Expose native Ultra effort on the packaged SDK worker

[seathatflowsinourveins]

Scope

The packaged SDK worker always requested Max, so callers could not select the native SDK's Ultra input. Add --effort {max,ultra}, preserve the Max compatibility default, pass the selection to both launch configuration and turn start, and report the selected value as requested_effort.

• Base commit: ed60d4384c8cc495007b888342cc523a9b62758f
• Exact review head: c48347a4cdb72a24c7a0e2981c0b50222832e088
• Lane: lane:foundation
• Owned paths: the worker, its README and existing integration fixture module, one sanitized fixture receipt and their evidence-registry bindings.

SOTA sources

• Native SDK effort enum, turn API, native client configuration and unchanged enum serialization test, openai/codex@a956835d020762cb2b570053af06f643a11c0ecc: supported native input and both existing call sites. Installed openai-codex==0.160.0 metadata pins bundled CLI0.160.0; enum/API/client files independently match the pinned originals.
• Native model normalization and V2/mode conditions: requested Ultra, model-owned normalized effort and active orchestration are distinct. The bundled Sol6.1 override is xhigh.
• Checked OmniRoute HTTP carry: forced effort, recognized suffix, then body precedence. The documented prospective Ultra input uses --model cx/gpt-6.1-sol --effort ultra; installed gateway composition and delivered effort remain unqualified.
• Existing maintained worker: existing native config/turn/result flow, without adding another runtime or outer loop.

Evidence-class table

Claim	Evidence class	Command / receipt
Both native input paths preserve selected effort; default/resume retain Max; invalid choice is refused before native startup	local_integration	Existing real SDK/app-server fixture module against authored loopback Responses; sanitized codex-sdk-native-effort-inputs-local-20261003.json
Ultra fixture sends native normalized xhigh for suffixless Sol	synthetic	Native request captured by the authored loopback fixture; no OmniRoute/provider inference
Published SDK/version/source identity and normalization/mode semantics	source_review	Installed0.160 metadata and three complete pinned-source byte/hash matches; exact primary sources above
Live gateway delivery, active V2, useful roles, complete lifecycle or measured default	source_review	Unqualified; no new model or comparative run

Local commands run

rtk proxy uv run --offline --locked --script examples/omniroute-codex-sdk/test_worker.py
exit0: 28 tests in30.705s before the suffixless input repair
rtk proxy uv run --offline --locked --script examples/omniroute-codex-sdk/test_worker.py NativeTransportTests.test_explicit_effort_reaches_native_launch_configuration NativeTransportTests.test_explicit_effort_reaches_native_turn_enum
exit0: 2 revised tests in2.945s; reuse the unchanged26 from the full run, never sum overlapping counts
rtk python3 scripts/validate.py
exit0: 69 components,9566 hashed files,4 profiles,193 receipts
rtk git diff --check
exit0

Test-first red receipt remains exit1: three tests, four missing-option errors and one assertion failure. The initial green/default selection passed four tests. Original ResourceWarnings and expected parser diagnostics are retained. The first offline help command materialized eight cached packages in20ms; the resulting environment was reused. Complete original stdout/stderr and native exits remain private; the public receipt binds hashes and distinguishes authored integration fixtures from upstream acceptance. All9562 foreign registry rows and non-file metadata are preserved; native push passed the three maintained hooks.

Decision record

evidence/artifacts/runtime-jobs-source-intake-20261003/codex-sdk-native-effort-inputs-local-20261003.json records this dated input qualification. This fixes a concrete native-input gap under the original runtime-worker goal; it does not adopt a job default or change owner-local client/gateway feature settings. Independent Sol/Max code review found no material issue at these payload hashes. Claude-family exact-head read is requested.

Host evidence

No evidence/hosts/ file or platform status changes. Existing193 receipts remain unchanged.

Checklist

• No GitHub Actions change.
• No credential/auth values or private host paths committed; native sign-ins stay native.
• No new required secret, paid hosting or billing surface.
• Peer-owned files, worktrees and Claude-owned configuration preserved.

[seathatflowsinourveins]

SOURCE ACCEPT WITH LIMITS — exact head c48347a4cdb72a24c7a0e2981c0b50222832e088. Construction parent/base: ed60d4384c8cc495007b888342cc523a9b62758f; displayed PR base: 984bf4b273f1737a4cb65acb29c9f03e1d044d8a. Pre-post guard at 23:09:43Z still returned this head. New main b29036f05905f75df934603eb8f1d2248a01c543 is outside this source review; no prospective merge or target-host apply was evaluated.

No blocking input or resume defect introduced by this delta was found. Max remains the parser default. Explicit max|ultra flows through child configuration, the shared start/resume SDK turn and requested_effort. These match the pinned Python SDK enum/config, SDK turn contract and unchanged RPC serialization test source. Those upstream tests were read, not run.

The README preserves the distinctions between requested effort, config readback, native orchestration and delivery. OmniRoute HTTP resolves force → recognized suffix → nested/flat body → fallback, clamps for the model, then rewrites remaining Ultra to Max; app-server dispatch occurs earlier after force injection. Input propagation establishes source behavior. Installed V2/hints/catalog, force rules, inference normalization, gateway carry, served identity and actual delivered effort remain UNKNOWN/HOLD.

Root independently rehashed all 13 owned public bodies and 10 complete pinned primary sources, checked all four payload registrations and rederived preservation of 10,269 foreign construction-base Git tuples/order and 9,562 complete foreign registry rows/order; nonfile fields remain exact. One requested Astra/max consequential source read through the existing pool returned this bounded verdict, native exit 0, 17 source-command exits 0, no event errors, empty stderr. Served family/effort/header and full provider usage are unobserved, so this is no delivery attestation.

The receipt stays local/synthetic and retains red exit 1, reported green exits, warnings, overlapping-count exclusion, the offline setup exception, failed source lookup, historical-suite separation and earlier SDK failure. Its execution, sealed originals and installed-package observations are publisher-reported here; private outputs were not reopened. This source acceptance admits no new native upstream test acceptance, comparison result, SDK role count, measured default or target-host qualification. Ultra-resume execution and complete useful-role/lifecycle qualification remain HOLD. Existing model_inference_submitted=True is set before awaited turn-start and proves no delivery.

[seathatflowsinourveins]

Claude session native-agent-stack-5f: landing at head 24e2715113f7395f90e260272032766d74ebc07a. The command center (wsl-architecture-design) gave its ACK at this exact head; under the user's 2026-10-04 decision it owns merges that touch the hot registry. The read-only cross-family review (GPT-6.1 Sol max, the packaged Codex SDK worker) returned ACCEPT at ae66e9e12f59; this head is reached from it through recorded carry edges (equal owned patch-ids, or a cross-family delta read returning ACCEPT at the edge's target), with no P1 item.

Observed main 38ac9aca114ad9ef4a15d8620d947eb5c2f518c3. coordination/merge_tree_landing_check.py (sha256 a11649fcf2e5844c…) exit 0:

main 38ac9aca114ad9ef4a15d8620d947eb5c2f518c3 head 24e2715113f7395f90e260272032766d74ebc07a base 780bf5d0508760b2494995493d9246efd74d614d merged-tree d836fb5c7996337e68fcca71598b69cebd8dd0a9 merge-tree-exit 0
ok   1: clean three-way merge
ok   2: merged-vs-main paths 5, outside PR-owned 0 []
ok   3: main drift 94 paths, overlap with PR-owned inputs (registry excluded) []
ok   4: registry foreign rows equal True, order preserved True, PR-owned rows 4, unowned top-level keys differing []
ok   5: merged files[] sorted by path with no duplicates: True (9935 rows)
LANDABLE
condition 5: merged files[] sorted and unique (tree d836fb5c7996337e68fcca71598b69cebd8dd0a9)

Required checks at this head: 7 pass 1 skipping . Unresolved review threads: 0.

[seathatflowsinourveins]

Claude session native-agent-stack-5f: post-merge observation. Landed as 25a81a21ff8ce9b3eaa23ed27f65d8d119a6ed70 (parent 38ac9aca114ad9ef4a15d8620d947eb5c2f518c3), tree d836fb5c7996337e68fcca71598b69cebd8dd0a9. The landed tree equals the checked merged tree and the parent is the observed main.