post-dev-recap

name	post-dev-recap
description	Post-development recap wrapper. Use when: AI/Codex just finished implementing a feature and the user wants a guided walkthrough with scope detection + doc generation + follow-up Q&A. Not for: generating only a doc (use /recap-doc), Q&A over an existing recap (use /recap-ask), technical share-out (use /tech-brief), first-principles reasoning of an existing doc (use /fp-brief), general code tracing (use /code-explore). Output: ScopeReport → briefing-recap-<date>.md → interactive Q&A session.
allowed-tools	Read, Grep, Glob, Write, Bash(node:*), Bash(git:*), Skill, AskUserQuestion, mcp__codex__codex, mcp__codex__codex-reply

/post-dev-recap — Guided Post-Development Recap

Trigger

• Keywords: post-dev-recap, 開發後導覽, walkthrough after dev, recap workflow, 本輪回顧

When NOT to Use

Scenario	Alternative
Only need a doc (no Q&A)	/recap-doc — emits the recap markdown and stops
Have an existing recap, want to ask questions	/recap-ask --context <path>
Share-out for other developers (not the author)	/tech-brief — focuses on knowledge transfer to strangers
First-principles reasoning of an existing doc	/fp-brief — explains why decisions hold
General project Q&A, no recent change	/ask — reopens the whole repo as context
Deep code tracing across modules	/code-explore — execution-path walker

Positioning vs /ask, /tech-brief, /fp-brief, /code-explore (AS-5)

Dimension	/post-dev-recap	/ask	/tech-brief	/fp-brief	/code-explore
Trigger	Right after AI/Codex finishes a feature	Any quick question	Share-out for teammates	Explain an existing doc	Diagnose unknown code
Input	ScopeReport (detected live)	Free-form question	A finished feature	A document	A code question
Output	Recap doc + Q&A session	Answer	Brief artifact	First-principles chain	Execution-path report
Bounded?	Yes — scope of the just-done change	No — whole repo	No — whole feature	Document-scoped	Free search

Command Signature

/post-dev-recap [<focus>] [--interactive] [--depth brief|normal|deep]

Flag	Default	Description
<focus>	""	Natural-language keyword to bias section emphasis, passed to /recap-doc --focus
--interactive	false	Enable per-step AskUserQuestion — user can choose one of 4 actions between phases: 繼續 (continue) / 提問 (ask) / 跳段 (skip) / 結束 (end)
--depth	normal	Recap depth — brief / normal / deep, passed to /recap-doc --depth

FR-4 Must — Q&A is mandatory. This wrapper has no --no-qa flag. Callers that only want the markdown must invoke /recap-doc directly. See the Q&A Gate (FR-4 Must) section below for the design rationale.

Prohibited Actions (AS-6)

❌ git add | git commit | git push | git stash | git reset — all git mutations are forbidden

This wrapper reads git state (git status, git diff, git log) for scope detection only. It never creates commits, pushes, or alters the working tree. See @rules/git-workflow.md.

Workflow

sequenceDiagram
    participant U as User
    participant W as /post-dev-recap
    participant S as scripts/detect-scope.js
    participant D as /recap-doc
    participant A as /recap-ask

    U->>W: /post-dev-recap [<focus>] [--interactive] [--depth ...]
    W->>W: Phase 0: Parse args + resolve CWD = git repo root
    W->>S: Phase 1: node scripts/detect-scope.js (no args)
    alt source === null (scope detection failed)
        S-->>W: ScopeReport with source: null + fallback_trace[]
        W-->>U: ⚠️ Need Human + fallback_trace + suggest manual /recap-doc --scope
        Note over W: exit non-zero — do NOT invoke sub-skills
    else source resolved
        S-->>W: ScopeReport v1 JSON
        opt --interactive
            W->>U: AskUserQuestion — "Scope detected (<N> files). 繼續 (continue) / 提問 (ask) / 跳段 (skip) / 結束 (end)?"
            U-->>W: 繼續 (continue) | 提問 (ask) | 跳段 (skip) | 結束 (end)
            Note over W: ask=prompt the user to phrase a clarifying question about the scope, echo answer, then re-prompt. skip=skip /recap-doc, jump to Phase 3 with user-supplied recap path. end=stop.
        end
        W->>D: Phase 2: /recap-doc --scope <json> [--focus <focus>] [--depth <d>]
        D-->>W: briefing-recap-<date>.md path (ephemeral default = <tmp>/sd0x-dev-flow-recap/)
        opt --interactive
            W->>U: AskUserQuestion — "Recap ready at <path>. 繼續 (continue) / 提問 (ask) / 跳段 (skip-qa-end) / 結束 (end)?"
            U-->>W: 繼續 (continue) | 提問 (ask) | 跳段 (skip-qa-end) | 結束 (end)
            Note over W: "skip-qa-end" is the only documented FR-4 bypass — explicit user consent, logged.
        end
        W->>A: Phase 3: /recap-ask --context <recap-path>  (FR-4 Must — non-interactive flow)
        loop Q&A turns (until user ends)
            U->>A: <question>
            A-->>U: answer + Sources + Thread
        end
        A-->>W: Thread id + promote outcome
        W-->>U: Summary: recap path, thread id, promote status
    end

Phase 0 — Arg Parsing

1. Treat the first positional non-flag argument as <focus>; subsequent positional args are a warning (extra args ignored with stderr note).
2. Resolve CWD to the git repo root via git rev-parse --show-toplevel. If the command is not run inside a git repo, emit ⚠️ Need Human — /post-dev-recap must run inside a git repository. and stop.
3. Validate --depth ∈ {brief, normal, deep}. Reject unknown values.

Phase 1 — Scope Detection (T1 reuse)

1. Invoke node scripts/detect-scope.js with no additional args (it auto-detects source: uncommitted → branch → session → null).

2. Parse stdout as JSON (ScopeReport v1). Validate version === 1.

3. Scope-fail path (AS-6 exit): if source === null OR files.length === 0, emit:

   • ⚠️ Need Human — scope detection inconclusive
   • The full fallback_trace array (why each heuristic failed)
   • Suggested manual command: /recap-doc --scope <explicit-json>
   • Exit before invoking Phase 2. 3b. scan_error gate: feature_context.scan_error !== false ⇒ the source sets behind the scope are unknown, not empty — report it and take the ⚠️ Need Human exit rather than recapping a change whose feature corpus could not be enumerated. detect-scope.js projects the field onto feature_context for exactly this check. Gate on !== false, not === true: a payload from a shell fallback or an older producer carries no such field at all, and a non-null key is not evidence the sets are complete — scan_error rides alongside a resolved key.

4. Interactive hook (opt-in): when --interactive is set and scope resolved, call AskUserQuestion with the scope summary (<N> files, source=<s>, confidence=<c>) and the 4-option set {繼續 (continue), 提問 (ask), 跳段 (skip), 結束 (end)}. ask = prompt the user for a clarifying question about the detected scope, echo the answer, then re-prompt the 4-option set. skip = skip /recap-doc and jump to Phase 3 (requires the user to supply an existing recap path). end = stop cleanly without invoking sub-skills.

Phase 2 — Recap Doc (T2 reuse)

1. Serialise the ScopeReport to a temp file (or pipe via stdin) and invoke /recap-doc --scope <path> [--focus <focus>] [--depth <depth>] via the Skill tool.
2. Capture the output recap path. The default is ephemeral under <tmp>/sd0x-dev-flow-recap/briefing-recap-<YYYY-MM-DD>.md (see @skills/recap-doc/SKILL.md ## Save Behavior).
3. On /recap-doc failure, surface its stderr verbatim and exit — do not silently fall through.
4. Interactive hook: when --interactive, call AskUserQuestion with the recap path and the 4-option set {繼續 (continue), 提問 (ask), 跳段 (skip-qa-end), 結束 (end)}. ask = prompt the user for a clarifying question about the recap doc itself (not a Q&A turn — that belongs in Phase 3), echo the answer, re-prompt. skip-qa-end is the only documented FR-4 bypass — explicit user consent, flagged in session log. end closes the session immediately after the doc is written.

Phase 3 — Q&A Gate (FR-4 Must)

1. Unless the user already chose skip-qa-end in Phase 2's interactive prompt (the only documented FR-4 bypass), collect the user's first question before invoking /recap-ask — the sub-skill's command signature requires <question> as a positional argument (see @skills/recap-ask/SKILL.md ## Command Signature). Use AskUserQuestion with a free-text prompt: 「針對本輪 recap 想問的第一個問題？」. This is a required data-collection step (not a bypass mechanism): a non-empty answer is mandatory. If the user submits an empty answer, re-prompt once with clarification; if they still decline, emit the canonical bypass reminder — 「依 FR-4，非互動模式無法跳過 Q&A；若需跳過請改用 --interactive + 選 skip-qa-end。」 — and re-prompt. Phase 3 does not proceed without a question. The "No bypass" cell at line 133 below remains authoritative for non-interactive runs.
2. Invoke /recap-ask "<first-question>" --context <recap-path> via the Skill tool using the collected question.
3. The Q&A session continues with /recap-ask --continue <threadId> turns until the user types 結束 / done or invokes /recap-ask --end (see @skills/recap-ask/SKILL.md Phase 4). The wrapper does not add extra termination signals — it forwards them as-is.
4. On session end, relay the thread id + promote outcome back to the caller so --continue is reusable in future sessions.

Q&A Gate (FR-4 Must)

Design rationale: a "recap" without follow-up Q&A is just a doc. The feature's value is the interactive exploration. Callers who truly only want the markdown must use /recap-doc directly. The wrapper intentionally has no --no-qa flag.

Mode	Bypass path
Non-interactive (default)	No bypass — Phase 3 always runs
--interactive	skip-qa-end option in Phase 2 AskUserQuestion (explicit consent, logged)

Performance

Target: NFR-2 applies to Phase 2 (/recap-doc ≤ 30s); Phase 1 scope detection and Phase 3 per-turn Q&A inherit from the respective sub-skills' targets. Wrapper overhead itself must be negligible (≤ 1s parsing + dispatch).

Path Security

Rule	Implementation
Repo boundary	CWD is resolved via git rev-parse --show-toplevel; wrapper refuses to run outside a repo
Sub-skill pass-through	/recap-doc and /recap-ask enforce their own path-security rules (see those files)
No project writes	Wrapper never writes inside the user's project tree. Phase 2 may serialise the ScopeReport to a short-lived temp file under the OS temp dir ($TMPDIR / os.tmpdir() / /tmp) purely as input to /recap-doc; the recap markdown itself is owned by /recap-doc (ephemeral default under <tmp>/sd0x-dev-flow-recap/).
Secret redaction	Inherited from /recap-doc Phase 5 (scripts/security-redact.js) and /recap-ask Phase 3 (same script)
Git write operations	Forbidden — see ## Prohibited Actions

Output Format

## Post-Dev Recap

**Scope**: <N> files, source=<source>, confidence=<high|medium|low>
**Recap**: `<path>` (ephemeral | committed)
**Q&A**: <N> turns | thread=<codex-threadId> | promote=<yes|no|n/a>

### Summary

- Key file: `<path>:<line>` — <one-line intent>
- Design decision: <one-liner>
- Blind spot flagged: <heuristic name>

### Next steps

- Resume Q&A: `/recap-ask "<follow-up>" --context <path> --continue <threadId>`
- Promote to ticket: already done | `/recap-ask --end` + confirm

Scope-fail output (⚠️ Need Human)

## ⚠️ Need Human — scope detection inconclusive

**Fallback trace**:
- uncommitted: <N> changes (if <0 or >threshold, fail reason)
- branch: <ahead-of-base>, <reason>
- session: <transcript-entries>, <reason>

**Suggested manual path**:
- Inspect `git status` and pick a scope manually
- Invoke `/recap-doc --scope '{ version: 1, source: ..., files: [...] }'` with explicit JSON

Verification

• First positional arg captured as <focus>; extras ignored with warning
• --depth rejected for unknown values (only brief/normal/deep accepted)
• CWD enforced via git rev-parse --show-toplevel; refuses outside repo
• Phase 1 emits ⚠️ Need Human + fallback_trace when source === null OR files.length === 0
• Phase 2 passes --focus and --depth through to /recap-doc unchanged
• Phase 3 always runs in non-interactive mode (FR-4 Must); only interactive skip-qa-end can bypass
• --interactive triggers AskUserQuestion between Phase 1↔2 and Phase 2↔3
• No git mutations (git add/commit/push/stash/reset) are invoked (AS-6)
• When NOT to Use contrasts /ask / /tech-brief / /fp-brief / /code-explore (AS-5)
• Pass /codex-review-fast

References

• @skills/recap-doc/SKILL.md — Phase 2 sub-skill (T2)
• @skills/recap-ask/SKILL.md — Phase 3 sub-skill (T3)
• scripts/detect-scope.js — Phase 1 ScopeReport producer (T1)
• @skills/ask/SKILL.md — Pattern comparison only (not reused here; /recap-ask reuses its Phase 2)
• @rules/git-workflow.md — Git mutation ban (AS-6)
• @rules/auto-loop.md — Post-execution review loop
• Tech Spec §3.3.1 Signature; §3.4.4 Interactive Mode
• Requirements FR-4 (Must), FR-5, FR-6, AS-1, AS-2, AS-4, AS-6, AS-7

Examples

Input: /post-dev-recap
Action: Phase 0 parse → Phase 1 detect-scope → source=uncommitted, 12 files → Phase 2 /recap-doc → ephemeral recap at <tmp>/sd0x-dev-flow-recap/briefing-recap-2026-04-17.md → Phase 3 /recap-ask --context <path> → interactive Q&A loop until user ends

Input: /post-dev-recap "auth middleware" --depth deep
Action: pass `--focus "auth middleware" --depth deep` to /recap-doc → deep recap with inline snippets → mandatory Q&A

Input: /post-dev-recap --interactive
Action: AskUserQuestion after scope detection (繼續/提問/跳段/結束) → after doc (繼續/提問/跳段-skip-qa-end/結束) → Q&A unless user chose skip-qa-end

Input: /post-dev-recap when no uncommitted/branch/session scope available
Action: Phase 1 returns source=null + fallback_trace → emit ⚠️ Need Human + trace → exit non-zero, no sub-skill invocation