Guides
Start Analyzing
After capture is installed, use the skill or CLI to discover a run, test hypotheses, resolve fixtures, and write a bounded RCA.
This is the first workflow to run after capture is wired and at least one run exists.
Agent Path
Use $isplay-analysis. Find my latest captured run in project <projectId>, inspect the catalog and context inventory, branch likely hypotheses, run replay experiments with recorded model policy and pause-for-fixture tool policy, resolve fixture requirements when safe, and return an RCA report with validity labels.The agent should return IDs and evidence, not just prose.
Manual Path
npx isplay runs list --project "$ISPLAY_PROJECT_ID"
npx isplay discover run <runId>
npx isplay runs context <runId>
npx isplay runs checkpoints <runId>Capture the catalog, context summary, and checkpoint list.
Write down:
- User/task input.
- Model provider, model, settings, and operation.
- System/developer/user prompt constraints.
- Tool schemas available to the model.
- Tool proposals, args, executions, outputs, and errors.
- Retrieval and memory items.
- Checkpoint state.
- Final response or side effect.
Turn possible causes into narrow tests:
Because <context/tool/state evidence> entered at <point>, the agent chose <behavior>.
If we <intervention>, then <metric or effect> should <direction>, under <policy/fixture assumptions>.npx isplay branch create --from <runId> --checkpoint <checkpointId> --project "$ISPLAY_PROJECT_ID"
npx isplay branch intervene <branchId> --project "$ISPLAY_PROJECT_ID" --kind <kind> --target-ref <refId> --patch patch.json
npx isplay replay <runId> --project "$ISPLAY_PROJECT_ID" --branch <branchId>
npx isplay replay requirements <replayId>If requirements are open, resolve them before interpreting final effects.
npx isplay replay diff <replayId>
npx isplay replay metrics <replayId>
npx isplay effects replay <replayId>
npx isplay analyze <runId> --project "$ISPLAY_PROJECT_ID" --replay <replayId>Your report should cite replay IDs, effect IDs, fixture provenance, comparability, and validity labels.
Manual RCA Template
The strongest tested explanation is <hypothesis>.
Under model policy <modelPolicy> and tool policy <toolPolicy>, intervention <intervention> changed <effect/metric>.
The first divergence was step <n>, comparing <base event> to <branch event>.
The result is <supported/inconclusive/invalid> because <validity labels>.
Fixture dependence: <none/provenance and scope>.
Residual uncertainty: <missing context, low-N, non-comparable trace, nondeterminism, or untested sibling hypotheses>.
Next experiment: <smallest follow-up>.Stop Conditions
Stop and report uncertainty when:
- There is no usable checkpoint before the suspected cause.
- Context inventory lacks the prompt, tool, retrieval, memory, or state evidence needed for the question.
- Replay is non-comparable after a broad intervention.
- Fixture requirements cannot be satisfied safely.
- The result depends on an AI or simulator fixture and no sensitivity check was run.
Bound the answer
The final output should read like an audit trail, not a confident story. Tie every claim to observed trace evidence or replay output.