| name | agent-observability |
| description | Add lightweight local observability to Codex/agent workflows. Use when tracking local runs, traces, evals, cost, token usage, cache efficiency, model comparisons, failure rates, or claudetop-style monitoring. Skip when designing production gateway policy, staged guardrails, shadow traffic, model budgets, or trace-to-regression loops; use `$auto-skill-build-agent-reliability-loop`.
|
Agent Observability
Use lightweight local observability before adopting a full platform.
Minimal Run Ledger
Create:
.agent-runs/
runs.jsonl
evals.md
costs.md
Each runs.jsonl entry should include:
{"ts":"","task":"","model":"","inputs":"","commands":[],"result":"","tokens_est":"","cost_est":"","status":""}
What To Track
- task type and outcome
- commands run
- tests passed/failed
- estimated tokens and cost
- repeated failure causes
- user corrections
- guardrail or safety decisions
Eval Loop
- Define 3-10 realistic tasks.
- Run the same task across workflow variants.
- Score correctness, time, tokens, and human intervention.
- Keep the cheaper path only if quality is comparable.
Skip Or Escalate
When observability needs gateway policy, staged guardrails, shadow traffic,
trace-to-regression feedback, or model budgets, switch to
$auto-skill-build-agent-reliability-loop. Keep this skill for local ledgers
and lightweight eval/cost tracking.
Use full platforms like Future AGI when you need hosted traces, simulations,
gateway routing, or production guardrails. Use local ledgers for personal workflows.
For appending or summarizing .agent-runs/runs.jsonl, use
$auto-skill-build-agent-run-ledger-dashboard.
For Codex Supercharge artifact grades and subagent/automation handoff status,
use $auto-skill-build-supercharge-feedback-loop; it writes a separate local
feedback stream under plugins/codex-supercharge/.agent-runs/.
Validation
- Each tracked run has a task, status, commands/tests, and outcome.
- Cost/token notes are marked as estimates unless measured.
- Repeated failures are summarized into a follow-up or eval case.