[devfive-relay-c70423]

What

/devup-retro: one command to type when a piece of devup-mcp work is finished. The agent reviews the conversation from the first devup-mcp call to now, extracts every follow-up prompt the user had to write to correct or complete the result, and turns them into ready-to-paste improvement prompts for devup-mcp development. The goal is that the same follow-up never has to be written twice.

Three phases, never merged:

1. review: each item carries the user's prompt quoted verbatim, what was wrong, the evidence (file / Figma node / tool output), and exactly one surface that should have prevented it: codegen, export-output, validation-rule, runtime-audit, skill-guidance, bootstrap-instructions, project-plan-intake, or out-of-scope:devup-ui|app|figma-data. Items with no user-prompt evidence are excluded. Items already fixed are marked fixed-in-installed / fixed-not-installed / already-recorded, or unverified when release notes cannot be read.
2. confirm: the list is shown as a table and the runtime question tool asks OK / Edit items / Drop items / Cancel. Nothing is written and no prompt is printed before OK.
3. prompts: one prompt per surface, in the user's language, each with goal, required behaviour, acceptance checks / regression tests, and scope boundary. If the project has a .feedback/ folder and the user chose to save, records are added in that folder's existing format, numbering and index.

How it ships

Investigated first: devup-mcp had no MCP prompts. /devup reaches clients as an embedded own skill that devup_skills install writes into the workspace skill root (Claude Code offers a project skill as a slash command of the same name), plus a command file for OpenCode. devup-retro uses the same path, and adds a prompt so it works without an install.

Client	How to invoke
Claude Code	/devup-retro [scope or notes] after devup_skills install (.claude/skills/devup-retro/SKILL.md). Without the install: the MCP prompt, /mcp__<server name>__devup-retro
OpenCode	/devup-retro: .opencode/command/devup-retro.md, generated from the skill body at install so the two cannot drift
Codex and other .agents/skills runtimes	the skill loads by name; no slash-command location is verified, and devup_skills says so instead of guessing a path
Any MCP client	prompts/list -> devup-retro (arguments scope, language, both optional); devup://skill/devup-retro; listed in devup_skills status

Server support (minimal, read-only)

devup_skills gets a third action, retro. It writes nothing and returns:

• server: installed version / commit / buildId (same identity every response carries)
• callLog: the devup-mcp calls this server process served, oldest first: tool, action, outputs, a short target (frameIds=...), UTC time, duration, ok. Bounded to 200. It keeps names and times only, never argument contents or results, and it says in the response that a reconnect empties it, so a short log is not read as "nothing was called"
• surfaces, phases: the same list the skill's table has (a test holds them together)
• feedback: whether .feedback/ exists, its index file, nextNumber / nextFilePrefix, the files to read first, and each record's title and status line
• releases: where release notes live and the installed tag, with the rule that an unreadable changelog means unverified, never a guessed "fixed"

The MCP prompts capability is now advertised with one prompt.

Name

devup-retro, defined once: the retro_name! macro in crates/devup-mcp/src/server/skills/commands.rs. The embedded path, the command, the OpenCode file, the prompt and the tool report all derive from it. A rename is that literal, git mv skills/devup-retro skills/<new>, and the name in the skill frontmatter and its manifest.json entry (plus bytes/sha256); the_retro_name_is_spelled_in_one_place fails until they agree. Alternatives considered: devup-review-next, devup-confirm. The tool action stays retro either way.

Other changes

• skills/commands.rs: the /devup command code moved out of installation.rs and generalised to a COMMANDS table. /devup behaviour and command_state output are unchanged; install now reports both commands under commands for OpenCode.
• mod_parts/error_envelopes.rs: to_mcp_error and structured_tool_error moved out of exploration_and_responses.rs, unchanged, to keep that file under 1000 lines after the prompt handlers were added (914 now).
• Bootstrap INSTRUCTIONS are untouched (1490 of 1500 bytes, and another PR is editing them). The pointer is a short section at the end of devup://guide/usage instead.
• README section, changepack (Minor).

Tests

Local (Windows), all green:

• cargo test --workspace --all-features --no-fail-fast: 2300 passed, 0 failed, 2 ignored
• cargo insta test --workspace --all-features --check: 2300 passed, no snapshots to review
• cargo clippy --locked --workspace --all-targets --all-features -- -D warnings: clean; cargo fmt --all -- --check: clean
• node --test ... (CI set): 82 pass, 1 skipped; python packaging/test_release.py: 15 OK

New tests (10):

• skills/retro_tests.rs (8): the name is spelled in one place; the skill states review -> confirm -> prompts with its honesty rules (doc contract); skill table and tool surfaces are identical; prompt arguments and body; call log ordering, bounding and that it keeps no argument contents; UTC timestamps; .feedback/ numbering read without writing; command install per runtime (OpenCode file generated / alreadyPresent / leftAlone, Claude Code via skill, Codex unsupported)
• tests/retro_command.rs (2, real stdio process): prompts capability, prompts/list has devup-retro with optional scope/language, prompts/get returns the three phases, unknown prompt is an error; devup_skills retro returns identity, logged calls (including a refused one as ok: false), feedback state, and leaves .feedback/ untouched; install flips the command to installed
• updated: skills_install.rs own-skill list; manifest well-formedness covers the new entry (bytes/sha256)

Sample run (synthetic, modelled on the Prime animation follow-up)

> /devup-retro

● devup_skills(action: "retro", projectRoot: "apps/prime")
  server 0.21.0 / 5436249 · callLog 14 calls (first: devup_figma_export 2026-11-09T05:12:40Z)
  feedback: .feedback/ exists, index 000-index.txt, nextNumber 34

검토 범위: 2026-11-09T05:12 devup_figma_export(frameIds 33:3550 외 2) 이후의 사용자 메시지 9개.
그중 후속 수정 요청은 4개(항목 3개)입니다. 나머지 5개는 새 요구사항·승인이라 제외했습니다.

| # | 사용자 프롬프트 (원문) | 무엇이 잘못됐나 | surface | 증거 | 상태 |
|---|---|---|---|---|---|
| 1 | "애니메이션을 추가해야 함" / "스크롤할 때 다시 재생되면 안 돼" | 내보낸 랜딩 섹션에 진입 모션이 없었고, 직접 추가한 reveal이 재진입 때 다시 재생됨 | validation-rule | devup_ui_validate(seq 6) errors 0 · apps/prime/src/components/pages/landing/Hero.tsx에 transition/keyframes 없음 | fixed-not-installed (0.22.0 릴리스 노트: entrance-motion rules) |
| 2 | "모바일에서 CTA 줄바꿈이 피그마랑 달라" | 360px CTA에서 '개선을'이 통째로 다음 줄로 넘어감 | codegen | node 33:3550 · devup_visual_compare(seq 11) diff 6.1% | already-recorded (.feedback/016-korean-wrapping-vs-figma.txt) |
| 3 | "CI 통과했다고 했는데 그 run은 이전 커밋이야" | 완료 보고가 HEAD가 아닌 커밋의 run ID를 인용 | skill-guidance | 완료 보고의 run ID ↔ git rev-parse HEAD 불일치 | open |

후속 프롬프트에서 나오지 않은 관찰(프롬프트로 만들지 않음): hero 이미지 alt가 비어 있음.

? 검토 목록          ❯ OK (Recommended) / Edit items / Drop items / Cancel
? .feedback/에 저장   ❯ Yes / No

> OK, Yes

설치된 빌드: devup-mcp 0.21.0 / commit 5436249 / buildId 5436249-…

── skill-guidance ─────────────────────────────────────────────
devup-mcp의 완료 보고 지침에 "CI 증거는 현재 HEAD의 run이어야 한다"를 필수 항목으로 추가해 주세요.

필수 동작
1. devup / devup-implementation-plan 스킬의 완료 조건에 "보고하는 CI run의 head SHA가 `git rev-parse HEAD`와 같아야 한다"를 명시합니다.
2. 다르면 완료로 보고하지 않고, 해당 SHA의 run을 기다리거나 unresolved로 남기도록 합니다.
3. 최종 보고 표에 head SHA와 run ID를 함께 적게 합니다.
근거: 사용자 프롬프트 "CI 통과했다고 했는데 그 run은 이전 커밋이야" · 완료 보고의 run ID와 HEAD 불일치.

수용 기준·회귀 테스트
- 스킬 문서에 위 문구가 없으면 실패하는 skills 단위 테스트를 추가합니다.
- development contract의 required terms에 "head SHA"가 포함됩니다.

범위 경계
- CI 워크플로 자체나 앱의 테스트는 다루지 않습니다(out-of-scope:app).
───────────────────────────────────────────────────────────────

프롬프트가 필요 없는 항목
- #1 fixed-not-installed → devup-mcp를 0.22.0으로 업데이트하고 devup_skills install로 스킬을 갱신하세요.
- #2 already-recorded → .feedback/016-korean-wrapping-vs-figma.txt

저장: .feedback/034-ci-evidence-must-match-head.txt, 000-index.txt에 1줄 추가.
요약: 후속 프롬프트 4개 검토 · 항목 3개 확인 · 프롬프트 1개(skill-guidance) · 프롬프트 불필요 2개 · 파일 2개 기록.

Open items

• The call log covers one server process; a resumed or reconnected session gets a partial log (stated in the response). Persisting it across processes was left out on purpose.
• "Already fixed" is judged by the agent from release notes; the binary carries no changelog.
• Codex has no verified slash-command location, so it gets the skill and the prompt, not a command file.

🤖 Generated with Claude Code

---

Submitted via Devfive hub. Employee: ; device: 3099154426bffbeef8b62999b80bb038