Imported from SavinRazvan/agent-colony (skills/drift-audit/SKILL.md). Install upstream with npx skills add SavinRazvan/agent-colony --skill drift-audit. Copyright stays with the author.

Drift audit

Use ASD-STE100: .ai_infra/docs/operations/asd-ste100-prose.md

Goal

Detect operational workflow drift and a falsifiable goal/doctrine pulse:

• plan ↔ tracker ↔ session-pointer incoherence
• DRIFT-004b session Board vs export; DRIFT-009 dual-write; DRIFT-010 board vs PRs
• DRIFT-011 .cursor/agents basenames == eight live kit agent ids
• DRIFT-011b (consumer) extra integrator agents — advisory only
• DRIFT-012 .local/plans/ snapshot-only under board_only (no live/current plan SSOT in that dir)
• DRIFT-013 git index must not track .local/, secrets, or .venv/ (consumer + kit-dev)
• DRIFT-014 token-efficiency anchor in agent cards (profile-aware)
• DRIFT-015 (kit-dev) plugin+workspace rule basename dup — WARN
• DRIFT-016 thin-index § headings for skills on disk (profile-aware skip)
• DRIFT-017 (kit-dev) audit independence WARN when Schema-1 Commissioned-By is missing/placeholder or equals Audited-By
• Prose goal pulse: board Acceptance/Notes vs plan pointers vs AGENTS.md / agent cards (flag gaps; hand off — do not rewrite architecture)

Does not replace auditor (CHK-* scorecard) or verifier.

When

• Substantive implementer slice closure (recommended)
• After plans / board Acceptance / agent doctrine docs change
• Optional pre-review drift pass before PR workflow
• When project_ssot.enabled — every pass should include board Status evidence

Entry checklist (goal pulse)

1. Board (when enabled): project api-ready then project entry (or project status + scoped list --status in_progress) — read Acceptance / Notes on In progress cards.
2. Plan pointers: .local/index-and-planning/current/plan.md (or board card body as SSOT) — Current focus / goals.
3. Doctrine: AGENTS.md skills/agents table vs .cursor/agents/*.md (script: DRIFT-011).
4. Docs freshness vs goals: note Probable if AGENTS.md / IMPLEMENTATION-STATUS look stale relative to Current focus (prose only).

Steps

1. Board first (when enabled): python3 -m agent_colony project api-ready then project entry (or project status + project list --status in_progress) — cite board Status in artifacts. Optionally refresh the read-only snapshot: python3 -m agent_colony project export (never writes Status).
2. Script: python3 -m agent_colony drift validate --directory . (or make drift-validate). On consumer app projects, use --profile consumer. Include DRIFT-004b / DRIFT-009 / DRIFT-010 / DRIFT-011 / DRIFT-012 / DRIFT-013 when kit-dev / board_only as applicable (ADR-007/008/010). Consumers always get DRIFT-013 + DRIFT-011b on consumer / consumer-board profiles.
3. Capture profile, check IDs, severities, and details from output.
4. Add prose Goal pulse section in drift-audit.md (board/plan/AGENTS gaps). Fuzzy “vision mismatch” stays Probable — not CI.
5. Write artifacts under .local/workflow-artifacts/drift/ only.
6. Board Exit: Shippable drift-pass (PR citation / [AUDIT] / P0|P1) → handoff --next verifier --to in_review before Done (CLI EXIT_VALIDATION without hop / allow-skip). Hygiene chores may → done with --agent drift-guard. For Confirmed dual-write or roster gaps, Notes on offending card or Ready handoff to board/implementer — do not auto-edit plan.md, work-tracker.md, or session-pointer.md.
7. Print handoff line with item_id when applicable.

Evidence contract

Universal canon: evidence-first skill · evidence-first.md. Drift-specific labels:

Artifact frontmatter (both files)

Audit-Schema: 1
Audit-Scope: kit
Named-Target: <repo or profile under review>
Commissioned-By: maintainer
Audit-Type: workflow-drift-pass
Audited-By: drift-guard
Action-By: <name>
GitHub-User: <handle>
Date: <ISO-8601>
Profile: kit-dev | consumer
Command: python3 -m agent_colony drift validate --directory . [--profile consumer]

Include ## Accountability summary and mandatory ## Audit limits. P0/P1 script findings mirrored in drift-todos.md need owner, due_slice, consequence_if_ignored, status (missing/open → Schema-1 FAIL), category, paths, recommendation, and evidence when stamped Audit-Schema: 1.

Consumer DRIFT-005: When IMPLEMENTATION-STATUS.md is absent (normal on plugin installs), DRIFT-005 PASSes (skip). A FAIL on the missing file is a kit false positive on older payloads — not a consumer app defect. See consumer-quickstart.md § Drift on consumer apps.

drift-audit.md

Summary, per-check table (ID, severity, pass/fail, detail, evidence path), Goal pulse subsection, verdict (GO / blocked on P0).

drift-todos.md

Open findings with id, severity, evidence, recommendation, status (open | fixed | deferred | accepted_divergence).

Severity handling

Overlap (do NOT duplicate)

Governance/debrand → check_governance_consistency.py. Agent/registry deep structure → integrate validate. Architecture / CHK-* → auditor. Claims → verifier. Agent doc/canvas content parity vs machine truth → auditor + agent-surface-parity (not this skill).

Exit criteria

Script run captured; both artifacts written; P0 count explicit; Goal pulse noted; handoff names next agent if blocked.

Peers (debugger)

SG rows in debugger publish packs are standards gaps, not DRIFT findings — open DRIFT only when doctrine/docs drift is confirmed by script output.