Skip to main content
RELAI reads traces, transcripts, or supporting context, finds durable patterns, and records requirements, risks, and open questions before you decide what to test.
Ingestion does not create agent tests.It updates EvalsWiki and stores selected evidence attachments. Create or discover a grounded test afterward.
Run ingestion from an initialized RELAI project with an existing .relai/evals-wiki/. If initialization skipped the Wiki, run relai wiki bootstrap first.

​
Pick the right workflow

​
1. Ingest traces into EvalsWiki

The source may be outside the repository. Remove secrets and unrelated private data first, then use review mode to inspect ordinary Wiki patches.
Replace the path and focus, then paste this prompt.

Use RELAI to ingest the trace folder at with Wiki review mode. Focus on . Summarize the proposed requirements, risks, open questions, and selected attachments. Do not create a test yet.

​
What RELAI reads and writes

Review mode is not a zero-write preview.RELAI pauses before ordinary Wiki patches, but selected attachments and risk-hypothesis updates may already be written. Review attachments before committing them.

​
2. Create a test from what RELAI learned

First inspect the curated knowledge. Then create one grounded agent test from the highest-impact requirement or risk.
First, review what RELAI learned.

Use RELAI to show what this trace ingestion added to EvalsWiki.

Then create grounded coverage.

Use RELAI to create one agent test for the highest-impact requirement or risk learned from these traces.

​
Resume safely

An ingestion has a one-hour limit per invocation and saves resumable progress on timeout. Resume rechecks the original source folder and preserves review mode. Restart discards only that unfinished session; it does not roll back durable Wiki or attachment writes already made.

Use RELAI to inspect and resume the unfinished trace ingestion. Preserve accepted changes and explain anything that still needs my decision.

Use --timeout 2h, --timeout off, or RELAI_TIMEOUT when the next invocation needs a different budget. Repeat the override on resume.

Ready to evaluate one learned behavior?

Create a grounded agent test and simulate the current agent.

​
Command reference

See relai ingest for all flags, defaults, and recovery details.