Skip to main content
For a guided workflow, see EvalsWiki. relai wiki manages EvalsWiki, RELAI’s local project knowledge base. Init bootstraps it by default, and later commands use it to ground agent tests, evaluators, benchmarks, and optimization work.

​
Project knowledge types

RELAI curates three user-facing categories:
  • Requirements are grounded behaviors the agent appears expected to satisfy.
  • Risk hypotheses are likely ways the current agent may fail. RELAI can turn them into agent tests through discovery.
  • Open questions are unresolved ambiguities. They are not treated as requirements until a user answers them.

​
Inspect project knowledge

Use --id {id} to inspect one item in detail and --json when another tool needs deterministic output.

​
Generate agent tests from Wiki items

Create an agent test from a selected risk:
Create an agent test from a selected requirement:
The risk flow runs relai test discover for the selected risk. Use relai test discover --risk {id} directly when you need discovery controls such as the invocation-wide --timeout or --no-review. The requirement flow runs relai test create --requirement {id} and supports normal creation controls such as --agent-target, --guidance, permissions and the invocation-wide timeout. Every relai wiki command runs for at most 1 hour by default, including the discovery or creation it starts; pass --timeout {DURATION|off} to change that (see Timeouts).

​
Resolve open questions

When you know the answer to an open question, resolve it through RELAI’s Wiki update flow:
You can also start from the list command:
Resolution treats the answer as explicit user intent. It may create or update a requirement and remove the resolved open-question page. The patch applies directly; pass --review to approve it first.

​
Bootstrap, update, and reconcile

relai init runs Wiki bootstrap by default. Use these commands directly when you need more control:
  • bootstrap builds or merges the initial Wiki from repository files, Git history, local RELAI artifacts, and local run/evaluation logs.
  • update records durable project decisions, corrections, constraints, and preferences from a bounded evidence envelope.
  • reconcile reviews agent tests, evaluators, and benchmarks whose grounded requirements changed.
To curate a folder of traces or other context, see Ingest traces and context. Use backlinks to find registered artifacts grounded in a requirement or other stable Wiki knowledge ID.