fleet-triage - SKILL.md Agent Skill

name: fleet-triage description: Cognitive triage of fleet audit findings. Read issue content, evaluate labeling accuracy, and determine open/close/dispatch/relabel actions for fleet-managed issues. Use when triaging undispatched issues or reviewing audit scan results. allowed-tools: run_command(gh:) run_command(fleet:) view_file write_to_file read_url_content

Fleet Triage

Cognitive triage of fleet-managed issues. The audit scan surfaces findings deterministically; you make the judgment calls.

When to Use

After running fleet audit scan and seeing cognitive findings
Before running fleet dispatch to prevent wasting sessions on non-actionable issues
When asked to triage, clean up, or review fleet issues

Core Principle

Tools are evidence, you are expertise. The scan tells you what exists. You read the content and decide what to do.

Process

Phase 1: Build a Lightweight Index

Pull all open issues as a summary table. Do NOT read issue bodies yet.

gh issue list --repo <owner>/<repo> --state open --limit 50 \
  --json number,title,labels,milestone

Classify each issue into buckets by title and labels alone:

Bucket	Signal	Typical Action
Insight	Title contains `[Insight]`, `[Fleet Insight]`, or label `fleet-insight`	Likely close — informational only
Assessment	Title contains `[Assessment]`, or label `fleet-assessment`	Evaluate — may be actionable or stale
Execution	Title contains `[Fleet Execution]`, label `fleet` without insight/assessment	Likely dispatch — has code work
Ambiguous	Mixed signals or unlabeled	Needs deep dive (Phase 3)

Record the index in a triage artifact (see Triage Artifact).

Phase 2: Batch Triage from Index

For Insight and Assessment buckets, you can often decide without reading the body:

Duplicate insights (same title pattern, sequential numbers like "Update 2", "Update 3", "Update 4") → Close older duplicates, keep latest
Insights with no milestone → Likely orphaned, close
Assessments with linked PRs → Check if PR is merged; if so, close

Update the triage artifact with decisions.

Phase 3: Deep Dive (One at a Time)

For Ambiguous or Execution issues only:

gh issue view <N> --repo <owner>/<repo>

Read the body. Evaluate:

Is there actionable code work? Look for "Files to modify:", "Proposed Implementation", specific code paths
Is it stale? Has the work already been done by another PR? Check linked PRs
Are the labels accurate? Does an "Execution" issue actually describe an insight?
Should it be dispatched? Only if there's concrete code work with clear acceptance criteria

Record your decision in the triage artifact before moving to the next issue.

Phase 4: Apply Decisions

After all decisions are recorded, present the triage artifact to the user for review. Then apply:

# Close issues
gh issue close <N> --repo <owner>/<repo> --comment "Closing: <reason>"

# Relabel
gh issue edit <N> --repo <owner>/<repo> --remove-label fleet --add-label fleet-insight

# Dispatch (only confirmed actionable issues)
fleet dispatch --owner <owner> --repo <repo>

Triage Artifact

Create a persistent artifact at triage-<repo>.md:

# Triage: <owner>/<repo>

## Summary
- Total open: N
- Reviewed: N
- Close: N | Keep: N | Dispatch: N | Relabel: N

## Decisions

| # | Title | Labels | Decision | Reason |
|---|-------|--------|----------|--------|
| 194 | [Fleet Insight] Coverage | fleet, fleet-insight | CLOSE | Insight, no code action |
| 141 | [Fleet Execution] Update Enums | fleet, fleet-assessment | DEEP DIVE | Mixed labels, need to read body |

Token Management

Never load all 25 issue bodies at once — each body can be 500-2000 tokens
Process one deep dive at a time, record decision, then move on
The triage artifact survives context resets — pick up where you left off
For batch closes, use gh issue close in a single command per issue

Common Patterns

Duplicate Detection

Fleet analyzers often create duplicate insights across runs. Look for:

Sequential "Update N" suffixes (keep only the latest)
Same title with different numbers
Same objective described in different words

Stale Assessment Detection

An assessment may have been addressed by a PR that didn't reference the issue:

Check if the described code changes already exist in main
Check if a similar PR was merged without Fixes #N

Mislabeled Issues

The analyzer sometimes labels insights as fleet (execution-worthy) when they're informational:

Body says "No source changes" → should be fleet-insight, not fleet
Body says "N/A - This is an Insight report" → not dispatchable