Skip to content

feat(miner): run the TARGET repo's own test/build commands before opening a PR — stop trusting the agent's self-attestation #8807

Description

@JSONbored

Context (AMS adversarial audit)

Nothing independently verifies a coding agent's work before submission: slop-assessment.ts is a static pattern scorer; the only real verification module (engine lint-guard.ts:94-131) is hardcoded to loopover's own monorepo commands AND never passed in production (driver-factory.ts:283 gates on options.lintGuard, which no production caller sets). coding-task-spec.ts:235-259 only TELLS the agent which commands to run. A known-bad PR passes every AMS-side gate.

Fix

Thread detectRepoStack's inferred test/build/lint commands into a real post-edit verification step (target-repo-aware generalization of lint-guard's pattern) that runs before checkSubmissionFreshness; a failure returns the attempt to the iterate loop (bounded) instead of submitting. Tests with a fixture repo whose test command fails.

99%+ patch coverage, branch-counted.

Metadata

Metadata

Assignees

Labels

maintainer-onlyOwner-only work — yields no Gittensor points.

Projects

No projects

Relationships

None yet

Development

No branches or pull requests

Issue actions