TaskRun is the auditable result of one AgentTaskManifest execution. It is separate from Library
Evolution RunReport v1.0: the latter explains a Data Pack state; TaskRun records an agent,
patch-policy decision, graders and runtime/visual evidence.
A TaskRun binds:
- task manifest bytes, taskId, revision and source contract tree digest;
- independent runner snapshots before/final and
sourceUnchanged; - complete source/candidate snapshots including
.git/**, before/after untrusted stages and at finalization; - digest-verified private staging copies of grader and tool trees plus original/staged integrity checks;
- any post-agent task/grader/tool/candidate/source integrity drift as a non-passing policy violation;
- provider/model metadata, argv, duration, exit/signal/timeout, stdout/stderr digests;
- agent patch bytes and PatchAudit;
- candidate tree digest and every GradeReport;
- every prompt, patch, audit, DOM, screenshot, diff and evidence artifact by SHA-256;
secondaryErrorsfor post-grade integrity/artifact problems that must not erase an already observed candidate failure;- explicit capabilities and missing isolation guarantees.
Verdicts are passed, agent-failed, patch-invalid, policy-violation, grader-failed, input-error, or infra-error. Provider authentication/unavailability and trusted dependency preflight failures are infrastructure failures. Candidate/subject timeout, malformed evidence, browser failure after subject execution, and post-agent integrity drift are completed non-passes and remain in the AI denominator. termination.processExitCode uses 0 for pass, 1 for completed non-pass, 2 for invalid input, and 3 for infrastructure failure.
TaskRun reports can prove what happened in one disposable workspace. They do not authorize promotion of the patch into the source library; promotion is a separate reviewed operation.