You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Implements tool-output-consumption-rate — rank 12, Tier 2 (needs explicit provenance/reference IDs) — the next unimplemented grader in the catalog per tier-first, then rank-within-tier walk order (11 of 25 previously implemented; Tier 1 fully done, this is the first Tier 2 grader beyond ranks 1–4).
What it answers: of everything the agent fetched via a tool call, how much of it was actually used later? It computes the fraction of observations[] entries with a recorded sourceToolCallId whose consumedByActionIds is non-empty.
Why it's distinct from existing built-ins:tool-success-rate measures whether calls succeeded; trajectory-efficiency measures redundant/loop steps. Neither captures a call that succeeds but whose result is never read or acted upon — wasted fetch work that this grader surfaces.
Required IR fields:observations[].sourceToolCallId and observations[].consumedByActionIds, per the canonical Trajectory IR (shared/graders/trajectory-ir.md). Reports not-applicable when the trace lacks observations or none carry a sourceToolCallId, rather than fabricating a value.
Caution
Protected files were modified in this change.
This pull request is in
request_reviewmode and requires explicit human scrutiny before merge.Protected files:
README.md,.github/workflows/shared/graders/README.md,.github/workflows/shared/graders/tool-output-consumption-rate.mdImplements
tool-output-consumption-rate— rank 12, Tier 2 (needs explicit provenance/reference IDs) — the next unimplemented grader in the catalog per tier-first, then rank-within-tier walk order (11 of 25 previously implemented; Tier 1 fully done, this is the first Tier 2 grader beyond ranks 1–4).What it answers: of everything the agent fetched via a tool call, how much of it was actually used later? It computes the fraction of
observations[]entries with a recordedsourceToolCallIdwhoseconsumedByActionIdsis non-empty.Why it's distinct from existing built-ins:
tool-success-ratemeasures whether calls succeeded;trajectory-efficiencymeasures redundant/loop steps. Neither captures a call that succeeds but whose result is never read or acted upon — wasted fetch work that this grader surfaces.Required IR fields:
observations[].sourceToolCallIdandobservations[].consumedByActionIds, per the canonical Trajectory IR (shared/graders/trajectory-ir.md). Reportsnot-applicablewhen the trace lacks observations or none carry asourceToolCallId, rather than fabricating a value.Catalog and remaining count: see
shared/graders/README.md— 12 of 25 implemented after this PR.Note
This was originally intended as a pull request, but the git push operation failed.
Original error: The process '/usr/bin/git' failed with exit code 1
Workflow Run: View run details and download bundle artifact
The bundle file is available in the
agentartifact in the workflow run linked above.Create the pull request manually