aoslisted

# Agent Operator Score Two operators run the same model, on the same repository, with the same task. One ships. One burns the budget and merges something that does not work. Benchmarks measure the half that was identical. This measures the other half — and states the conditions every number is bound to. There are no runtime dependencies, so `node bin/aos.mjs` works from a bare clone. Nothing is uploaded, telemetry is off, and there is no switch to turn it on. ## Which half to run | the operator asked | run | |---|---| | "what went wrong in that session?" | `review` | | "what do I keep doing wrong?" | `review --since 12` | | "how good am I at running this agent?" | `assess` | `review` costs nothing and calls no model. `assess` spends model quota — confirm before running it. ## review ```bash node "${CLAUDE_PLUGIN_ROOT}/bin/aos.mjs" review # the session that just finished node "${CLAUDE_PLUGIN_ROOT}/bin/aos.mjs" review --since 12 # what recurs node "${CLAUDE_PLUGIN_ROOT}/bin/aos.mjs" review --list # pick one node "${CLAUDE_PLUGIN_ROOT}/bin/aos.mjs" review --json # machine-readable ``` Six rules: a completion claimed without verification, a session ended on stale evidence, edits outside the working directory, a destructive command, secret material, and a long stretch with no operator input. Reading the output: - Every finding names the step that produced it. Carry that step into your answer — the operator should be able to check the finding