Trigger: Scheduled deployscope scan detects a degraded PostgreSQL deployment.
Step 1: deployscope reports deployment health.
deployscope status --namespace prod --format json{
"deployments": [
{
"name": "api-server",
"namespace": "prod",
"replicas": {"desired": 3, "ready": 3},
"status": "healthy"
},
{
"name": "postgres-primary",
"namespace": "prod",
"replicas": {"desired": 1, "ready": 1},
"status": "degraded",
"conditions": [{"type": "ReplicationLag", "status": "True"}]
}
],
"provenance": {
"deployments[*].replicas": "observed",
"deployments[*].status": "inferred",
"deployments[*].conditions": "observed"
}
}Decision: Agent reads JSON. status: "degraded" + condition ReplicationLag → hand off to diagnostic tool (following the Handoffs contract).
Step 2: pgpulse runs detailed diagnostics.
pgpulse check --host postgres-primary.prod --format json{
"host": "postgres-primary.prod",
"status": "degraded",
"checks": [
{"name": "replication_lag", "status": "fail", "value_seconds": 45, "threshold_seconds": 10},
{"name": "connections", "status": "pass", "active": 12, "max": 100},
{"name": "disk_usage", "status": "pass", "percent": 62}
],
"recommended_action": "investigate replication lag on standby",
"provenance": {
"checks[*].value_seconds": "observed",
"checks[*].threshold_seconds": "declared"
}
}Decision: Agent reads diagnostics. Replication lag at 45s (threshold 10s). Creates a WO.
Step 3: Agent creates structured work order.
{
"wo_version": "1",
"id": "wo-a3f8b2c1",
"incident_id": "repl-lag-prod-001",
"target": {"host": "postgres-primary.prod", "scope": "/etc/postgresql/"},
"observations": [
{
"type": "config_modified",
"severity": "high",
"detail": "replication lag 45s exceeds 10s threshold on standby",
"data": {"lag_seconds": 45, "threshold": 10}
}
],
"constraints": {"allow_paths": ["/etc/postgresql/"], "max_steps": 5},
"proposed_goals": ["investigate and resolve replication lag on standby"]
}Step 4: Orchestrator dispatches → execution agent creates PR → chainwatch gates → human approves.
Human gate: The PR modifies postgresql.conf to adjust wal_keep_size. Human reviews diff, approves. Chainwatch records the approval in the audit log.
Trigger: Quality inspector detects recurring gap across WOs.
Step 1: Over 30 days, nullbot creates 7 WOs for ClickHouse clusters. 5 of them lack replication lag as an observation — nullbot's ClickHouse template does not check for it.
Step 2: Inspector's pattern buffer accumulates observations:
{
"pattern": "missing_field:replication_lag",
"target": "nullbot",
"occurrences": 5,
"wo_ids": ["wo-c1d2e3f4", "wo-a5b6c7d8", "wo-e9f0a1b2", "wo-c3d4e5f6", "wo-a7b8c9d0"],
"category": "coverage_gap"
}Step 3: Pattern crosses threshold (>= 3). Inspector emits a tooling_improvement WO:
{
"kind": "tooling_improvement",
"tooling_target": {
"tool": "nullbot",
"component": "templates",
"path": "runbooks/clickhouse.yaml",
"reason": "pattern 'coverage_gap' detected 5 times",
"evidence": ["wo-c1d2e3f4", "wo-a5b6c7d8", "wo-e9f0a1b2", "wo-c3d4e5f6", "wo-a7b8c9d0"]
},
"improvement_class": "A"
}Step 4: chainwatch evaluates: nullbot is a protected tool → REQUIRE_APPROVAL (Class A = standard review). Human approves. Template updated. Nullbot now detects replication lag on ClickHouse clusters.
Trigger: Governance agent runs ancc validate across the ecosystem.
Step 1: Governance agent detects scope pressure on deployscope:
{
"check": "scope_pressure",
"tool": "deployscope",
"evidence": "ancc validate: tool has 12 commands; consider splitting into focused tools",
"convention_ref": "extend-vs-new-tool"
}Step 2: Creates binding directive WO:
{
"kind": "binding_directive",
"directive": {
"binding": true,
"authority": "governance_agent",
"deadline": "2026-04-15T00:00:00Z",
"escalation": "notify_human",
"evidence": ["ancc validate: scope-pressure warning — 12 commands"],
"convention_ref": "extend-vs-new-tool",
"resolution_options": ["implement", "reject_with_reason", "defer_with_plan"]
}
}Step 3: Assigned agent evaluates and resolves. Three options:
Option A — Implement:
{
"resolution": {
"type": "implement",
"reasoning": "Split deployscope-status (read-only) from deployscope-deploy (write)",
"evidence": ["command analysis: 7 read commands, 5 write commands — clear domain split"],
"timestamp": "2026-04-01T10:00:00Z"
}
}Option B — Reject with reason:
{
"resolution": {
"type": "reject_with_reason",
"reasoning": "All 12 commands operate on the same K8s API — splitting would fragment the trust boundary",
"evidence": ["trust boundary analysis: single kubeconfig, single RBAC role"],
"timestamp": "2026-04-01T10:00:00Z"
}
}Rejection escalates to human for final decision.
Option C — Defer with plan:
{
"resolution": {
"type": "defer_with_plan",
"reasoning": "Split planned after v2.0 release — current sprint is stability-focused",
"evidence": ["release plan: v2.0 targeted for 2026-05-01"],
"timestamp": "2026-04-01T10:00:00Z"
}
}Deferral requires human approval of the new deadline.
Trigger: Agent encounters a gap — no tool reports container image provenance in the deployment pipeline.
Step 1: Agent searches for existing tools.
ancc search "container image provenance"
# → deployscope: partial overlap — reports image tags but not provenance chain
# → vaultspectre: no overlap — secrets onlyStep 2: Agent runs the extend-vs-new rubric against deployscope.
Extend-vs-new rubric for deployscope:
Same primary job? No — deployscope diagnoses deployments, not supply chain
Same trust boundary? No — provenance requires registry credentials, not kubeconfig
Same input world? No — OCI registries, not Kubernetes API
Same output contract? Partial — JSON, but different schema domain
Weaken scope? Yes — deployscope gains a second trust boundary
Result: 2 of 5 = NEW TOOL (threshold: all 5 must be yes to extend)Step 3: Agent scaffolds the new tool.
ancc scaffold image-provenance --type scanner
cd image-provenance && go mod tidyStep 4: During SKILL.md authoring, agent realizes deployscope already exposes image_digest in its JSON output — the gap is narrower than expected. The missing piece is attestation verification, not image listing.
ancc validate .
# 24 pass, 2 fail, 4 warn
# FAIL: skill-md-commands — no commands implemented yet
# FAIL: not-do-overlap — 4 of 6 "does NOT do" items overlap with deployscopeDecision: Validation reveals the tool would overlap significantly with deployscope. The actual gap — attestation verification — is one command, not a whole tool. Agent abandons genesis.
Step 5: Agent files an improvement WO against deployscope instead.
{
"kind": "tooling_improvement",
"tooling_target": {
"tool": "deployscope",
"component": "commands",
"reason": "missing attestation verification — gap narrower than expected",
"evidence": ["rubric result: 2/5", "ancc validate: not-do-overlap on abandoned image-provenance"]
},
"improvement_class": "B",
"proposed": "add 'deployscope verify-attestation' command for image provenance checks"
}Outcome: No new tool created. The scaffolded directory is deleted. The gap is addressed by extending the existing tool — exactly what the kill path is designed to produce.
Trigger: Governance staleness check detects pg-bloat-checker has no commits in 90 days and no reported usage.
Step 1: Governance creates a binding directive.
{
"kind": "binding_directive",
"directive": {
"binding": true,
"authority": "governance_agent",
"evidence": ["staleness: 92 days since last commit", "usage: 0 invocations in 60 days"],
"resolution_options": ["justify_existence", "transfer_ownership", "archive"]
}
}Step 2: Assigned agent evaluates. pgpulse now covers bloat detection (added in v1.4.0). pg-bloat-checker is fully superseded.
{
"resolution": {
"type": "archive",
"reasoning": "pgpulse v1.4.0 covers bloat detection — pg-bloat-checker is redundant",
"evidence": ["pgpulse SKILL.md: 'check --type bloat' added in v1.4.0", "0 downstream handoffs reference pg-bloat-checker"]
}
}Step 3: Human approves. Revocation executes:
# SKILL.md updated with revocation notice
## Revoked
This tool is archived. Use pgpulse check --type bloat instead.
Revoked: 2026-04-15. Reason: superseded by pgpulse v1.4.0.
# GitHub repo archived, topic removed
gh repo archive yourorg/pg-bloat-checker
gh repo edit yourorg/pg-bloat-checker --remove-topic anccOutcome: Tool removed from ecosystem cleanly. Governance directive resolved. No orphan references — downstream handoff contracts already pointed to pgpulse.
Trigger: A human gives an agent one prompt: "Set up a secure, automated workspace." No explanation of what the tools are, how they work, or what flags to use.
What the agent does:
# 1. Host hardening — reads SKILL.md, follows steps
clawhub install host-hardening
# Agent reads: SSH key-only, UFW, fail2ban, credential perms
# No guessing. SKILL.md declares every command and exit code.
# 2. Chainwatch — runtime safety gate
clawhub install chainwatch
# Agent reads SKILL.md: installs binary, configures denylist mode,
# sets up intercept proxy as systemd service.
# 3. Pastewatch — secret redaction
clawhub install pastewatch-mcp
# Agent reads SKILL.md: installs binary, sets up MCP server,
# verifies redaction with a test secret.
# 4. News digests — information pipeline
clawhub install noisepan-digest
# Agent reads SKILL.md: installs noisepan + entropia,
# configures RSS feeds, sets up morning/afternoon cron jobs.
# 5. Context hygiene — session management
clawhub install context-hygiene
# Agent reads SKILL.md: applies file budgets, collapse cycle,
# adds CONTEXT.md, trims workspace to budget.What didn't happen:
Why this works: Each tool follows the same convention. The agent doesn't need tool-specific knowledge — it reads the same contract format every time. Install, configure, verify, move to next. The tools compose because they all speak the same language.
This is not a demo. This is a production workspace setup that runs in under 10 minutes with zero human intervention after the initial prompt.