/chief-of-staff:batch

Generate or process a visual HTML batch triage interface.

Usage

/chief-of-staff:batch              # Generate HTML interface (incremental if state exists)
/chief-of-staff:batch --generate   # Same as above
/chief-of-staff:batch --process    # Process downloaded decisions
/chief-of-staff:batch --digest     # Process reading digest decisions (archive/delete summarized emails)
/chief-of-staff:batch --retry      # Retry failed items from last batch
/chief-of-staff:batch --reset      # Clear sync state, full fetch
/chief-of-staff:batch --full       # One-time full fetch, save new state

Arguments

• --generate (default): Create the HTML batch triage page
• --process: Process decisions from downloaded JSON file
• --digest: Process reading digest decisions (archive/delete summarized emails)
• --retry: Retry only failed items from previous batch
• --reset: Clear sync state completely, do full fetch (starts fresh)
• --full: One-time full fetch but save new state for future incremental calls
• --days N: Override lookback period (default: 7)
• --limit N: Override email limit (default: 100)

Workflow

Generate Mode

1. Incremental sync: Fetches only NEW inbox emails since last triage (or all on first run)
   • Uses JMAP queryChanges for Fastmail; falls back to list_emails for other providers
   • --reset clears state and re-fetches everything
   • --full does a one-time full fetch but saves state for next time
2. Classifies into categories (Top of Mind, Deliveries, Newsletters, etc.)
3. Generates standalone HTML file with embedded data
4. Opens in browser for review

HTML Interface Features:

• Emails grouped by category
• Default actions pre-selected
• Dropdown to change action
• Text field for steering notes
• "Submit All" downloads decisions JSON

Process Mode

1. Finds decisions JSON (Downloads folder or scratchpad)
2. Groups decisions by action type
3. Executes inline actions (archive, delete, keep)
4. Delegates batch actions to sub-agents (parcel, newsletter, reminder)
5. Reports summary with success/failure counts

Example Output

Generate Mode

Generated batch triage interface.

Summary:
- Top of Mind: 3 emails
- Deliveries: 5 emails
- Newsletters: 8 emails
- Financial: 4 emails
- Archive Ready: 6 emails
- Delete Candidates: 2 emails
- FYI: 12 emails

Total: 40 emails

Browser opened. Review emails, adjust actions, click 'Submit All'.

After submitting, run: /chief-of-staff:batch --process

Process Mode (with Summarize)

Processing 40 decisions from batch-2026-02-02-abc123...

INLINE ACTIONS
--------------
Archive: 15 emails
  - 8 to Financial
  - 5 to Orders
  - 2 to Travel
Delete: 5 emails
Keep: 2 emails (flagged)

DELEGATED BATCHES
-----------------
Parcel: 5 packages → Added to Parcel
Newsletters: 8 → Unsubscribed from 6, 2 failed
Reminders: 3 → Created in Apple Reminders
Calendar: 2 → Created events

SUMMARY
-------
Total: 40 decisions
Successful: 38
Failed: 2

Failed items saved for retry.

Digest Mode

Processing 5 digest decisions...

DIGEST PROCESSING
-----------------
Archived: 3 emails (unflagged)
Deleted: 2 emails
Remaining in digest: 0 items

Decisions recorded as summarize_archive/summarize_delete.

Implementation

CRITICAL: This command MUST delegate to sub-agents via the Task tool. Do NOT attempt to perform actions directly in this command.

Generate Mode (default or --generate):

Use the Task tool with:
  subagent_type: "chief-of-staff:batch-html-generator"
  model: sonnet
  prompt: |
    Generate the HTML batch triage interface.
    Parameters:
    - Look back: [days] days (default: 7)
    - Limit: [N] emails (default: 100)
    - Sync mode: [--reset | --full | (default: incremental if state exists)]

    If --reset was passed: Clear sync-state.yaml before fetching. Full re-fetch.
    If --full was passed: Do full fetch this time, save new state for future incremental.
    Otherwise: Use incremental sync if sync-state.yaml has query_state.

    IMPORTANT: If email MCP fails, report the error immediately.
    Do NOT generate sample data or placeholder HTML.

Process Mode (--process):

Use the Task tool with:
  subagent_type: "chief-of-staff:batch-processor"
  model: sonnet
  prompt: |
    Process batch triage decisions from the downloaded JSON file.

    Check these locations for the decisions file:
    1. ~/Downloads/inbox-triage-decisions-*.json
    2. Scratchpad directory

    IMPORTANT:
    - Check steering text for reply intent even if action is not "reply"
    - Use reply_to_email with markdownBody for drafts
    - Initialize data files from .example templates if missing
    - Record all decisions to decision-history.yaml for learning
    - CRITICAL: If any decisions have action "summarize", you MUST launch
      the reading-digest-generator sub-agent AFTER writing reading-digest-state.yaml.
      This generates the newspaper-style HTML digest and opens it in the browser.
      Do NOT skip this step — the user expects the digest to appear automatically.

Digest Mode (--digest):

Use the Task tool with:
  subagent_type: "chief-of-staff:batch-processor"
  model: sonnet
  prompt: |
    Process reading digest decisions from the downloaded JSON file.

    Check these locations for the decisions file:
    1. ~/Downloads/reading-digest-decisions-*.json
    2. Scratchpad directory

    The decisions JSON will have type: "digest" with archive/delete actions per email.
    For each decision:
    - Archive: move to Archive folder, unflag the email
    - Delete: delete the email
    After processing, update reading-digest-state.yaml (remove processed items, add to history).
    Record decisions as summarize_archive / summarize_delete in decision-history.yaml.

Retry Mode (--retry):

Use the Task tool with:
  subagent_type: "chief-of-staff:batch-processor"
  model: sonnet
  prompt: |
    Retry failed items from the previous batch.
    Read failures from ~/.claude/data/chief-of-staff/batch-state.yaml and reprocess only those items.

Why Task tool is required:

• Sub-agents have specialized tools and prompts
• Direct execution would miss important context (patterns, rules, MCP setup)
• Learning system expects consistent agent behavior