Search BrigadeStack

Search BrigadeStack pages and docs.

BrigadeStack
Menu
Docs menu

Docs / Checked work

Score skills and agents

Use check records to see which skills and notes help your agents get work right, and which ones to drop.

Once your checks leave records, brigadestack outcome turns them into data. It ties each real exit code to the skill or card that was in play and ranks them, so you can see what helps. The model never grades its own work.

Record a result

Run a check and capture it against the skill or card that produced the work, in one step:

brigadestack work verify run --manifest <id> --capture my-skill

Or capture a verify run you already did with brigadestack outcome capture my-skill --run-id latest. You name the artifact, because BrigadeStack does not guess which skill caused a result. Only a run from a registered verifier manifest scores it, so ad-hoc --command runs stay audit-only. See Which checks count.

Read the data

brigadestack outcome rank
brigadestack outcome rank --by-capability
brigadestack outcome explain my-skill

rank orders artifacts by a lower confidence bound, so something that passed twice never outranks something vetted across twenty runs. --by-capability ranks by what worked under your current tool and model. explain prints the signal trail behind a score.

Act on it

brigadestack outcome reconcile is dry-run by default. With --apply it installs a skill that earned it across your agent tools, or rolls back one that regressed. Both are reversible. The full flag list is on the outcome reference.

Model scorecard

brigadestack model scorecard ships in 0.31.0. It reads the run artifacts under .brigade/runs/ (the roster and worker results of each brigadestack run) and never looks at check records. It is read-only and groups the runs by CLI and model: runs, worker seats, how many workers finished ok, suspected no-op runs and mean duration. --since YYYY-MM-DD limits the window, --runs-dir points at other run folders, and --json prints it as data. It tells you which tool and model completed workers, not whether the work passed a check.

0.32.0 beta. In observe mode, hooks record the raw checks agents run. brigadestack work checks reports each tool’s and model’s first attempts as passed, failed or unknown, and --group-by harness combines models. These are observed checks, not signed check records, and a rate appears only after five reported results. It is not in 0.31.0. In 0.31.0 the check results are the records that outcome rank reads, one command at a time, as above.