类别 4：输入/输出契约

### Output
# Pitch Review — CodePilot

**Format:** 5 slides, ~3 minutes | **Overall:** 16/25 (Adequate)

## Scorecard

| Criterion | Score | Feedback |
|-----------|-------|----------|
| Hook | 2/5 | Title slide alone is not a hook. Start with a pain point: "Your PR has been waiting 3 days. Sound familiar?" |
| Problem / Solution | 4/5 | Problem is clear and quantified. Solution is specific. Add one metric: "reduces review time by 60%." |
| Demo Flow | 4/5 | Live demo in VS Code is strong. Add a before/after comparison to make impact visible. |
| Technical Clarity | 4/5 | Architecture diagram is clear. Mention latency and cost briefly for judge credibility. |
| Call to Action | 2/5 | "Thank you" is not a CTA. End with: "Try it today — install from VS Code marketplace." |

## Suggested Improved Outline

| Slide | Content | Time |
|-------|---------|------|
| 1 | **Hook:** "Your PR has been waiting 3 days." + problem stat | 0:00–0:30 |
| 2 | **Solution:** CodePilot overview + key metric (60% faster reviews) | 0:30–1:00 |
| 3 | **Demo:** Live VS Code refactor suggestion → PR comment | 1:00–2:00 |
| 4 | **How It Works:** Architecture + cost/latency | 2:00–2:30 |
| 5 | **CTA:** "Install now" or "Join our beta" + team slide | 2:30–3:00 |

Complete — pitch scored 16/25, 5 actionable improvements provided.

Output the results in a structured way. Include the scores and feedback.
Make it clear and easy to understand. Use a table if appropriate.

### Timeout & Error Handling
Every Bash command specifies a timeout. On timeout:
1. Retry ONCE with the same timeout
2. If retry fails, apply graceful degradation:
   - **Phase 1 (metadata) timeout/failure:** STOP — no manifest, cannot proceed
   - **Phase 2 (collect) timeout/failure:** Check if manifest was partially written.
     If `changed_files` present → re-run Phase 3 with `--api-only`. If missing → STOP
   - **Phase 3 (finalize) timeout/failure:** Check if partial diffs exist.
     If some → proceed with available diffs. If none → STOP
   - **Agent timeout/failure:** Log "Agent {name} did not complete" and continue with
     remaining agents' results. If ALL 5 agents fail → STOP
   - **Posting (Step 8.1) timeout:** Report partial results and offer retry

Common error patterns:
- `az CLI not found` → tell user to run `az login`
- HTTP 401/403 (ADO) → token expired, re-run `az login`
- HTTP 404 → PR not found, verify URL
- `No manifest found` → Phase 1 did not complete, re-run from Phase 1

### Partial Result Handling
- Missing diff for a file → skip that file, note in output
- Missing local context → proceed without (review still works, just shallower)
- Agent returns no findings → treat as "no issues found" (not an error)
- Agent returns malformed output → attempt salvage (infer severity from keywords)

If something goes wrong, try to recover gracefully. Report any errors to the user.
If a tool fails, try an alternative approach.

## First-Time Setup

On every invocation, check whether saved configuration exists before doing anything else.

### Step 1: Load Saved Configuration

Check for the config file. Use `$HOME` which works across all platforms:

    cat "$HOME/.config/ado-flow/config.json" 2>/dev/null

If the file exists and contains valid JSON with all required fields (`organization`,
`work_item_project`, `pr_project`), skip ahead to the relevant task workflow.

If the file does not exist or is missing fields, proceed with setup steps below.

### Step 2: Check Prerequisites

Verify Azure CLI and the DevOps extension are installed:

    az version 2>/dev/null && az extension show --name azure-devops 2>/dev/null

If Azure CLI is missing, display install instructions per platform.

### Step 3: Collect Configuration

Ask the user for these details one at a time:

1. **Organization name** - "What is your Azure DevOps organization name?"
2. **Project for work items** - "What project should I create work items in?"
3. **Project for pull requests** - "What project are your repositories in?"

### Step 4: Save Configuration

    mkdir -p "$HOME/.config/ado-flow"
    cat > "$HOME/.config/ado-flow/config.json" <<EOF
    {
      "organization": "{ORG}",
      "work_item_project": "{WORK_ITEM_PROJECT}",
      "pr_project": "{PR_PROJECT}"
    }
    EOF

Confirm: "All set! Your configuration has been saved."

If this is the user's first time, ask them for their Azure DevOps settings.
Save the settings somewhere for next time.

## Platform Support

| Platform | TTS Engine | Focus Detection |
|----------|-----------|-----------------|
| Windows | OneCore (Mark) with SAPI fallback | Windows Terminal tab inspection (UI Automation) |
| macOS | `say` command (system voices) | `osascript` frontmost app + window title |
| Linux | espeak-ng / spd-say / festival / notify-send | `xdotool` (X11), Wayland stub |

## Cache Location

- **Windows:** `$env:TEMP\claude_voice\{session-id}.txt`
- **macOS/Linux:** `${TMPDIR:-/tmp}/claude_voice/{session-id}.txt`

Play a notification sound when done. Use whatever audio tool is available on the system.