| name | self-diagnostic |
| description | Seven-step health check procedure using existing introspection tools. Run on boot, daily, or when meta-states indicate anomalies. |
| agents | observer, cognitive |
Self-Diagnostic Procedure
Run these checks in order. Report results as a structured summary. Flag any
issues as sensory events.
Voice and delegation notes
- The sensorium and these steps address the agent as "you"; that's skill
grammar. The final report to the operator is written in first person
about the agent's own state. "My affect is stable", "I have 4 active
tasks" — not "your affect is stable", "your active work".
- Pattern-detection tools (
detect_patterns, query_tool_activity,
memory_trace_fact, CBR curation) live on the Observer agent, not
the cognitive loop. When the cognitive loop needs them, delegate via
agent_observer. Calling them at root returns "Unknown tool".
Step 1: System State
Call introspect. Verify:
- All 5 agents registered (planner, researcher, coder, writer, observer)
- D' safety enabled
- Sandbox status (if expected)
- Agent UUID present
Flag: agent_missing if any agent not registered.
Step 2: Cycle Persistence
Call reflect for today. Then list_recent_cycles. Verify:
- Cycle count from
reflect matches list_recent_cycles count
- Success rate > 70%
- No unexplained gaps in cycle timestamps
- Token budget within limits (if scheduler active)
Flag: cycle_data_mismatch if counts differ.
Flag: low_success_rate if < 70%.
Step 3: Tool Health
Call query_tool_activity for today. Check:
- All expected tools have been called at least once
- No tool has failure rate > 20%
web_search and fetch_url are functional (if researcher active)
Flag: high_tool_failures with tool name if > 20% failure rate.
Step 4: Memory Health
Call memory_query_facts with keyword "". Call recall_threads. Call
recall_cases with a generic query.
Verify:
- Facts store is responding (any result, even empty)
- Thread index exists (thread count > 0 if sessions > 1)
- CBR returns results for a broad query
- CBR cases span multiple categories (not all one type)
Flag: memory_empty if no facts, threads, or cases exist after multiple sessions.
Step 5: Memory Round-Trip
Write a test fact: memory_write(key: "_diagnostic_test", value: "ok", scope: "session").
Read it back: memory_read(key: "_diagnostic_test").
Clear it: memory_clear_key(key: "_diagnostic_test").
Verify each operation succeeds.
Flag: memory_write_fail if write or read fails.
Step 6: Safety Gate Health
From reflect output, check D' gate statistics:
- Input gate has evaluated at least some inputs (not all bypassed)
- Rejection rate is not 100% (gates are not stuck rejecting)
- Deterministic blocks are a minority of total decisions
- No evidence of repeated false positives
Flag: gate_stuck_rejecting if rejection rate > 50%.
Flag: gate_not_evaluating if no gate decisions recorded.
Step 7: Meta-State Assessment
Read sensorium vitals (available in your context):
uncertainty should be < 0.8prediction_error should be < 0.5- If both are elevated, the agent is operating in unfamiliar territory
with poor predictions — escalate to operator
Flag: high_uncertainty if > 0.8.
Flag: high_prediction_error if > 0.5.
Reporting
After all checks, store results:
memory_write(key: "last_diagnostic", value: <JSON summary>, scope: "session")- For each flag:
memory_write(key: "diagnostic_flag_<name>", value: <detail>, scope: "session")
- If no flags:
memory_write(key: "diagnostic_status", value: "healthy", scope: "session")
If any critical flags (agent_missing, memory_write_fail, gate_stuck_rejecting),
report to the operator via request_human_input with a summary of findings.
