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 rank 12 (Tier 2) from the trajectory grader catalog: tool-output-consumption-rate.
Distinct from existing built-in graders: tool-success-rate only measures whether a tool call itself succeeded/failed. This grader instead measures, for observations that succeeded, whether their output was ever actually referenced by a later action (consumedByActionIds non-empty). It surfaces wasted tool calls whose results were fetched but never used — a signal tool-success-rate and the other built-ins do not capture.
IR fields required (per shared/graders/trajectory-ir.md):
toolCalls[].outputRef — links a tool call to the observation it produced
observations[].id and observations[].consumedByActionIds — whether/what later action referenced the observation
The grader filters observations[] to those originating from a toolCalls[].outputRef, then computes consumed / total where consumed counts observations with non-empty consumedByActionIds. Returns applicable: false-equivalent (value: null, explanatory message) when no observations or no tool-originated observations are present in the IR.
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 rank 12 (Tier 2) from the trajectory grader catalog:
tool-output-consumption-rate.Distinct from existing built-in graders:
tool-success-rateonly measures whether a tool call itself succeeded/failed. This grader instead measures, for observations that succeeded, whether their output was ever actually referenced by a later action (consumedByActionIdsnon-empty). It surfaces wasted tool calls whose results were fetched but never used — a signaltool-success-rateand the other built-ins do not capture.IR fields required (per
shared/graders/trajectory-ir.md):toolCalls[].outputRef— links a tool call to the observation it producedobservations[].idandobservations[].consumedByActionIds— whether/what later action referenced the observationThe grader filters
observations[]to those originating from atoolCalls[].outputRef, then computesconsumed / totalwhereconsumedcounts observations with non-emptyconsumedByActionIds. Returnsapplicable: false-equivalent (value: null, explanatory message) when no observations or no tool-originated observations are present in the IR.Catalog: see
shared/graders/README.md— 12 of 25 graders now implemented.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