AGENT SKILL

flow-implement

Implement a requested software change through verification, corrections, and its authorized delivery endpoint.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
flow-implement
description
Implement a requested software change through verification, corrections, and its authorized delivery endpoint.

02 资源文件 1

03 SKILL.md 正文

Implement a change

For non-trivial work, first restate in your own words the user's goal, the problem it solves and how you plan to approach it (in-your-own-words); keep going unless that surfaces a real ambiguity. Recover the requested outcome, constraints, working surface and authorized endpoint. Reuse settled decisions and repository context relevant to this change. Follow-up corrections steer the whole task unless the user replaces it. Define completion by observable behavior and delivery, not by producing a first diff.

For non-trivial repository work, use coding-orchestration for ownership, delegation and verification. Reuse an already selected orchestrate-work method instead of adding another execution loop. Clear work proceeds without another planning or approval ceremony. A bounded worker executes its supplied packet. Use the exact-default execution contract for ordered actions, expected observations and receipts. Adapt the packet to available capability without reducing acceptance. Resume model/quota changes through its linked continuity handoff, preserving valid evidence and ownership.

For a stateful UI or routing change, preserve the relevant state owners and transitions: drafts, selection, filters, focus, dismissal and permissions. Check the affected transition early; two static endpoints do not establish continuity. For a visual redesign, recover the approved assets and criteria through design-direction before implementation. Refinements preserve the current identity.

Use checks that establish the changed behavior and required repository gates. Add no automated tests; real verification observes the changed behavior. A regression claim needs evidence that distinguishes the broken and corrected behavior. Preserve command exit codes, resolve actual findings and rerun affected checks. Reuse still-valid evidence. Continue through authorized implementation, verification and correction without asking for another continuation message after the first pass.

For an on-screen outcome, use the repository's existing product-verification capability to exercise the affected flow. Capture a useful baseline before edits where possible; missing baseline evidence remains a declared gap. Invoke maintain-product-verification only when that verification capability itself needs maintenance. When visual quality is acceptance, design-quality owns independent critique before evidence-capture packages final media. Record the affected behavior with useful data and the intended role. A backend change with no on-screen outcome uses relevant behavior checks without a visual workflow.

Finish every requested criterion and the installation/publication steps already authorized. A missing authorization for a distinct external action does not stop preparation or independent work. State partial or blocked work plainly. The final response covers the complete outcome, checks, remaining gaps, candidate and requested review link; code, Preview and production are separate claims.

For a requested standalone completion report or a durable media package, read completion-report.md.

Before integrating a PR

  1. Write or refresh the system-recap block from the diff; update docs/architecture/primitives.json in the same PR when a primitive changes.
  2. Review the finished PR once, never each push. One independent Sol 6.1 High review of the ready head; a docs-only or small low-risk PR skips it. A second round only for a blocking finding (security, authorization, data loss, money), reviewing only the fix diff. Non-blocking findings are fixed without re-review or recorded as follow-ups. No third round without the operator.
  3. Codex review: development-system pr-review-status <pr>. pending → run it with --request once and wait (T3 watch_pull_request when available); findings → fix or answer each one in its thread. Ask again at most once after the fixes; after that second review, answer its findings, push only their fixes and integrate on budget-spent, naming the later fix commits in the report. Integrate on no-findings or budget-spent only. unavailable is not approval.

Leave nothing running

Stop every dev server, model server or watcher you started before the final report: kill the PIDs you launched, never by name, and never another thread's.

Friction

Before the report, look back at the session for paper cuts: slow or flaky startup, missing test data, a workaround for a missing tool, a doc that was wrong. Record each with development-system friction add --id <slug> --incident <ref> --evidence <ref> --repository <repo> --summary <text>. Frictions feed the Steward; confirmed root causes still go to mistake add.

Close every task

Every task that changed files ends with a report:

  1. Write the concise packet (references/completion-report.md). Its Markdown opens with the sections Qué se hizo, Hallazgos, Qué sigue, Detalle in that order (English: What was done, Findings, What's next, Detail); the command rejects other shapes and tables without allowTables.
  2. Render it with development-system document --input <packet> --json.
  3. Deliver it durably: in T3, render the page in the owning thread; otherwise give the persistent htmlPath the command returns.
  4. A working-backwards reader-live.mjs --tunnel link is optional; when you add one, state its expiry next to the durable delivery, never instead of it.

A Stop hook asks once when files changed and no report was produced. If a report is truly not useful, say why in one line. When the user asks for the answer in the chat only, produce no document and no tunnel.

Automated tests and evidence

No automated tests anywhere: do not create, run or restore them; delete them when found (check-no-tests and the guard enforce this). Every task includes real verification without being asked: computer use, browser, and the repository's verification CLI and feature map. Report passed / failed / not reached with evidence.

Headroom transport evidence

Headroom is an explicit same-account per-invocation option. Keep the existing Codex binary and CODEX_HOME, preserve caller arguments, and request lossless, cache-conservative operation. Provider cache reuse can still change request bytes; do not promise byte identity. Keep native usage from the parent and every actual descendant separate from proxy counters, and separate requested model or tier from observed identity. Use launch/completion events instead of polling. Proxy counters and token deltas are descriptive; without complete controlled evidence, do not claim Headroom caused savings.

Spec and ticket handoff

Use the shared definition contract for stable criteria, linked tickets, dependencies, rules and required evidence. Preserve the selected parent. With an OpenAI parent, Opus 5.5 authors the plan and a fresh independent Sol 6.1 High reviews it; with an Anthropic parent, Sol 6.1 High authors and a fresh independent Opus 5.5 reviews it. Use the existing read-only routes and honor provider restrictions. Missing cross-family contrast stays not reached; unresolved blocking objections prevent acceptance. Fast researchers supply bounded source facts. When implementation is authorized, use coding-orchestration and its advisory Jev recipe across the complete feature, including nonvisual behavior. Retrieve all tickets linked by the spec; preserve settled decisions across threads.

04 同仓库更多 Skills