AGENT SKILL

sdlc-deploy

Prepare a verified change for human PR review by running REVIEW.md passes through the sdlc-reviewer subagent and posting ranked findings. Do NOT use for intent or scope decisions, accepting organizational risk, self-approval, merging, deploying, or bypassing branch protection.

使用场景运维与部署
License 许可
Apache-2.0
最近更新
2026-11-14
Stars
★ 1.6k
Forks
491

01 前言 Front Matter

name
sdlc-deploy
description
Prepare a verified change for human PR review by running REVIEW.md passes through the sdlc-reviewer subagent and posting ranked findings. Do NOT use for intent or scope decisions, accepting organizational risk, self-approval, merging, deploying, or bypassing branch protection.

03 SKILL.md 正文

SDLC Deploy

Run Stage 5 of 6 in the AI-Native SDLC playbook described in docs/PLAYBOOK.md.

Purpose

Prepare the pull request and review evidence. This stage posts findings and stops for a human; it never merges or deploys.

Do not judge whether the original intent was worthwhile. Do not accept risk for the organization.

Entry gate

Proceed only when sdlc-test has completed with green verification evidence and the code is ready for review.

Read the test report. Confirm it includes actual passing output and applicable eval evidence for any changed CLAUDE.md, skill, or hook.

If verification is incomplete, stale after changes, missing, or red, return to sdlc-test and stop.

Procedure

  1. Identify the slug for this change and read .sdlc/<slug>/spec.md and .sdlc/<slug>/plan.md. The PR description must link them, so a reviewer can see the accepted intent, spec, and plan behind the diff without leaving the PR. A fast-path change has only plan.md: link that, and quote its ## Fast path declaration in the PR so the reviewer sees who called the change low-risk and why.
  2. Locate the project's REVIEW.md.
  3. If it exists, read every applicable review pass, the depth setting, the round cap, the sensitive-path list, and the nit cap.
  4. If it does not exist, record that fact and continue with the required review passes using sound judgment.
  5. Read .sdlc/<slug>/review-findings.md if it exists. It carries the round number, the SHA each earlier round reviewed, and every finding already posted.
  6. If the change took the fast path, compare the real diff against the sensitive-path list now. A diff touching any sensitive path revokes the fast path: say which path it reached, and route the change to sdlc-plan to enter through Stage 1 and Stage 2, exactly as sdlc-build would have refused the declaration had the file been declared. Do not review it onward.
  7. Determine the round number. If this would be round 4, stop reviewing: record that the cap is reached and return the author to plan.md to reconsider the approach rather than fixing findings one by one.
  8. Determine the review scope. Round 1 reviews the whole change; later rounds review only the diff since the previous round's recorded SHA.
  9. Assemble the changes in scope, verification evidence, and REVIEW.md instructions.
  10. Invoke the sdlc-reviewer subagent in review mode with the changes and the passes REVIEW.md names for this depth. At Standard and Deep that is Bugs, Security, and Compliance; at Quick it is Bugs and Security. Security runs at every depth.
  11. Drop every finding an earlier round already posted. Count them.
  12. Invoke the sdlc-reviewer subagent again in verify mode, in a fresh context, with the surviving findings and the diff they cite, and without the reasoning that produced them. Discard every finding it judges not real. Count them.
  13. Require each surviving finding to carry a REVIEW.md severity: Important, Minor, Nit, or Pre-existing.
  14. Enforce the nit cap from REVIEW.md; for example, if it says max 5 nits, include at most five nits.
  15. If REVIEW.md has no cap, use judgment and keep nits actionable.
  16. Append this round to .sdlc/<slug>/review-findings.md using the plugin's template: round number, stage, reviewed SHA, scope, passes run, findings, suppressed count, and verifier-rejected count.
  17. Scan .sdlc/*/review-findings.md for a finding class that now appears in two or more changes, and record it under Recurring in this change's file, with what encodes it.
  18. Create or update the pull request with context, test evidence, and findings.
  19. Post the findings as a PR comment or include them in the PR body.
  20. Cite the review findings and explicitly stop for human approval.

Give the reviewer only evidence and review scope, not authority to waive risks.

Reviewer request

Tell the sdlc-reviewer subagent to inspect the actual diff and report:

  • Bugs found by the Bugs pass.
  • Security concerns found by the Security pass.
  • Compliance concerns found by the Compliance pass.
  • Severity for each finding: Important, Minor, Nit, or Pre-existing.
  • File and line references where available.
  • Concise rationale and a suggested remediation when appropriate.
  • How many findings it suppressed as already posted in an earlier round.

A real defect in code this change did not introduce is Pre-existing. The reviewer reports it; it does not block merge and is not counted against the author. A pre-existing defect that is a security finding stays under Security at its real severity and blocks merge like any other.

The verifier runs as the same subagent in verify mode, in its own context, and sees only the findings and the cited code. Do not pass it the reviewer's reasoning, and do not let it change a severity or add a finding. It answers one question per finding: does this failure actually happen?

Pass through the project's REVIEW.md rules. Do not replace them with an invented review rubric.

If no findings exist, record that the passes found none rather than fabricating comments.

Findings and PR content

Separate Important findings from Nits in the PR content.

Keep the number of nits within the REVIEW.md cap. Important findings are not nits and must not be hidden by the cap.

List every Pre-existing finding in the PR under its own heading, with the location and a proposed disposition: fix now, a backlog reference, or accepted with a reason. A pre-existing finding is never silently dropped. Nothing else in this loop looks for old defects, so this list is the only place they surface.

Include test evidence from Stage 4 so a human can evaluate the change.

State whether REVIEW.md was found and, if so, which passes governed the review, at which depth, and which round this is.

State the round's counts: findings posted, suppressed as already reported, and rejected by the verifier.

Do not alter code merely to make findings disappear without returning to the test stage for fresh verification.

Running in CI

The same review runs unattended. In CI, invoke Claude non-interactively inside a sandbox with filesystem and network isolation, scoped to the repository and the review passes. A sandboxed CI run posts findings exactly as an interactive run does; it gains no authority to merge, deploy, or waive a finding.

Treat a CI review as evidence for the human reviewer, never as a substitute for one.

Prohibitions

Never merge the PR.

Never deploy the code.

Never bypass, disable, or work around the approval-gate hook. It fires on its own, and branch protection enforces the merge gate; an agent cannot override either.

Never accept a security, compliance, operational, or product risk on behalf of the organization.

Never treat your own PR post as human approval.

Exit gate

Before posting, check Recurring in .sdlc/<slug>/review-findings.md. Every finding class now seen in two or more changes must name what encodes it: a line in CLAUDE.md, a lint rule, or a rule in REVIEW.md. A row with no encoding holds this gate. Fixing the instance is not fixing the class — the third occurrence is the one this prevents.

Post the PR and cite the ranked findings. Then explicitly stop.

The next step must be a human approval, either through branch protection or an explicit LGTM.

If human approval has not happened, do not hand off to maintenance.

Handoff

Only once a human approves merge through branch protection or explicit LGTM, proceed to sdlc-maintain after the code is deployed.

Carry forward the PR link, approval evidence, review findings, and deployment identity if available.

04 同仓库更多 Skills

hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-review-with-annotations

Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.

★ 1.6k
hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-find-and-synthesize-context

Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.

★ 1.6k
hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-create-or-update-docs

Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.

★ 1.6k
hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-create-or-manage-records

Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.

★ 1.6k
hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-create-or-manage-interactive-html

Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.

★ 1.6k
hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-collaborate-threads

Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.

★ 1.6k