[tzhouam]

Evidence and scope

Two exposed historical Direct/cursor/auto findings missed interprocedural evidence:

• Omni #6371, reviewed at 69211d003019f42578a26c9de3f10c74431323f0, claimed sampling metadata never reached the consumer. The concrete subclass caller supplies it explicitly, and the superclass forwards the kwargs.
• Omni #6477, reviewed at 05916a1fe58f31866c4c29051705269bf9c7279f, proposed unregistering after successful shared-memory unlink. CPython 3.12.13 unlink already unregisters on the POSIX success path. This is version-specific source evidence; the historical deployment interpreter and a runtime duplicate-removal traceback were not established.

Both were fallback Direct outputs, not quality evidence for the originally assigned Strict provider. These public cases informed development and are not sealed holdouts. No claim that their other runtime paths are correct.

Change

Add three targeted items to the canonical DIRECT_REVIEW_CHECKLIST:

• Trace the concrete caller, subclass overrides, explicit keywords, kwargs and superclass forwarding to the actual consumer before an absence claim.
• Inspect version/platform-specific library callees and success/exception ownership; check proposed cleanup for duplicate non-idempotent side effects.
• Keep inaccessible callers/callees or unverified dependency assumptions unresolved, including sparse-checkout limits, while preserving independently demonstrated reachable defects.

Add a normal public-SDK regression and update the corresponding SPEC. No new schema, attestation field, finding filter, semantic completion gate, routing/model configuration, bot edit or deployment change. This targets Direct, separately from #449 cancellation ownership and #450 peer-verifier context.

Delivery verification

Base: 2e8123235d2ee58927f973be16efc2600fa8c398.

Current bot source was checked at JiusiServe/omni-reviewbot@43364408d42e100243e3ab40329f9aa425ef0bc2: its Direct adapter returns plan.to_dict(), the shared Codex prompt embeds the full context JSON, and Cursor calls that same builder with cursor_read_only=True. Updating only the standalone Cursor instructions would miss this path.

Six focused offline checks passed using selected production-function ASTs with resource discovery, containment and tracing isolated by doubles:

• Canonical raw bundle.
• SDK plan method and real dataclass serializer.
• The new committed SDK regression function body.
• Bot Direct adapter through shared prompt JSON.
• Cursor prompt construction stopped by a sentinel before any process launch.
• Empty-context negative transport control.

With the original provider source, the five guidance checks fail and the empty-context control passes. This establishes prompt delivery in the isolated functions, not a full-package SDK execution or model compliance. Python 3.12.14 compilation, whitespace checks, unchanged production AST outside the checklist, and scoped SPEC mapping/date checks also pass.

The sparse local snapshot lacks pytest and full-package dependencies. Normal-import execution of the committed regression, full repository tests, installed-wheel checks and whole-tree/history SPEC validation were not run locally; exact-head CI must supply those results. Standalone host contract tests were inspected and impose no checklist-copy requirement.

No model/replay calls, user's Codex execution, live review assignment, retry job, GPU run, merge or deployment was invoked. No measured precision/recall or production quality improvement is claimed. Keep draft pending exact-head CI and a new independent dot review.