Skip to content

S26-E0.5: remove the self-written Codex harness - #280

Closed
michaellchung wants to merge 2 commits into
s25-finalfrom
s26-e05-harness
Closed

michaellchung wants to merge 2 commits into
s25-finalfrom
s26-e05-harness

Conversation

@michaellchung

Copy link
Copy Markdown
Contributor

Removes the legacy self-written Codex harness (task loading, sandbox, tool execution, grading, Codex runtime; ~29k lines incl. tests). The Harbor + recording proxy + boat path replaces it.

  • Moved still-used pieces: Responses event formatting into the recording proxy, refresh_segment_counts into yeto/rl/compaction.py, SecRLEnv outcome checks into yeto/rl/secrlenv_outcome.py.
  • Launcher rejects the removed Codex/SecRLEnv flags; island-side parsing and ssh_harness eval paths for SecRLEnv are removed.
  • Two progress keys stay (value None) so pre-existing checkpoints still resume.
  • Pre-removal state is tagged s26-pre-harness-removal.

Tests: full suite 6179 passed / 0 failed (remote CPU runner); Harbor, boat, recording proxy and tools/eval tests pass.

🤖 Generated with Claude Code

chowhsu and others added 2 commits October 20, 2026 14:44
Harbor (task loading, sandbox via the boat provider, agent, verifier) plus the
records proxy replace it. Removed: yeto/rl/harness/codex (agent, sandbox
broker, TB2 provider, tool exec, grading, Codex bridge), the Miles codex glue,
SecRLEnv retry, the RLF family, the TB2 reward-env adapter and Modal reward
sandbox, the old eval runners, the launcher Codex bundle/contract path, the
SSH-harness Codex/SecRLEnv daemon path, and their tests.

Kept or moved: the Responses SSE writer into the records proxy,
refresh_segment_counts into yeto.rl.compaction, the SecRLEnv outcome check
into yeto.rl.secrlenv_outcome (trajectory evidence of that kind still
verifies), the CompactionRL launch check into yeto.launch.harness_env.
Removed Codex/SecRLEnv launcher flags are refused. The last commit with the
old harness is tagged s26-pre-harness-removal.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The removed agents made these paths unreachable:
- --codex-backend-profile, --codex-reasoning-effort,
  --codex-exclude-reasoning-from-loss and
  --secrlenv-max-infrastructure-replacements become launcher-only flags (still
  refused); the island no longer parses them or --codex-harness-contract, and
  no longer copies them into miles_args.
- The SSH harness drops its SecRLEnv-only eval path: eval checkpoint
  contract, prepare options, checkpoint deploy, eval-only argv, syncer
  lifecycle evidence, eval collect/verify, and the codex fields of the plan.
  Plans that still carry eval, daemon or Codex fields are refused.
- The island checkpoint config keeps its codex/secrlenv keys as None so
  checkpoints written before the removal still resume.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@michaellchung

Copy link
Copy Markdown
Contributor Author

Closing: all commits of s26-e05-harness (head 9ced3bf4) are already in main (verified with git merge-base --is-ancestor; main = 1f83be1d). The content landed through the stacked merges of #269 / #279 / #292. GitHub cannot retarget this PR to main ("no new commits"), so it is closed rather than merged. The branch is kept.

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.

2 participants