the-agency-ai/.claude/skills/flag-triage

<!-- Flag #62/#63: allowed-tools removed. Inherits Bash(*) from .claude/settings.json. Restricting to specific subcommand patterns at the skill level silently blocks agents on permission prompts the agent cannot see — see dispatch #171 for the devex incident that surfaced this trap. -->

Flag Triage

Structured review session for accumulated flags. Agent pre-categorizes into three buckets, principal approves dispositions, then each bucket is processed.

Arguments

• $ARGUMENTS: Optional. No arguments needed — operates on all unprocessed flags.

Instructions

Step 1: Gather flags

Run bash $CLAUDE_PROJECT_DIR/claude/tools/flag list to get all unread/read flags.

If no flags are waiting, report "No flags to triage" and stop.

Step 2: Categorize into three buckets

Review each flag and assign it to one of three buckets. Present the categorization to the principal as a numbered table:

Bucket 1 — Resolved: Items that are already done, no longer relevant, or overtaken by events. For each, cite the evidence (commit, dispatch, code change) that resolves it.

Bucket 2 — Autonomous: Items the agent can handle independently — writing seeds, sending dispatches, updating docs, filing bugs, creating tools. No principal collaboration needed. For each, state the proposed disposition (what you'll do).

Bucket 3 — Collaborative: Items requiring principal input, joint decision-making, or 1B1 discussion. For each, state why it needs discussion.

Format:

## Flag Triage — [N] items

### Bucket 1: Resolved ([n] items)
| # | Flag | Evidence | Disposition |
|---|------|----------|-------------|
| 1 | ... | commit abc123 fixed this | Mark resolved |

### Bucket 2: Autonomous ([n] items)
| # | Flag | Proposed Action |
|---|------|----------------|
| 4 | ... | Write seed to iscp workstream |

### Bucket 3: Collaborative ([n] items)
| # | Flag | Why discuss? |
|---|------|-------------|
| 7 | ... | Needs prioritization decision |

Step 3: Principal review

STOP and wait for principal approval. The principal may:

• Approve all bucket assignments
• Move items between buckets
• Add context to specific items
• Reject categorizations

Do NOT proceed until the principal approves the bucket assignments.

Step 4: Process Bucket 1 (Resolved)

For each approved Bucket 1 item:

• Run bash $CLAUDE_PROJECT_DIR/claude/tools/flag clear if all flags are being resolved
• Report: "[n] items marked resolved"

Step 5: Process Bucket 2 (Autonomous)

For each approved Bucket 2 item, execute the proposed disposition:

• Write seed: Create seed file in the appropriate workstream's seeds/ directory
• Send dispatch: Use dispatch create --to <agent> --subject <text> --body <text>
• Update docs: Edit the relevant file directly
• File bug: Use /agency-bug or create a dispatch
• Create tool: Write to usr/{principal}/{agent}/tools/

Report each action taken with the artifact created (file path, dispatch ID, etc.).

Step 6: Process Bucket 3 (Collaborative)

Enter 1B1 discussion mode for collaborative items. For each item:

1. Present the item with context
2. Three possible dispositions per item:
   • Action now — do the work in this session
   • Flag to project — create a seed, dispatch, or backlog item for later
   • Bin — discard as not worth pursuing
3. Execute the chosen disposition before moving to the next item

Use the standard 1B1 protocol: present → feedback → resolve → next.

Step 7: Summary

After all buckets are processed, report:

## Triage Complete

- Resolved: [n] items
- Autonomous: [n] items ([n] dispatches sent, [n] seeds written, ...)
- Collaborative: [n] items ([n] actioned, [n] flagged to project, [n] binned)
- Total: [N] flags processed

Run bash $CLAUDE_PROJECT_DIR/claude/tools/flag clear to mark all processed flags.

Key Rules

• All bucket assignments require principal approval — never auto-process
• Bucket 2 items must produce artifacts — dispatches, seeds, PRs, docs. Not just "noted"
• Bucket 3 uses 1B1 protocol — one item at a time, resolve before moving on
• Evidence required for Bucket 1 — don't mark resolved without citing proof

OFFENDERS WILL BE FED TO THE — CUTE — ATTACK KITTENS!