Skip to content

Complete experimental Task IR P1 acceptance and recovery - #71

Draft
Aram164 wants to merge 5 commits into
codex/system-improvements-20261027from
codex/task-ir-p1-experiment
Draft

Aram164 wants to merge 5 commits into
codex/system-improvements-20261027from
codex/task-ir-p1-experiment

Conversation

@Aram164

@Aram164 Aram164 commented Oct 27, 2026 •

Copy link
Copy Markdown
Owner

Task IR P1 now returns useful bounded context/material/session results and preserves the exact reviewed route-patch envelope through submission and recovery. A lost response or abrupt process exit is reconciled through the existing gateway replay-only path and strict receipt verifier; missing proof remains uncertain instead of triggering a new write.

The experimental surface remains four operations over existing Core capabilities. Inputs reject undeclared fields, reads bind verified snapshots, incomplete material discovery cannot select a route, and review artifacts bind the task scope, payload, root, session and identities. Preview rejects symlink/hardlink redirection and derives set-like no-op results from structured check evidence. Operator sealing does not authenticate human approval. Legacy previews without retained original bytes require a new reviewed request; interrupted requests are never resealed.

This draft is stacked on codex/system-improvements-20261027 (PR #64). It preserves the original P1 commits and the separate verification repair. It adds no operation catalogue, approval protocol, journal, requirement/session entity, plan.compose, P2 benchmark, learner-data mutation or installation.

Validation:

  • 73 Task IR tests, including authored-runtime integration, review/payload/target/guard attacks, real committed-response loss, abrupt exits, exact replay, idempotency conflicts, field preservation and path-safety regressions.
  • 276 owning runner/report/runtime/seal/paired-recovery tests and 94 architecture tests passed.
  • Ruff/reachability, strict mypy, import boundaries, system map, validator and warning baseline passed. UI npm run check passed and both worktrees were clean.
  • Separate agent review found no unresolved blockers; its implementation/test/documentation diff digest matches the committed tree. This is agent review, not human approval.
  • The canonical error-vocabulary checks also passed (3 tests); recovery compares imported domain constants.
  • Full hosted gates passed for Core e895758126e332f8a8a9d7b34d4f1916337cee45 and UI 3c9a5f664c147790c490fae0a256e7764b4b9157: exact-pair release run 40599114508 and PR validation run 40599065662 both concluded success with 8,896 tests, 0 failures, 0 errors, 9 skipped (8,887 passed) and static gates green. Pair artifact learningos-pair-e895758126e3-3c9a5f664c14 receipt was downloaded and independently verified (release_pair_receipt.py verify returned ok: true for the exact Core/UI pair). See P1-ACCEPTANCE-REPORT.md for commits, counts, skips and receipt.

Addresses #65, #66, #67, #68, #69 and #70. Actual agent round-trip/token savings remain a P2 measurement question. No merge is requested.

Four typed operations compile into existing machinery: reads become one
agent plan, route.patch preview/apply rides the reviewed seal flow.
No new los verbs, approval format, or transaction engine.
Covers the one writable workflow end to end on disposable fixtures,
plus rejection, purity, stale, unchanged, and uncertain-dispatch
contracts; registers the file in its group and shard leg.
Close read shapes and retain bounded context, material provenance and runtime
session proposals under verified read snapshots. Bind reviewed artifacts to
the task and submit the original envelope bytes; reconcile interrupted writes
through replay-only and strict receipt verification. Cover authored-runtime,
scope/guard/review substitution, interruption, replay and persistence cases.

LOS-CI-Bootstrap: codex/task-ir-p1-experiment
Hosted A14 caught a literal IDEMPOTENCY_CONFLICT comparison. Import it and
UNCONFIRMED from the existing domain vocabulary without changing recovery
behavior or weakening the architecture ratchet.

LOS-CI-Bootstrap: codex/task-ir-p1-experiment
Aram164 pushed a commit that referenced this pull request Oct 28, 2026
Final decision record: P1 retired as a production candidate, PR #71 left unmerged, the plan runner remains the supported interface. Stage conclusions, preserved references and artifact hashes, limitations, and the two candidate runner improvements are recorded in one place. Documentation only.

LOS-CI-Bootstrap: codex/task-ir-p3-feasibility-20261028
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