Skip to content

fix(agent): account for repair signals delivered at boundaries - #453

Draft
zhengdaqi wants to merge 3 commits into
codex/convergence-v24r11-source-apply-handofffrom
codex/convergence-v24r12-boundary-signal-accounting
Draft

fix(agent): account for repair signals delivered at boundaries#453
zhengdaqi wants to merge 3 commits into
codex/convergence-v24r11-source-apply-handofffrom
codex/convergence-v24r12-boundary-signal-accounting

Conversation

@zhengdaqi

Copy link
Copy Markdown
Collaborator

Summary

V24R12 fixes a domain-neutral Progress Lease accounting defect isolated by the V24R11 Case 09 product gate. When an applied boundary and a new opaque repair signal occur in the same PreModelRequest, the model request already receives that signal; Lease now records that delivery so a strictly later preparation signal can receive the existing bounded apply turn.

This is stacked on #452. Keep #451, #452, and this PR unmerged until the new V24R12 product campaign passes Gate 0, paired smoke, Case 05, and full Case 09.

Changes

  • Record repairAdvanced and repairPreparationAdvanced as delivered in the existing boundary_grace transition instead of resetting both delivery markers.
  • Add fail-first Lease counterexamples for boundary-co-delivered feedback, later preparation, and no-replay co-delivery.
  • Calibrate the AgentLoop runtime regression to the real ordering: feedback appears in the forced-boundary request, preparation in the next request, then semantic progress.
  • Add reproducible preserved-session replay and disposable legal-repair projection evidence.

Boundaries

  • No Legal Coverage production code or validator rule changed.
  • No decision type, retry, model request, threshold, deadline, prompt, Router/Memory setting, model, Skill, corpus, transaction, receipt, or completion authority changed.
  • Lease remains domain-neutral and compares opaque monotonic ordinals only.
  • Limits remain maxInitialStagnantObservations=8 and maxStagnantObservations=2 in the product gate.

QA & Evidence

  • Counterexample-first Lease test: V24R11 failed only with fail_closed instead of repair_preparation_grace; after the two-assignment fix, 25/25 passed.
  • AgentLoop ordering: baseline -> stagnant -> boundary_grace -> repair_preparation_grace -> renewed; combined Core suite 33/33 passed.
  • Preserved V24R11 trajectory: all first 20 of 21 decisions unchanged; only the terminal decision changes from fail_closed to repair_preparation_grace.
  • Disposable legal projection: the blocked repair remains validated: true, exact apply succeeds, a durable receipt is written, state advances, input tree remains unchanged, and the frozen run is not mutated.
  • Focused Legal/Gateway/Lease/O1: 87/87 passed.
  • Complete repository: 311/311 passed.
  • Patch hygiene: git diff --check passed.

Reviewer-readable evidence: .omo/evidence/20260727-v24r12-boundary-signal-accounting/QA.md.

Risks & Residuals

Local tests and preserved-session replay prove the isolated state transition and surrounding regressions. They do not prove the production model will complete Case 09 after receiving the newly available apply turn. That remains the blocking product gate; V25 and the 85-case campaign are not authorized yet.

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