Agent Observability

What Gets Installed

Installation

Step 1 — Copy files

WORKSPACE="${OPENCLAW_WORKSPACE:-$(pwd)}"

cp references/throughput-dashboard.js  "$WORKSPACE/scripts/"
cp references/decision-audit.js        "$WORKSPACE/lib/"
cp references/failure-tracer.js        "$WORKSPACE/lib/"
cp references/drift-guard-auto.js      "$WORKSPACE/scripts/"

Or manually copy each file from the references/ directory in this skill.

Step 2 — Add to heartbeat/cron (weekly)

In your heartbeat or weekly cron script:

node "$WORKSPACE/scripts/throughput-dashboard.js" "$WORKSPACE"
node "$WORKSPACE/scripts/drift-guard-auto.js" "$WORKSPACE"

Step 3 — Wire decision-audit into high-stakes decisions

const { logDecision } = require('./lib/decision-audit');

logDecision({
  task_type: 'code_generation',
  decision: 'spawn CoderAgent',
  reasoning_summary: 'Multi-file edit blocks chat >5s',
  session_channel: 'discord'  // optional
}, workspaceRoot);

Step 4 — Wire failure-tracer into quality validation (optional)

The failure-tracer fires automatically when you call it after scoring subagent output:

const { captureFailureTrace } = require('./lib/failure-tracer');

// Call after scoring any subagent output
if (qualityScore < 7) {
  captureFailureTrace('AgentLabel-task', qualityScore, agentOutput, workspaceRoot);
}

Reading the Data

Query examples

# Recent decisions
tail -20 memory/decisions-audit.jsonl | jq .

# All failure traces
ls memory/traces/

# Latest drift report
cat memory/drift-reports/$(ls memory/drift-reports/ | tail -1)

Tool Descriptions

throughput-dashboard.js

Aggregates weekly metrics: tasks routed, subagents spawned, estimated cost, quality ratio, routing distribution. Reads from session-metrics.js (if installed) and drift-guard-auto.js. Degrades gracefully if data sources are missing — every section is independent.

decision-audit.js

Append-only JSONL log at memory/decisions-audit.jsonl. Each entry: { id, ts, task_type, decision, reasoning_summary, outcome, session_channel }. Use updateOutcome(id, 'success', workspaceRoot) to close the loop after a decision resolves.

failure-tracer.js

Fires when quality score < 7. Writes structured JSON to memory/traces/. Each trace includes: tool call sequence hints, output snippet, inferred failure reason. Use to post-mortem why a subagent underperformed.

drift-guard-auto.js

Scores recent agent outputs against behavioral rules (sycophancy, social cushioning, unprompted explanations, hallucination hedges). Reads INTENT.md for custom criteria if installed. Writes weekly report to memory/drift-reports/.

References

• references/throughput-dashboard.js — Full script implementation
• references/decision-audit.js — Full lib implementation
• references/failure-tracer.js — Full lib implementation
• references/drift-guard-auto.js — Full script implementation