refactor(skill): плейбук возвращён к оптике тонкого мастера

- Зачем:
  - субагенты должны сохранять внимание мастера и помогать результату, а не обслуживать обязательный конвейер.
- Что:
  - семиступенчатый процесс заменён эвристиками выбора исполнителя, проверяющего и независимого ревьюера.
  - права агентов отделены от канала результата, а полная полезная нагрузка выведена из контекста мастера.
  - добавлены авторитетные договоры промптов, полевые заметки и handoff следующей сессии.
- Проверка:
  - git diff --cached --check.
  - свежий обзор через writing-great-skills, все принятые находки закрыты.
This commit is contained in:
ddadmin committed 2026-07-12 23:50:37 +03:00
1 parent 7110a075ea
commit f9ea5b8c44
10 files changed
+344 -290

No files matched your search

@@ -0,0 +1,14 @@
# Narrow re-check
Re-check only your original findings and the changes immediately adjacent to
their resolution. Do not perform a full review.
For every original finding report `CLOSED`, `NOT_CLOSED`, `REJECTION_VALID`,
or `REJECTION_INVALID`, with concise evidence. If this narrow work exposes a
new adjacent defect, report it in the normal findings format; the scope bound
limits exploration, not honest reporting.
Return the full re-check through the channel named by the caller. For a file
channel, write it to the supplied path and return only
`RECHECK: ALL_CLOSED | ISSUES_REMAIN`, counters, and `REPORT: <absolute path>`.
For a message channel, return the concise finding statuses once.
@@ -0,0 +1,14 @@
## Result contract
Write the full result through the channel chosen by the caller:
- `STATUS: complete | questions | blocked`
- `EVIDENCE:` changed files, executed commands, and final counters
- `DECISIONS:` only questions for the orchestrator or user
- `FINDINGS:` severity, location, essence, and verification method
- `NEXT:` one concrete action or `none`
Omit empty sections. Do not include raw output or a reasoning narrative. For a
file channel, put the full result in the named file and return only status,
counters, and `REPORT: <absolute path>`. For a message channel, return one
concise result without a report path.
@@ -0,0 +1,21 @@
## Mandate
- Review and test hypotheses; do not edit source files or persistent
environment state.
- Use the working permissions needed for checks, caches, temporary artifacts,
documentation and primary-source research. The caller verifies Git before
and after the review.
- Do not run `git add`, `git commit`, or commands that discard changes.
- Report only the findings index and verdict through the channel named by the
caller.
## Findings index
For each finding:
`SEVERITY(BLOCKER|MAJOR|MINOR|NIT) | file:line or task section | failure scenario | evidence | how to verify`
If there are no findings, list the reviewed scope and every executed check or
tested hypothesis with its outcome. A bare no-findings assertion is invalid.
Final line: `VERDICT: APPROVED | APPROVED_WITH_MINORS | CHANGES_REQUIRED`
@@ -0,0 +1,4 @@
## Runtime language
Write every runtime prompt result, report, findings index, and exchange file in
English. Do not infer the output language from the repository or conversation.
@@ -0,0 +1,15 @@
# Triage blind-review findings
The review indexes below are immutable input. Process every finding and change
nothing outside their scope. For each finding, either:
- fix it and name the contract or check that proves closure; or
- reject it with evidence from the contract and a check, never a bare claim.
Do not commit or stage changes. Propose accepted rejections and residual risks
for the orchestrator to own and record as durable task decisions. The
orchestrator cannot silently accept risk; authority must come from the user or
already be explicit in the task.
Write a full resolution report to the path supplied by the caller. Return only
status, counters, executed checks, and that path in the final message.