Skip to content

Pinned verification commands keep the CI check that proves them - #190

Merged
czpython merged 1 commit into
mainfrom
agent/ENG-825
Aug 6, 2026
Merged

Pinned verification commands keep the CI check that proves them#190
czpython merged 1 commit into
mainfrom
agent/ENG-825

Conversation

@czpython

@czpython czpython commented Aug 6, 2026

Copy link
Copy Markdown
Owner

Building the effective profile from the operator's pinned commands hardcoded ci_check to None on every one, discarding what the profiler detected. The evaluator runs any command that names no check — "A configured command with no CI check name runs locally. Never infer that another check covers it." So a repo pinning its verification in .druks/ship/config.yml ran every command in the sandbox on every evaluation and could never read its own CI, however often it was re-profiled.

VerificationProfile.with_ci_checks pairs each pinned command with the check detected for that same command. Pinning chooses which commands gate a PR; it no longer decides what proves them.

Acceptance criteria

  • AC1: A pinned command matching a detected command carries that command's ci_check in the stored effective profile.
    • Verification: test_pinned_command_keeps_the_check_detected_for_it pins ("pytest", "make e2e") against a profile that detected pytest under Backend / tests, and asserts the stored entries.
  • AC2: A pinned command with no matching detected command carries ci_check: null.
    • Verification: the same test's make e2e entry, plus test_pinned_verification_replaces_the_detected_one and test_skips_the_agent_and_reapplies_the_pin, which pass unchanged.
  • AC3: With no pinned verification, the effective profile is the profiler's own output, unchanged.
    • Verification: test_persists_baseline_and_effective, unchanged.

Closes ENG-825.

🤖 Generated with Claude Code

@czpython
czpython force-pushed the agent/ENG-825 branch 2 times, most recently from 88e3d9a to 33dc2ef Compare August 6, 2026 18:58
Building the effective profile from the operator's pinned commands hardcoded
ci_check to None, discarding what the profiler detected. The evaluator runs any
command that names no check, so a repo pinning its verification ran every one
in the sandbox and could never read its own CI.

Pinning chooses which commands gate a PR; it no longer decides what proves them.
@czpython
czpython merged commit 225ef26 into main Aug 6, 2026
@czpython
czpython deleted the agent/ENG-825 branch August 6, 2026 19:02
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant