Skip to content

feat: add inactive bounded KV pilot launcher - #276

Merged
Dong1017 merged 10 commits into
mainfrom
feat/kv-live-launcher-inactive
Oct 31, 2026
Merged

Dong1017 merged 10 commits into
mainfrom
feat/kv-live-launcher-inactive

Conversation

@Dong1017

@Dong1017 Dong1017 commented Oct 31, 2026 •

Copy link
Copy Markdown
Collaborator

Bounded inactive delivery

Refs #248; builds on merged #275. This PR delivers only the ten-file inactive launcher, not a paid run or an efficacy result.

  • Validate the exact public freeze, run/source identities and hosted actor qualification; repeat isolation and tool preflight on the execution host.
  • Reuse durable full-pair admission and one-grant/one-HTTP-attempt transport. Unknown usage or abnormal worker completion stops further requests without refunding reservations.
  • Retain immutable source revisions, linked feedback and bounded first-run semantic JSON/executor outcomes, ABBA/quality data, public freeze and preflight evidence—including partial or interrupted states.
  • Grade and project in separate secret-free steps. The always() projection reads existing artifacts only; it never reruns the candidate or model.

Inactive by design: the locator is disabled and credential bindings are absent. This delivery made no provider request, live ledger mutation or dispatch. PR #275 remains the sole prospective charge identity; this PR creates no additional spending allowance. Execution still needs reviewed operational configuration, refreshed freeze/accounting, exact qualification and separately authorized activation. Broader #248 acceptance remains open.

Verified content, tests and evidence limits

Published head 68fa4f405c6b010c0626a4e4c9d000906673e739 and clean local tested commit 1c2f89cfb47efcc9b88868f35c29efdfae616ff6 share tree 059d3023fa5f02513f2bb0dcbb7f1a50229f5c48. Against main 19c3df8311f55ab9efbcd6fa60e29a4bbce95dd1 (the #275 squash merge), GitHub compare confirms exactly ten launcher files, +2,808/-0. A normal two-parent merge incorporated that main ancestry without changing the tested tree; generic model/harness fixes belong to #275.

  • Frozen-head full local aggregate: 3,169 passed, 12 skipped, 284 subtests passed, one expected duplicate-ZIP fixture warning; 455.50 seconds with no source changes during the run.
  • Complete affected selection: 480 passed. Whole-tree Ruff and diff checks passed.
  • Independent review passed 137 focused tests for the evidence-retention patch; final integration rechecked all five projection-file hashes, the unchanged worker and both independently reviewed shared fixes. No open substantive finding for this inactive scope.
  • Current-head Ready run 41154288295 completed successfully: both Python 3.10/3.11 full checks, both installed-wheel checks and the model-free task/actor qualification job. Its test-merge commit 01e2d6f35b67cb32c2cf3bcaec1664984f1ef037 has the same tree above. Draft lint and knowledge-declaration policy also passed.
  • The native code-review summary completed on this exact head with no findings; native security review was not run. Independent source review remains separate.
  • The earlier opened-event Ready run was superseded/cancelled when this draft became Ready, not counted as PASS. An earlier concurrent local aggregate lifecycle failure remains recorded; its isolated rerun and subsequent clean aggregates passed. None of these software checks establishes paid readiness or efficacy.

Retention was checked against all 77 frozen semantic expected outputs (including large integers) and 34 final benchmark/quality outputs. Allowed task JSON and numeric evidence stay within fixed file/size bounds. Invalid arbitrary text is not exported as a log; bounded observation metadata records byte count/hash, failure reason and truncation or unavailability instead. Missing/malformed/interrupted evidence never becomes PASS; death before an executor returns leaves its outcome unknown. This is not a claim to retain every raw byte or survive complete host/artifact loss.

The new task/actor artifacts are retained locally with exact run/head/tree metadata. Outer/nested archive hashes, fixed member inventories and unchanged qualifier dependencies were checked; prior independent raw audits are reused only for the unchanged source/protocol. This is not a fresh raw semantic/timing recomputation of run 41154288295.

Public projection excludes arbitrary logs, environment contents, provider responses and credentials. The frozen source binds the sandbox profile; hosted/current runtime inventory hashes are retained separately, with independent current-host checks rather than incidental inventory equality. Task-fixture anchors remain explicit; campaign/charge identities come from the future verified plan. No raw ZIP is published by this PR.

Authorship

Actual configured implementation and independent review: Codex / gpt-6-astra / xhigh. Publication preserves the tested tree through an API-created wrapper; commit identity, independent review, CI and operational acceptance remain distinct.

{"schema_version":1,"head_sha":"68fa4f405c6b010c0626a4e4c9d000906673e739","human_only":false,"material_ai_authors":[{"tool":"Codex (inactive KV launcher and bounded evidence retention)","model_id":"gpt-6-astra","effort":"xhigh","role":"author"}]}

Refs #248. Publish reviewed software-only tree 302ada9 from local commit 6cfbc5ea1782c4d697db74b71fdf408abc4497fa.

Implementation and independent review: Codex, gpt-6-astra, xhigh. This API-created publication wrapper preserves tested file bytes; local commit identity and hosted qualification are distinct.

No provider calls, secret binding or efficacy acceptance.
Publish reviewed local 5e4a78084c85425e6015e804387e526b7c14b311 tree unchanged. Read raw authoritative RECORD; retain actual anchored source/editable package trees and independent installed copies. Fail closed on missing, ambiguous or unsafe nodes.

Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh]

Prior hosted qualification remains bound to its original tree. New exact-head qualification required.
Retain first successful task-read request index so later redundant reads do not reject valid edits, checks or benchmarks. Preserve rejection of a first read and action in the same response; add both positive and negative regressions.

Local reviewed commit f9f76365ead9bcdb339488974f4862414b28679e; exact same tree. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh].

Addresses native review discussion_r4426397229; fresh exact-head qualification remains required.
Stack ten launcher-only files on reviewed adapter. Locator disabled and workflow credential bindings absent. No execution or paid readiness is claimed.

Exact local combined tree from 7d7f3d8cdf3f35b5b456d41f0648c81d10e478e3. Implementation and independent review configured Codex / gpt-6-astra / xhigh.
Only accept completed status and an actual integer zero exit code. Fail closed on positive exits, malformed codes and conflicting states even with completed worker JSON. Preserve consumed-request accounting and add adversarial controls.

Reviewed local commit ccf307ca44f93093ce6ca2db22346885d31aabd6; published tree identical. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh].

Addresses discussion_r4426926161. Prior hosted receipts remain bound to prior trees; new current-head qualification required.
Keep latest adapter head as ancestor and preserve exactly ten launcher-only files. Same tree as local 57e0b88491502a1cccdd25be9bb4de6d5b329950; 320 affected tests passed.

Locator disabled, no credentials or runtime activation. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh].
Atomically persist only status, returncode and elapsed_seconds after the executor returns, before timeout/marker/JSON branches. Preserve existing oracle, metric, record and verdict semantics; no argv, environment or error text retained.

Reviewed local commit 6bf845295850b4e0d671594cc690a0ac5cba4afc; published tree identical. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh].

Preserve prior receipts and require fresh exact-tree qualification.
Preserve existing literal substring checks for request and response while representing the key in the same canonical JSON encoding. Cover allowed quote/backslash shapes using real transport plus mocked HTTP; preserve ordinary-key protection, fixed errors and unknown-usage handling.

No observed real-service disclosure, credential access or worker secondary scanner. Local reviewed commit f53f0c007311ed73858c806004cda94ab122b6a2, identical tree. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh].
Merge latest reviewed adapter dependency. Add fixed data-only projection of semantic execution outcomes, raw task JSON, timing/quality records and public preflight/freeze evidence. Preserve interrupted/missing/redacted states without rerunning candidates or exporting arbitrary logs.

Exact tree from clean local 1c2f89cfb47efcc9b88868f35c29efdfae616ff6; 3169 aggregate tests and 480 affected tests passed. Locator disabled; no secret binding or activation. Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh].
Preserve the independently reviewed and tested launcher tree while incorporating the merged adapter ancestry. No source bytes, authorization locator, credentials or execution behavior changed.

Agent-assisted-by: Codex / gpt-6-astra [effort=xhigh]
Signed-off-by: Codex <codex@openai.com>
@Dong1017
Dong1017 marked this pull request as ready for review October 31, 2026 01:13
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Oct 31, 2026 •

Copy link
Copy Markdown

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-10-31T01:15:32.279998Z 68fa4f4 Draft marked ready
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

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