Repository navigation
Conversation
…sion WC-01 of M24: gap analysis of the static Codex/Claude version and model allowlists, the admission-subject, guarantee, evidence/decision and registry contract, drift and expiry rules, bounded versus governed modes, baseline and legacy policy handling, recertification procedure, plan, roadmap section and docs index entry. Documentation only; no allowlist, worker, routing or configuration change. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c0c21925e2
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".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| A host-local registry, because the facts describe the clients installed on one | ||
| host and the Runtime has no host-level database; PostgreSQL authority is | ||
| unaffected. |
There was a problem hiding this comment.
The premise that the Runtime has no host-level database is false: docs/architecture/overview.md:618 documents <runtime-home>/global.sqlite, runtime-paths.ts resolves it, and the daemon opens it at startup. Since this premise is used to reject database persistence and introduce a separate object/manifest store with its own locking, integrity, backup, and audit semantics, the ADR should either explain why the existing database's ownership contract disqualifies it or compare a dedicated/extended SQLite store against the proposed filesystem registry.
AGENTS.md reference: AGENTS.md:L24-L28
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Partially confirmed. The 'no host database' premise was already corrected, but the rationale only rejected reusing global.sqlite and never compared a dedicated SQLite store against AGENTS.md (authoritative state in SQLite behind the Runtime). bd832fc rewrites the Alternatives and records the choice as a new open owner decision O6 that blocks acceptance and WC-02. Left open on purpose: not settled by this PR.
Narrow the self-certification claim to what the Runtime provides and state the same-account limit; specify legacy_policy as a closed source-shipped predicate with its own admission rule; exempt baseline records from the sufficiency table and forbid registry decisions with baseline or legacy bases; define the full admission step order, closed reason set and reason-to-public-code table; keep the per-dispatch adapter checks; make supersession single-active; retain every manifest and record the last accepted generation; reword harness provenance as a recorded claim; fix facts, naming and the dependency graph. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
…cedence Add a model-admissibility step ahead of per-dispatch observations so the public error codes and the current model-before-inspect order are preserved, with an exact reason-to-code table; give exact decisions (active or revoked) precedence over the legacy predicate and never fall back after a revocation; treat an absent registry as empty only before any generation was accepted; correct the pre-release clause of the Claude predicate; keep the execution-time Codex login re-admission; refuse certified decisions with unobservable dependencies until O1 is decided. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
Extend the model-admissibility step to the run's persisted parameters so that today's supportsModel refusals (output-token cap, unsupported reasoning effort) remain refusals; add the parameter clause to the claude-legacy-floor@1 predicate and parameters to the differential matrix; name one reason for the explicit-requirement case; specify the state reset as planned, approved and audited, and add it to the WC-02 command shapes. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
After a record matches in step 4, require the run's parameters to lie within that record's own parameter scope so one subject's scope cannot admit another's runs; state the parameter scope of the Codex baseline import. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
…external review Refuse the snapshot when a committed manifest or revocation is missing or invalid, honor only revocations listed by committed manifests, make manifests cumulative for revocations, check the legacy predicate record for revocation, select the single active exact decision so a superseded predecessor cannot block its successor, enforce notBefore with a typed reason, and make the implicit bounded requirement set one named set that includes the routed-model guarantee. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
…sion An invalid decision listed by a committed manifest now blocks its subject instead of letting the legacy predicate admit it; a decision named by a committed revocation is revoked even if still listed as active. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
Limit rollback and tampering detection to inconsistencies while the recorded state is intact and name the independently protected anchor as M10 scope; scope the "new version stays refused" statement to subjects needing an exact certification and describe the Claude legacy exception; make blocked subjects explicit in model admissibility; record the WC-02 gate on the blocked-subject scope. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
… parity Bind claude-legacy-floor@1 to the source-shipped configuration profile, output protocol and supported platforms; pass explicit requirements to admit separately and derive the mode-implied set inside it so the legacy exception tests for empty explicit requirements; list drift, validity-window and registry-integrity refusals as expected differences of the shadow comparison; scope the blocked-subject rule to the current manifest's active set and forbid dropping an active decision without a revocation. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
…ock, order model first Take the subject used for blocking, matching and revocation from the manifest binding so a corrupted decision cannot move its subject to the legacy predicate; make registration, revocation and reset share one publication exclusion with the planned generation rechecked under it; run model and run-parameter admissibility before platform and mode so unsupported models keep WORKER_MODEL_UNSUPPORTED on every platform; clarify bundle-citing decisions and that re-pinning the legacy predicate needs owner approval. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
…nd storage rationale gaps - A reload that meets a committed manifest or revocation it cannot verify fails closed; the previously loaded snapshot stays in force only when no newer committed object was read, so a revoked decision cannot be readmitted. - admit takes an explicit dispatchContext (persisted run parameters, per-dispatch observation outcomes, evaluation timestamp) so steps 1, 3 and 5 have no hidden state. - Correct the rationale: a host-level global.sqlite exists, but it is the reusable-memory store and is deliberately not reused for admission facts. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
…y baseline checks Re-evaluate ADR-0037 against the Codex review threads on PR #184. Compare the file registry with the existing global.sqlite and a dedicated host-level SQLite store against AGENTS.md, and leave the choice to the owner as O6. State that the deterministic sufficiency checks govern registered decisions while source-shipped baselines follow decisions 9 and 10, and that a baseline declares its metadata dependencies as unobservable. Claude-Session: https://claude.ai/code/session_01Q21NUKiVwrCvpZ5YgzjM5W
…O6 options A failed reload no longer keeps the previously loaded snapshot in force: a state file that records no newer generation cannot prove that no newer revocation was published (a writer can publish the pointer and stop before the state is updated). Any reload that does not end in a fully verified snapshot refuses every subject with registry_unreadable until a later reload verifies. A pointer ahead of the recorded state is accepted only after the whole chain verifies; an older pointer is a rollback. Persistence stays open decision O6; a non-normative comparison of the file and dedicated-SQLite options is added. Claude-Session: https://claude.ai/code/session_01FKM37BV4LS8Rc8ny5GCwXy
… advanced A pointer ahead of the recorded state is accepted only after chain verification; if the state then cannot be durably advanced, a later pointer rollback would match the stale state and drop a revocation. The reload now fails closed with registry_unreadable. AGENTS.md citations in the ADR and the O6 comparison say "project state" as the rule is worded. Claude-Session: https://claude.ai/code/session_01FKM37BV4LS8Rc8ny5GCwXy
… mode_unsupported reachable Review hardening. Ancestry of the recorded manifest validates against rollback only; admission uses the snapshot built from the generation the pointer names. Step 1 no longer matches a mode without a governed contract, so a model lookup cannot mask mode_unsupported. The O6 comparison states the anchor update order and the verified-successor recovery instead of claiming the crash window disappears. Claude-Session: https://claude.ai/code/session_01FKM37BV4LS8Rc8ny5GCwXy
…rsisted model Review hardening. admitModel runs before any observation so an unsupported model still launches no process, probe or login read; admit repeats the step. Unrouted and historical runs have no exact subject and are admitted only through the legacy predicate on the resolved model, with provenance basis legacy_policy, and join the WC-05 differential matrix. Claude-Session: https://claude.ai/code/session_01FKM37BV4LS8Rc8ny5GCwXy
|
Update for head 8ac0ae9 (docs only, ADR-0037 stays Proposed): ChatGPT P1 confirmed and fixed. A failed reload never keeps the previous snapshot; admission uses only the snapshot of the pointed-to generation; a pointer ahead of the state is accepted only after chain verification and a durable state advance, otherwise |
avalla
left a comment
avalla
left a comment
There was a problem hiding this comment.
Independent ChatGPT ADR review — HEAD 8ac0ae9 (static source review, not Codex CLI)
Technical verdict: no additional confirmed P1/P2 defect found in the reviewed revised contract. Revisited the registry loading/revocation invariants, admission preflight and ordering, baseline imports, and bounded legacy Claude exception. The prior fail-closed contradiction is addressed: a failed reload refuses all new admissions, and a prior snapshot cannot silently remain authoritative. The model preflight is explicitly staged ahead of launching a worker.
Outstanding design gate: O6 (content-addressed files vs dedicated SQLite registry) remains undecided and the corresponding review thread intentionally open. The comparison is non-normative. Do not mark ADR-0037 Accepted or start WC-02 until the owner selects the persistence model and the normative ADR is reconciled. O1–O5 should also be recorded as explicit owner decisions rather than inferred from recommendations.
Evidence: GitHub workflows on this HEAD are green; no Bun tests were executed by this reviewer. The PR description's consolidation note names an earlier HEAD and should be refreshed.
Gate: This is a static ChatGPT review, not a successful codex review --base origin/main. It does not replace the configured external-review gate.
… rule - The plan's ASCII graph attached WC-06 to the bus that WC-02 also feeds, which contradicted the dependency table (WC-06 depends on WC-03 only). Replace it with an adjacency list identical to the table. - ADR-0037 listed the six *_mismatch values as reasons and also used them as dimensions of no_matching_audit, without saying which one selects the public code. State that the code follows the primary reason, so a known model with an unaudited version stays WORKER_UNAVAILABLE as today. - Realign the open-decisions table, which the O6 row had left misformatted. ADR-0037 stays Proposed; O1-O6 stay open for the owner. Claude-Session: https://claude.ai/code/session_01FKM37BV4LS8Rc8ny5GCwXy
|
Hardening cycle closed on |
Summary
WC-01 of M24 (Worker Compatibility & Capability Admission): the gap analysis and the contract that will replace the static Codex/Claude version and model allowlists with exactly-bound, evidence-backed, owner-certified admission. Documentation only. Proposes ADR-0037; nothing is implemented and no allowlist, worker, routing or local configuration changes.
Task
99b23a02-59a3-4a5a-b82e-bc21874299f3— WC-01 Capability Admission Contract (milestone M24, requirements WC-R01–WC-R14)Scope
docs/adr/ADR-0037-…(Proposed),docs/development/worker-compatibility-admission.md, an M24 section indocs/development/roadmap.md, an entry indocs/README.md. Runtime records (14 requirement descriptions and titles, 6 task descriptions, 7 task-requirement links) were aligned outside Git.Dependencies
origin/main(c3fdb25), not stacked.Design notes
Normative contract is the ADR; the plan does not restate it. Key choices, all following the owner's decisions of 2026-10-16 (D1–D7, recorded in the ADR):
baseline_importfor the current Codex combination,legacy_policyfor the Claude floor) reproduce today's behavior without being presented as certification, introduced in shadow mode with a differential test.Open decisions O1–O5 for the owner are listed in the ADR (unobservable metadata dependencies, reading of the legacy exception, sufficiency table,
local_operatorapproval class until M10, catalogue freeze).Verification
bun run checkbunx prettier --checkon the four filesgit diff --check origin/main..HEADbun run checkReview
Limitations and follow-ups
deliverypipeline is enforced, and starting a run would bind the task to stage performers I cannot be assigned to; the task-delivery gates were followed instead. Owner to decide if a run is wanted.Merge
Do not merge without explicit authorization. READY FOR MERGE is not DONE:
post-merge verification is still required.
🤖 Generated with Claude Code
https://claude.ai/code/session_016ZEHMaJc8oY8hS3FUSQw1t
Hardening update (2026-10-19) - HEAD
d9c7f4dDocumentation only; ADR-0037 stays Proposed; O1-O6 stay open for the owner; nothing is implemented.
*_mismatchvalues as reasons and also used them as dimensions ofno_matching_audit: it now states that the public code follows the primary reason (a known model with an unaudited version staysWORKER_UNAVAILABLE, as today).Evidence:
bun run checkexit 0 ond9c7f4d(3470 passed, 242 skipped); GitHub CI 13/13 green. Independent read-only review of the delta and the whole PR: no confirmed findings; constants checked against the code (auditedCodexVersions, Claude floor,run.tsgate,global.sqlitedocs) and links resolved.The
global.sqlitereview thread stays open on purpose: it is open decision O6, owned by the owner (see the non-normative comparison).External review (Codex): not satisfied on the final head. The Codex CLI and the GitHub connector are unavailable (usage limit); no other review stands in for it. The independent review below is by a subagent of the same model family, not Codex.