Agent Analysis

Use the bundled isplay-analysis skill to let an AI analyst discover traces, form hypotheses, run replays, resolve fixtures, and write bounded RCA reports.

The isplay-analysis skill turns the CLI and API into a repeatable analyst workflow. It is useful when you want an agent to investigate a captured run without skipping trace discovery or overclaiming causality.

Install The Skill

npx skills add combined-ai/isplay --skill isplay-analysis -a codex

The 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 so $isplay-analysis is discoverable.

Prompt Template

Use $isplay-analysis. Analyze run <runId> in project <projectId>. Discover the run catalog and context inventory first. Build a factual baseline narrative, form narrow hypotheses, run recorded-only model replay with pause-for-fixture tool policy, resolve fixture requirements only when safe and explicitly labeled, then return an RCA report with diffs, effects, statistics, fixture provenance, comparability, validity labels, and next experiments.

If the run ID is unknown:

Use $isplay-analysis. Find my latest captured run for project <projectId>, then analyze it with the default evidence-bounded workflow.

What The Skill Loads

SKILL.md

platform-map.md

tutorial.md

hypothesis-playbook.md

trace-reading.md

recipes.md

report-template.md

Agent Duties

Establish ISPLAY_API_URL, ISPLAY_PROJECT_ID, and runId. Ask one concise question only if those cannot be inferred.

Run discover run and fetch context inventory before forming explanations.

Build a baseline narrative from observed facts: prompts, settings, tools, retrieval, memory, state, tool args/results, checkpoints, and final behavior.

Separate observed facts from candidate causes.

Prioritize hypotheses by proximity to divergence, isolation, fixture burden, actionability, and safety.

Create branches or a hypothesis batch with one primary intervention per hypothesis.

Resolve fixture requirements with recorded or analyst fixtures only when provenance is clear; otherwise ask the user.

Read results, statistics, effects, diffs, requirements, and validity labels.

Write a bounded report that states what changed, under which policy and fixture assumptions, and what remains uncertain.

Good Agent Outputs

The report should include:

Section	Contents
Executive finding	Strongest tested explanation, status, and validity labels.
Scope	Project, run, checkpoint, experiment/replay IDs, policies, trial plan.
Baseline	Factual trace narrative before hypotheses.
Hypotheses	Statements, targets, interventions, expected effects, trial counts.
Ranked effects	Effect titles, scores, status, evidence refs, recommended actions.
Diffs	First divergence, changed descendants, tool sequence distance, tool args/output changes.
Fixture notes	Requirements, fixtures, provenance, dependency rate, sensitivity.
Residual uncertainty	Missing context, non-comparability, nondeterminism, low-N risks.
Next experiments	Smallest follow-ups that would reduce uncertainty.

Guardrails For Agents

Do not state root cause certainty unless replay evidence actually supports the wording.
Do not run live tools or models unless the user explicitly approves and policy allows it.
Do not hide fixture dependence in a footnote. Put it in the finding.
Do not patch broad prompt sections when one clause or tool schema can be targeted.
Use expected base hashes when patching observed context through experiment specs.
If context inventory is sparse, report the missing evidence instead of guessing.

Best use

Give the agent a run ID and the outcome you want explained. Let it drive discovery and experiments, but require final claims to cite replay IDs, effect IDs, fixture provenance, and validity labels.