewer agent that ran.

8. Deliver

Follow references/deliver.md: branch and PR per the project, live verification, and every outward step only on the owner's word in the conversation (D052, D053). After a merge: bedrock log merged --block <id> --ref <PR>.

Done when: the change is checked live, or the report says plainly that it was not, or it names the outward step that waits for the owner's word.

9. Close the block

  • bedrock log closed --block <id> and bedrock minutes <id> into the report.
  • Update the knowledge base for every module touched; bedrock kb drift shows what is left.
  • Report to the owner in plain Russian (D021): what changed for the product, checked live or not, what is needed from the owner, what is next.
  • bedrock session set state=null task=null plan=null autocontinue=null.

Done when: the log line, the base update and the report are all done.

End of day

references/start-of-day.md § End of day. Done when the owner has the day summary.

04 同仓库更多 Skills