fleet-methodology-compliance

How fleet-ops detects methodology violations during review — code during analysis, skipped stages, readiness jumps, missing contributions. The quality enforcement backbone.

SKILL.md

GitHub

Methodology Compliance — Fleet-Ops Quality Backbone

Your review is the last gate before "done." The methodology exists to ensure agents think before they build. Your job is to verify they followed it — not just that the code works, but that the PROCESS was sound.

What You're Checking (5 Compliance Dimensions)

1. Stage Traversal

Every task should progress through stages in order. A task that jumps from conversation (readiness 10) to work (readiness 99) without analysis or reasoning skipped the thinking.

How to check: Use acct_trail_reconstruction(task_id) or read board memory trail events:

  • Trail should show: stage transitions (conversation → analysis → ... → work)
  • Trail should show: artifacts at each stage (analysis_document, investigation_document, plan)
  • Red flag: only plan_accepted and task_completed events — no analysis, no investigation

Acceptable exceptions:

  • Subtasks and fix tasks may start at work stage (PM triaged them correctly)
  • Spikes and concerns may only go through analysis/investigation (no work stage)

2. Tool Usage at Correct Stages

The methodology blocks certain tools at certain stages (methodology.yaml). The MCP layer enforces this at call time. But defense-in-depth means you should verify the trail doesn't show violations:

Stage Blocked Allowed
conversation fleet_commit, fleet_task_complete, fleet_task_accept fleet_chat, fleet_artifact_create (drafts)
analysis fleet_task_complete fleet_commit (analysis docs OK), fleet_artifact_create/update
investigation fleet_task_complete fleet_commit (research docs OK), fleet_artifact_create/update
reasoning fleet_task_complete fleet_task_accept, fleet_task_create, fleet_commit (plans OK)
work (nothing blocked) ALL tools available

Red flag: Trail shows fleet_commit events during conversation stage. The MCP layer should have blocked this, but if it somehow got through, flag it.

3. Contribution Consumption

For stories and epics, the synergy matrix requires contributions. Check that:

  • Required contributions were received (typed comments on the task)
  • The implementation actually FOLLOWS the contributions (design_input pattern used, TC-XXX criteria satisfied)
  • Security requirements were followed, not ignored

How to check: Read the task's typed comments. Look for **Contribution (design_input)**, **Contribution (qa_test_definition)**, etc. Then compare the implementation/PR against what was contributed.

Red flag: design_input says "adapter pattern" but implementation uses direct API calls. QA defined TC-003 testing timeout handling but no timeout test exists.

4. Phase Standards

Each delivery phase (poc/mvp/staging/production) has quality bars from config/phases.yaml. The phase_advance tool checks these, but you should verify during review:

Phase Tests Docs Security
poc Happy path only README Basic scan
mvp Main flows + edges Setup + usage Auth + validation
staging Comprehensive + integration Full docs Dep audit
production Complete + performance Everything + runbook Certified

Red flag: Production-phase task with only happy-path tests. MVP task with zero documentation.

5. Trail Completeness

A complete trail tells the story of the task. Minimum trail events for a standard task:

  1. plan_accepted — agent committed to an approach
  2. Stage transition events (if applicable)
  3. fleet_commit events — the work itself
  4. task_completed — agent submitted for review
  5. Contribution events (if story/epic)

Red flag: Only task_completed in the trail. No plan, no commits, no stage transitions — the agent may have done everything in one untracked burst.

The Review Decision

After checking all 5 dimensions:

  • ALL PASS → fleet_approve(decision="approved", comment="[specifics of what passed]")
  • MINOR GAPS → Approve with feedback: comment="Approved. Note: trail could be richer — consider recording stage transitions."
  • SIGNIFICANT GAPS → fleet_approve(decision="rejected", comment="[specific gap]: design_input says X but implementation does Y. Reground in verbatim and fix.")

Never rubber-stamp. Never approve without reading the work. The PO trusts your review — make it real.