Repository navigation
Conversation
Add deterministic random-single initiation and an all-agent completion barrier. Record policy evidence and document the reproducible four-arm experiment procedure.
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Change and reason
Goal Mode always started every agent and accepted the first completion claim. This made it impossible to test whether eager initiation and early termination caused duplicate work or premature close.
This change adds explicit, backward-compatible policies:
--initiation-policy all|random-single--termination-policy first|all-agents--policy-seed <uint32>for deterministicrandom-singleselectionall-agentsaccepts one exact completion claim from each joined agent, then runs a fresh semantic verifier. A failed verifier clears the confirmations and starts another repair round. Hard time, token, count, idle, and owner-close bounds still stop the run.episode.jsonandpolicy.ndjsonrecord the selected starter, starts, confirmations, quorum, ignored or duplicate claims, and verifier resets.The default remains
all + first. Frozen--harness-modeprofiles reject policy overrides. The CLI also rejects completion settings that could bypass the all-agent quorum.The new goal policy ablation runbook gives controlled-English instructions for the 2×2 experiment, artifact audit, randomized arm order, frozen controls, and hidden-grader boundary. It identifies the token threshold as soft because the current runner observes usage after a turn.
Verification
pnpm check:export— 95 allowlisted filespnpm typecheckpnpm test— 502 Vitest tests passed, 1 skipped; 16 process tests passedpnpm buildpnpm test:package— installed-package smoke passedbash -nThe no-model 2×2 coverage is a runner integration with fake Docker and service boundaries. It covers all four arms, exact start behavior, cross-turn all-agent quorum, no early close, verifier failure resets, evidence files, token bounds, and owner close. No model call or paid experiment was run.
Compatibility and release
Existing commands keep the
all + firstbehavior. The change adds fields toepisode.jsonand a newpolicy.ndjsonartifact. It does not change credential formats or require a newer server.A source merge does not publish a package. This PR does not enable auto-merge or release anything.