Measure ten-stage English delegation against native Codex shell - #14
Merged
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Expand the Codex harness from one small task to a continuous ten-stage incident packet, with efficient native-shell, ten fresh English-worker, and two grouped-worker strategies. Capture timestamped host events, reopen all saved outputs, replay scripts on changed inputs, and reject savings for failed or source/fixture-mismatched runs.
Real Codex pilot: normal shell completed 10/10 stages on both inputs in 94.0 seconds. Ten handoffs took 200.4 seconds and two handoffs took 156.6 seconds; both failed all stage checks. The report preserves these failures and separates token counts, tool/worker timing, and conditional arithmetic from successful-task savings. Private transcripts only informed synthetic patterns; raw evidence remains local.
Validation: compileall; 16 core/comparison invariant tests; three serial live Codex sessions with the existing FP16 specialist; saved original/alternate outputs and comparison reopened. The release metadata command requires GitHub event context and is left to hosted CI. No model/adapter changes.