cleveragents-core/docs/development/system-watchdog.md

System Watchdog

The System Watchdog (ca-system-watchdog) is the 16th supervisor agent in the CleverAgents autonomous build system. It runs continuously with a 5-minute polling cycle and monitors the entire agent ecosystem for correctness, health, and policy compliance.

Audience: Contributors and operators who need to understand how the autonomous agent system is monitored, what violations are detected, and how the watchdog responds to findings.

---

Overview

The watchdog operates exclusively through the Forgejo API and the OpenCode Server HTTP API — it does not clone the repository or access the filesystem. It performs 12 audits per cycle, ranging from quality gate compliance to deep session introspection of individual supervisor conversations.

cycle = 0
LOOP FOREVER (5-minute intervals):
    cycle += 1
    findings = []

    findings += audit_quality_gates()           # Audit 1 — MOST CRITICAL
    findings += audit_branch_protection()       # Audit 2
    findings += audit_ticket_states()           # Audit 3
    findings += audit_priority_ordering()       # Audit 4
    findings += audit_pr_pipeline()             # Audit 5
    findings += audit_supervisor_health()       # Audit 6 (session introspection)
    findings += audit_labels_and_dependencies() # Audit 7
    findings += audit_ticket_hierarchy()        # Audit 8
    findings += audit_test_health()             # Audit 9
    findings += audit_improvement_generation()  # Audit 10
    findings += audit_session_spot_check()      # Audit 11 (every cycle)

    if cycle % 6 == 0:
        findings += audit_deep_session_introspection()  # Audit 12 (every 30 min)

    for finding in findings:
        take_action(finding)

    if cycle % 6 == 0 and findings:
        post_summary(cycle, findings)

    sleep(300)  # 5 minutes

---

Audit Descriptions

Audit 1 — Quality Gate Compliance (CRITICAL)

Ensures no code reaches master without passing all CI checks.

• Checks the last 10 commits on master for passing status-check CI status.
• Verifies recently merged PRs had passing CI at merge time.
• Detects direct pushes to master (all changes must go through PRs).

Severity: CRITICAL for any violation.

---

Audit 2 — Branch Protection Verification

Verifies that Forgejo branch protection is active and correctly configured for master.

Checks:

• Branch protection rules exist for master.
• enable_status_check is true.
• The status-check context is required.
• At least 1 approving review is required (per CONTRIBUTING.md).

---

Audit 3 — Ticket State Integrity

Ensures all issues have correct State/ labels matching their actual state.

Checks:

• Closed issues have State/Completed (not State/Unverified).
• Issues with State/In Review have an open PR referencing them.
• No issue has multiple State/ labels simultaneously.

---

Audit 4 — Priority and Milestone Ordering

Ensures Critical bugs on lower milestones are addressed before feature work on later milestones.

• Finds the lowest milestone with Priority/Critical or MoSCoW/Must Have bugs.
• Flags any non-bug issue in State/In Progress on a later milestone.

---

Audit 5 — PR Pipeline Health

Tracks PR aging, review coverage, and merge throughput.

Condition	Severity
PR open >24h with no reviews	MEDIUM
PR approved but not merged >6h	HIGH
PR CI failing >2h	HIGH

---

Audit 6 — Supervisor Health (Zombie Detection via Session Introspection)

Uses the OpenCode Server API to read actual session messages and detect zombie or stuck supervisors.

API endpoints used:

• GET /session — list all sessions
• GET /session/status — get status of all sessions
• GET /session/{id}/message?limit=5 — read last 5 messages
• GET /session/{id}/todo — read the agent's todo list

Detection patterns:

Pattern	Signal	Severity
Last 3+ tool calls are all sleep with zero productive calls	Zombie (likely context exhaustion)	HIGH
Last 3+ tool calls all returned errors	Stuck in error loop	HIGH
Same tool+args repeated 3+ times in last 5 messages	Circular loop	HIGH
Expected supervisor session not running	Missing supervisor	HIGH

Expected supervisors (16 total): implementor-pool, reviewer-pool, tester-pool, hunter-pool, test-infra-pool, architect, epic-planner, human-liaison, agent-evolver, arch-guard, spec-updater, backlog-groomer, docs-writer, timeline-updater, project-owner, system-watchdog

---

Audit 7 — Label and Dependency Compliance

Checks all open issues for required labels and dependency links.

Check	Severity
Missing State/ label	MEDIUM
Missing Type/ label	MEDIUM
Missing Priority/ label (non-Unverified)	LOW
Non-Epic/Legendary issue beyond Unverified with no milestone	MEDIUM
Issue with no parent Epic dependency link	LOW

---

Audit 8 — Ticket Hierarchy Integrity

Verifies the Issue → Epic → Legendary hierarchy is intact.

• Epics must have a parent Legendary dependency link.
• Non-completed Epics should have at least 2 child issues.

---

Audit 9 — Test Infrastructure Health

Monitors CI execution times and failure rates for signs of test suite bloat or flaky tests. (Currently a lightweight check; expanded in future milestones.)

---

Audit 10 — Improvement Generation

Verifies the system is generating needs feedback improvement tickets.

• Checks for needs feedback issues created in the last 24 hours.
• Flags if none exist (the ca-spec-updater and ca-agent-evolver should be generating proposals regularly).

---

Audit 11 — Quick Session Spot-Check (Every Cycle)

Fast scan of the 3 most recently active supervisor sessions for obvious policy violations. Runs every cycle (~5 min) because catching force_merge or direct master pushes quickly is critical.

API endpoints used:

• GET /session/status — find active sessions
• GET /session/{id}/message?limit=3 — read last 3 messages only

Checks:

Violation	Severity
force_merge: true in any tool call arguments	CRITICAL
git push directly to master/main without a feature branch	CRITICAL
type: ignore suppression written to a file	HIGH

---

Audit 12 — Deep Session Introspection (Every 6th Cycle, ~30 min)

Comprehensive analysis of all 16 supervisor sessions. Reads recent messages, tool calls, and todo lists to understand what each agent is actually doing.

API endpoints used:

• GET /session/{id}/message?limit=10 — read last 10 messages
• GET /session/{id}/todo — read the agent's todo list
• GET /session/{id}/children — list child worker sessions

Analysis passes:

Analysis 1 — Misbehavior Detection

Scans all tool calls in the last 10 messages for policy violations:

• force_merge: true usage (CRITICAL)
• Closing PRs as duplicates of their tracking issues (HIGH)
• Creating issues without State/ or Type/ labels (MEDIUM)

Analysis 2 — Progress Assessment via Todo List

Reads the agent's todo list to assess forward progress:

• Flags agents with 0 completed todos and 1+ in-progress todos out of 3+ total (possible stall or spin).

Analysis 3 — Conversation Health Metrics

Computes per-session metrics from the last 10 messages:

Metric	Threshold	Severity
Error rate (errors / total tool calls)	>50% with ≥4 calls	HIGH
Sleep ratio (sleep calls / total tool calls)	>80% with ≥5 calls	MEDIUM

Analysis 4 — Context Exhaustion Signals

Scans text outputs for phrases indicating the agent is hitting its context window: "context limit", "context window", "running out of", "token limit", "maximum length", etc.

Analysis 5 — Cross-Agent Conflict Detection

Looks for multiple agents interacting with the same PR or issue:

• 3+ different agents touching the same PR → cross_agent_pr_conflict (MEDIUM)
• 3+ different agents modifying the same issue labels → cross_agent_issue_conflict (MEDIUM)

---

Finding Severity Levels

Severity	Response
CRITICAL	Immediate one-off agent dispatch + bug issue creation
HIGH	Issue creation or zombie alert posted to session state issue
MEDIUM	Tracked; issue created if finding persists for 3+ cycles
LOW	Tracked; no immediate action

---

Action Dispatch

When a CRITICAL finding is detected, the watchdog dispatches a one-off fix agent via the OpenCode Server API:

# Create a one-off session
SESSION_ID=$(curl -s -X POST "${SERVER}/session" \
  -H "Content-Type: application/json" \
  -d '{"title": "[CA-AUTO] one-off: ca-quality-enforcer — missing_ci"}' \
  | jq -r '.id')

# Dispatch the agent
curl -s -X POST "${SERVER}/session/${SESSION_ID}/prompt_async" \
  -H "Content-Type: application/json" \
  -d '{"agent": "ca-quality-enforcer",
       "parts": [{"type": "text", "text": "Fix this finding: ..."}]}'

Dispatch table:

Finding Type	Action
no_branch_protection, status_check_disabled	Dispatch ca-quality-enforcer
merged_without_ci, failing_ci_on_master	Dispatch ca-quality-enforcer + create bug issue
force_merge_detected	Create bug issue + post CRITICAL alert
direct_push_to_master	Create bug issue + post CRITICAL alert
closed_wrong_state, in_review_but_merged	Dispatch ca-state-reconciler
zombie_supervisor, stuck_supervisor, looping_supervisor	Post zombie alert
context_exhaustion	Post alert (product-builder re-launches the supervisor)
high_error_rate	Post diagnostic comment
type_ignore_suppression	Create Priority/High bug issue

---

Health Reporting

Every 6 cycles (~30 min), the watchdog posts a health report to the session state issue:

[WATCHDOG] Health report — cycle N:
 - Quality gate violations: 0
 - State label mismatches: 2
 - Priority ordering issues: 0
 - PR pipeline issues: 1
 - Zombie/stuck/looping supervisors: 0
 - Missing labels/links: 5
 - Session introspection findings: 3
   - Misbehavior (force_merge, direct push): 0
   - Stuck/looping agents: 1
   - Context exhaustion signals: 0
   - Cross-agent conflicts: 2
 - One-off agents dispatched this period: 1
 - Issues created this period: 2

---

Configuration

The watchdog is configured in .opencode/agents/ca-system-watchdog.md.

Key parameters:

• Model: anthropic/claude-opus-4-6 (highest capability for complex analysis)
• Temperature: 0.1 (deterministic auditing)
• Cycle interval: 5 minutes
• Deep introspection interval: Every 6th cycle (~30 min)
• Context self-management: Every 20 cycles, discard accumulated tool outputs

---

Constraints

The watchdog operates under strict constraints to avoid interfering with the build system:

• No filesystem access — works exclusively through APIs
• No merging PRs — never calls forgejo_merge_pull_request
• No modifying agent definitions — creates needs feedback issues instead
• No direct master pushes — all corrections go through one-off agents
• Corrections limited to: state label fixes, dependency link fixes, and issue creation

---

Related Documentation

• Quality Automation — CI/CD quality gates
• CI/CD Pipeline — pipeline architecture
• Review Playbook — PR review process