Get the coding-agent leaderboard for an eval suite
Ranks the suite’s completed coding-agent runs, one row per (agent, model, effort). Only runs that recorded an agent_name participate — a QA or transcription run of the same suite has no agent to attribute a score to and is not a board row. accuracy is the mean of metric (default reward) over the row’s scored trials, and accuracy_ci is the 95% Wald interval around it; both are derived per request rather than stored, so rerunning a failed trial moves them instead of leaving a stale copy behind. agent_version is reported but is not part of the row identity, so an agent upgrade does not split a row in two. Rows are ordered by accuracy descending, nulls last. Pass suite_version when the suite has more than one, otherwise the board mixes task sets.
Authorizations
Gateway auth: send Authorization: Bearer <mka1-api-key>. For multi-user server-side integrations, you can also send X-On-Behalf-Of: <external-user-id>.
Headers
Optional external end-user identifier forwarded by the API gateway.
Path Parameters
Query Parameters
Restrict the board to one suite version. Omit to aggregate every version, which mixes task sets — prefer pinning it when the suite has more than one.
1 <= x <= 2147483647The scores key to average into accuracy. Defaults to reward, what the harness writes for a pass/fail task. Matched as a single top-level key, so reward.avg looks up the literal key.
128^[\w.-]+$Restrict to one agent name.
128Restrict to one model.
255Maximum rows returned.
1 <= x <= 200Response
OK
The version this board was restricted to, or null when it spans every version.
-9007199254740991 <= x <= 9007199254740991The scores key accuracy averages. Echoed back so a board rendered from a non-default metric is self-describing.
Rows ordered by accuracy descending, nulls last, then by agent, model, and effort.