The operator requested GPT 6.1 Sol across AutoCoder where appropriate, with high as the normal reasoning default. Medium through xhigh is the usual range; low is mainly for overriding budget or latency needs. High is the default and xhigh is the maximum permitted Codex effort; the operator explicitly excluded max. Future Codex requests normalize max and ultra to xhigh, including unstarted queued work. Explicit allowed effort is retained. Dedicated compatible-provider categorization, Phoenix and embeddings retain their separate contracts. Existing account admission, spending ceilings and automation stops remain effective.
fa642a12110d4617c07613c537a7fb355ee165a4; dev integration: 4152143e89a243b18540b8ce59ce5c3c0ac53734; automatic main integration: 26bb5d34e540a452fd2c30e447420b9b1cebfafc (PR 6793).production_live.4152143e89a243b18540b8ce59ce5c3c0ac53734-4058-1.gitea.app.lan/autocoder/autocoder-engine@sha256:5b93f22eab257a242a8fb6210df15b8f305b881eda6bb3e7e2be94d0d106461b.gitea.app.lan/autocoder/autocoder-engine-frontend@sha256:79dcfbf04364cb682fbe44b4cd9348979aaec822bbd7e85f4317ed07633ace75.Selected established context kubernetes-admin@kubernetes, API https://192.168.1.9:6443, namespace autocoder-engine. Rendered repository-managed configuration, compared environment differences (only requested model values), verified immutable canonical instructions, created and reread two immutable ConfigMaps, then updated only the deployment metadata preparation annotation using resourceVersion compare-and-swap. The configured promotion workflow performed the application rollout. No Secret, RBAC, infrastructure or persistent-data changes were made.
autocoder-env-2fc696e110cbb2943601a774b0750d0b4a00b8881c77344fd96cf1d75e0c4f4b.autocoder-root-9f5f5590186a8c8aa4caa39d37aada9dd9acb8af587144f8a8a0e807cae9797f.68ce537eba1224622a937187f0dfc850c0dd2a417688c2d01ddc7b341e330125.3a495837d58a9a04e87de96d827be392c5ee27dcd3607064c28146dc95eae0c9.gpt-6.1-sol, reasoning high.Repository workflow publish-worker-image.yaml run 4063 passed on the worker host, publishing and selecting gitea.app.lan/autocoder/autocoder-engine@sha256:132e789bac1e222a4a99dfa41b4e7a3c707f370db476c5a284fd9a3f3803299f. The three ordinary subscription accounts still return catalogs without GPT 6.1 Sol. The existing fallback Pro account does return it, confirmed using its stored engine credentials in a restricted memory-backed diagnostic home; no credential import or account setting change was needed. The legacy pay_as_you_go field designates this fallback subscription and must not be interpreted as evidence of paid API billing.
The fallback worker was blocked by namespace CPU quota: 10.95 of 11 cores used, with a new pod requesting a 2 CPU limit inherited by its instruction init container from LimitRange. The actual task container only needs its configured 500m limit. Required scoped repair PR 6794 gives instruction initialization requests25m/64Mi and limits100m/128Mi. Its generation label participates in the existing worker mismatch logic, draining busy workers and replacing idle ones safely. Worker execution resources and namespace quota were not increased.
Reviewed repair head d24685cab8fec92d4be7d98f90bdf178d55e1771; branch checks run 4065 passed. Full candidate checks/test delivery run 4066 passed. Dev integration 7804310a166d661078ac019e04b0de6a1504dea1; automatic main integration 2aa497395148d013435dbba9576178b64367cd7b (PR 6795); production promotion run 4068 passed. Final production candidate 7804310a166d661078ac019e04b0de6a1504dea1-4066-1, backend gitea.app.lan/autocoder/autocoder-engine@sha256:e640930f26707ee0c5a2cf9612b61ef84ce7946970ce7505eb1b4ecb2f115314, frontend gitea.app.lan/autocoder/autocoder-engine-frontend@sha256:02378b1e6f309a0d91aca7fc8743446c32527b0f480384283421c2b7dd4b5392. The unchanged immutable configuration contract reused its verified preparation attestation. The worker binary image remains the already validated model-policy image; this repair changes engine-side pod rendering and reconciliation.
drg-k8s-worker-01.auth.home.lan; every instruction initializer has a 100m CPU limit and completed.autocoder-agent-operator-chat-4545036932c655024ca784b2618ba84d reports Codex CLI 0.159.0 and a fresh account-bound catalog including gpt-6.1-sol.Retained previous production candidate 6fbc3938c39bf1e89f1f897863a056af9229085b-4031-1 and its immutable configuration; rollback uses the documented candidate promotion/recovery route, restoring its preparation attestation. Worker workflow serializes with existing Lease autocoder-database-migration and restores the previous worker image on failed rollout. This change did not delete old maps or candidate evidence.
The operator subsequently excluded max entirely for Codex workloads. Reviewed PR 6798, head 65a88c8e0480a35f6272ea579b97ebcd183dc4ae, integrated into dev at 09af0b99da76ae2870ed243611b0f6fa6f15f711. Shared later candidates include that change. Full candidate run 4106 passed in 492 seconds at source c9b808013cac1fbfe97a9a13fa9c2f6a9d4ff93d; automatic main integration 09daea7cd2278e5a95b5da8f7ec156f4d75e2a6b and promotion run 4110 passed. Live backend is ready at that candidate source.
Common effort handling applies to enqueue, queued compare-and-swap migration, critical reconciliation, local invocation, worker exec, goal start/resume and chat. Missing Codex effort becomes high; max/ultra becomes xhigh. Dedicated provider/categorization interfaces and running/historical records are retained. Candidate acceptance logs explicitly passed default/high/max normalization and goal start/resume model tests. Live unstarted queue has two GPT 6.1 Sol/high tasks and no queued max/ultra. Model variables remain GPT 6.1 Sol/high; automation false, implementation workers false, operator chat workers true.
Worker publication run 4111 was dispatched from exact main revision above against the exact ready production source. Worker image and final worker behavioral acceptance are still pending; the previous 132e789... image remains selected at this checkpoint. Do not treat the backend deployment as completed worker delivery. Publication retains source/tree checks, existing migration Lease serialization and restoration of the previous image on rollout failure.
Pipeline monitor watch 14 retained successful exact pipeline evidence but still failed continuation with an active writer conflict. Further investigation verified that the original thread is owned by a standalone interactive CLI and not the monitor's managed daemon. Correct owner detection is under a separate bounded host repair; no writer lock was removed and no client was stopped. The worker publication is supervised in the owning administrator session while that transport limitation is investigated.
Publication run 4111 passed in 323 seconds, selecting immutable worker image gitea.app.lan/autocoder/autocoder-engine@sha256:053112704014e9d1615bbee847c71feab7630de1a17cf5f88865120b4f841085. All five enabled eligible operator chat workers are running and ready on drg-k8s-worker-01.auth.home.lan using that exact image. The fallback worker reports Codex CLI 0.159.0. Production backend remains ready at c9b808013cac1fbfe97a9a13fa9c2f6a9d4ff93d. Live model environment and both personal Codex profiles use GPT 6.1 Sol/high; two queued tasks remain GPT 6.1 Sol/high with no queued max/ultra and no running tasks. Background automation and implementation-worker stops remain false, operator chat workers true. The completed full candidate acceptance includes default high, max-to-xhigh normalization, provider-contract preservation and goal start/resume model binding; all builds and publication ran through worker CI. No paid inference or quota/reserve override was used for acceptance.
The requested model/effort correction is delivered. A separate host continuation transport repair remains open: the original standalone interactive CLI owns its kernel writer lock. A single retained terminal event was accepted through the supported CLI queue interface; autonomous consumption after the owning turn ends is not yet verified. Queued acceptance is not completed continuation. Do not replay that event or treat the transport branch as complete.
The blocked host handoff is repaired and verified through the installed scheduled monitor, not only a manual trial. Watch17 observed real publication4111 at main 09daea7cd2278e5a95b5da8f7ec156f4d75e2a6b; it queued one terminal event and automatically resumed the original CLI in turn 01a0eef9-4575-73a2-91b6-f67ceda3d80a. Exact client/text history evidence shows one consumed message, an empty thread queue and registry delivered with no error. Reviewed source, deployment and acceptance evidence.
Current acceptance recheck: both personal profiles and all Codex runtime role defaults remain GPT 6.1 Sol/high; all five published workers are ready. Later ready backend revision 18e0be36d26a91f96878c63745088fbad458af03 contains the reviewed model change. Background, requirements and implementation automation stops remain effective; dedicated-provider contracts are preserved. Model delivery and continuation repair are complete.