Proactive Watch — Drift Detection Skill

Continuous enforcement of the Executable Contracts meta-principle (ADR-002). Catches drift across 10 universal drift classes on every session start.

---

Skill Purpose

Detect silent drift between declared reality (docs, manifests, inventories, ADRs) and actual reality (filesystem, git state, dependencies, hooks). Report findings with severity, triage routing, and evidence — so drift is fixed while it's cheap, not months later when it's entrenched.

---

Activation

Trigger	Behavior
SessionStart hook	Runs watch --light silently; appends summary block to session banner
/mxm-watch	Runs LIGHT manually; shows full report
/mxm-watch --full	Reserved for v1.0.0 — FULL phase with agent triage
mxm-context.watch_run MCP tool	Programmatic invocation from other skills / agents

---

Dispatch Path

TASK arrives with "check drift" / "audit" / "session start"
  │
  ▼
Maxim skill .claude/skills/proactive-watch/ matches
  │
  ▼
Load config/watch-profile.yml (per project)
  │
  ▼
For each ENABLED checker:
  ├── Run checker (bash watch.sh <checker> OR pwsh watch.ps1 <checker>)
  ├── Capture JSONL report line per drift
  └── On checker error: apply fail-mode policy (hybrid)
  │
  ▼
Aggregate report → .mxm-skills/watch-report.jsonl
Emit summary block to stderr / chat (LIGHT phase)
Route critical drift to triaged office (FULL phase, v1.0.0)

---

Behavioral Framing

Apply Fogg + COM-B + Present-Bias principles when surfacing drift:

1. Fogg B=MAT — run at SessionStart when developer is engaged (max M × A)
2. COM-B — provide clear Evidence + Action path, not just "something's wrong"
3. Present-Bias Mitigation — frame every finding in terms of "fix now vs cost later"

Output language:

• ✗ Bad: "Drift detected"
• ✓ Good: "Declared 90 agents, filesystem has 91 — add to documents/ledgers/AGENT_SKILL_INVENTORY.md §1 (60s fix)"

---

Proactive Loops

The skill is itself a proactive loop. Additionally:

• After fixing an inventory drift, proactively scan for recurring-pattern opportunities — if 3+ inventory drifts happened in one month, recommend a stronger pre-commit gate
• After fixing a compliance drift, auto-route to CSO for root-cause analysis
• After fixing a contract drift, check whether the underlying ADR needs superseding

---

Confidence Tagging

Scenario	Tag
LIGHT run with all checkers passing	🟢 HIGH
LIGHT run with drift detected, remediation suggested	🟢 HIGH
FULL run (v1.0.0+) with agent triage complete	🟢 HIGH
Any checker errored (fail-open tier)	🟡 MEDIUM — report incomplete
Compliance or contract checker errored (fail-closed)	🔴 LOW — action required
super_user.enabled = true suppresses advice (not drift detection)	🔵 SUPER USER

---

Output Format

LIGHT summary (SessionStart banner)

Maxim WATCH (light)
  Drift detected : 3
  ⚠ inventory-drift: declared 90 agents, filesystem has 91 (severity 3)
  ⚠ version-drift: README badge v1.0.0, registry v1.0.0 (severity 3)
  ⚠ stale-handoff: handoff.md 9 days old (severity 2)
  Checker errors : 0
  Report         : .mxm-skills/watch-report.jsonl

JSONL report (machine-readable)

{"ts":"2026-04-18T12:00Z","phase":"light","project":"maxim","drift_class":"inventory-drift","severity":3,"declared":90,"actual":91,"evidence":"agents/MXM/ceo/new-agent.md","triage":"coo","action":"review-queue"}

---

Checker Execution Model

The skill ships two unified drivers that implement all 11 checkers (Classes 1–10 from v1.0.0; Class 11 surface-claims-drift added v1.0.1 with companion tool bootstrap/sync-counts.{sh,ps1} for mechanical propagation):

• watch.sh — bash (Linux/macOS/Git-Bash)
• watch.ps1 — PowerShell 7+ (Windows primary, cross-platform)

Each driver accepts one argument: the checker name. Running watch.sh all runs every enabled checker and aggregates output.

# Run single checker
.claude/skills/proactive-watch/watch.sh inventory-drift

# Run all enabled checkers (LIGHT phase default)
.claude/skills/proactive-watch/watch.sh all

---

Files in this Skill

File	Purpose
SKILL.md	this file — skill contract
watch.sh	unified bash driver implementing all 11 checkers
watch.ps1	unified PowerShell driver implementing all 11 checkers
schemas/watch-profile.schema.json	validates per-project config/watch-profile.yml
schemas/watch-report.schema.json	validates JSONL output lines

---

Self-Adoption (maxim repo)

This repo ships its own config/watch-profile.yml in v1.0.0 — the framework's first consumer is itself. Any drift that sneaks into this repo across multi-session work should surface at the next /mxm-watch or SessionStart.

---

Extending

Projects may add custom checkers by:

1. Writing a script at .claude/skills/proactive-watch/custom/<name>.sh (+ .ps1)
2. Registering it in config/watch-profile.yml under custom_checkers:
3. Emitting JSONL lines matching schemas/watch-report.schema.json

Custom checkers run after standard checkers.

---

Related

• Framework doc: composable-skills/frameworks/proactive-watch.md
• ADR-002: documents/ADRs/ADR-002-executable-contracts.md
• Command: .claude/commands/mxm-watch.md
• Hook: .claude/hooks/session-start.{sh,ps1} calls watch --light
• MCP: mxm-context tools watch_run, watch_report, watch_configure (v1.0.0+)

---

Copyright (c) 2026 iSystematic Inc. Maxim product. BSL 1.1.