Main session model selection

Main-session model policy has its own controls in the Decisions model view. It is independent of automatic subagent routing. The Decisions Model selection usage switch, automatic Decisions setting, advisor configuration, monthly budgets, current owner grants and account locks still apply. The main assessment switch controls model decisions; the bounded retry switch controls same-model recovery independently, including when the advisor is not configured.

At the start of an accepted session launch, an existing initial task can be assessed before the native process starts. An empty launch does not request a TaskGoal, invent a task or send a provider prompt. It retains the selected model. After a resume, the host verifies the same account, native conversation and process generation and reads bounded native user-task/compaction context. It does not export tool output or a full transcript. Common credentials are redacted before hashing, history recording and advisor evaluation; task text is treated as untrusted data rather than policy instructions.

Resume assessment has one accepted launch window, lasting at most 45 seconds from the verified process start. It binds the existing task revision and closes after its decision/application. A fresh real human turn closes that window; completion of an ordinary turn cannot authorize lowering for an unknown next task. An empty start retains its baseline without a later idle assessment.

Owned native-submit hashes identify automatic permission/file/resource notices, status probes and recovery continuations. These messages do not replace the real human task. The current matching native goal and accepted ordinary workqueue goal can supplement that task. Missing provenance for a versioned controller packet or clipped task evidence retains the baseline; ordinary human questions about permissions remain valid task context. A later real human task supersedes an earlier goal's pause/completion state.

A native temporary-capacity error first asks whether a lower eligible model can complete the current task to the baseline's quality. The baseline remains an explicit option. An unsuitable, uncertain or unavailable decision retains it and uses the existing bounded retry schedule: 30, 60, 120 and 300 seconds, at most four attempts within 20 minutes. Quota, authentication, egress, billing and context length failures are not temporary-capacity evidence.

Decisions can only lower a known capability tier and both known API token tariffs, on the same provider and account. Actual installed-account model IDs, compatible tools and reasoning effort are required. The context gate uses native current context used, with exact identity and generation. Cumulative lifetime tokens are never used as the current context size. Unknown models, default selections without actual model evidence, unavailable native context, stale catalogs and unsupported native interfaces retain the current model. Diagnostic evidence-only, policy-review workers and pinned review sessions retain their restricted model profile.

The host waits for native idle state, no active subagents, no approval/question and an empty composer. Codex 0.164 uses its verified session-only inline model command and genuine native acknowledgement. Claude Code 2.1.292 uses the documented session-only model-picker action and requires refreshed actual model/effort statusline evidence. It does not save a new account-wide default. Other native interfaces remain conservative until qualified.

Model application and continuation are separate operations. Only after a native model confirmation for a correlated capacity-interrupted task may the host send one short continuation and one Enter. Start/resume assessment alone does not restart idle work. A durable receipt is written before each ambiguous input effect. Unknown delivery, a lost completion response or a restart cannot replay the model command or fall through to another capacity retry.

Decisions history distinguishes retained, proposed, reserved, confirmed, skipped and uncertain application. It records the trigger, from/to models, confidence, reason, bounded retry time and continuation state. A command appearing in the composer is not confirmation, and a submitted continuation is not evidence that the task completed. Future task token consumption is unknown, so automatic decisions do not claim measured savings. Subscription cash savings remain unknown; any priced comparison is an API-equivalent forecast.

The same owner-scoped policy is discoverable through control_operations and control_request in Control Room, and through the CLI:

devctl api operations main-sessions
devctl api call < main-session-model-settings.json

For a read, the JSON input is:

{"operation":"GET /v1/decisions/main-sessions/settings","pathParameters":{},"query":{}}

For a change, read the current policy and use the discovered PUT operation with its published body schema. Saving the policy does not submit a task or alter a running session immediately. The internal host observation, reservation and completion endpoints are not operator application receipts.

Offline fixtures exercise native input phases and controller boundaries without provider traffic. Default verification makes no live provider calls. Live provider validation has not been run; it requires a separately authorized disposable fixture. Offline fixtures do not establish live account availability or paid task quality.