Tutorials
Agent-Led Analysis
Install the bundled isplay-analysis skill and let an analyst agent run a bounded investigation.
The repo includes an isplay-analysis skill for Codex-style agents. It teaches the agent to start the stack, select a project, discover runs, form narrow hypotheses, run replay or experiments, resolve fixture requirements, and write validity-bounded reports.
1. Install The Skill
npx skills add combined-ai/isplay --skill isplay-analysis -a codexThe command installs directly into Codex. Omit -a codex if you want the Skills CLI to prompt for another supported agent. Restart the agent app after installation.
2. Start With A Direct Prompt
Use $isplay-analysis. Start isplay if needed, create or select a project, wire the right adapter for this project, capture my next run, discover the trace, replay likely hypotheses, resolve fixture requirements when safe, and return an evidence-bounded RCA report with validity labels.3. What The Agent Should Do
Verify
ISPLAY_API_URL and ISPLAY_PROJECT_ID.Capture or select the target run.
Run
discover run, context inventory, checkpoint listing, and event inspection.Write an observed baseline before forming a cause claim.
Create one branch per hypothesis.
Use recorded model replay and pause-for-fixture tool replay by default.
Resolve fixture requirements only with explicit provenance and scope.
Return a report with first divergence, metrics, effects, comparability, validity labels, and next actions.
4. How To Review The Output
Check that the final answer includes:
| Required item | Why |
|---|---|
| Base run ID and project ID | Makes the investigation reproducible. |
| Replay or experiment IDs | Lets you inspect evidence later. |
| Replay policy | Defines evidence strength. |
| Fixture provenance | Explains conditional dependencies. |
| First divergence | Shows where behavior changed. |
| Metrics and effects | Connects trace differences to the conclusion. |
| Validity labels | Prevents overclaiming. |
| Next actions | Makes uncertainty actionable. |
If the report jumps straight to a cause without these items, ask it to rerun discovery and cite evidence.