Account models and Antigravity quota fallback
To create a profile first, press a in Accounts. Choose its agent with Up/Down and Enter, then enter a display name. Web and desktop use the same two-step form with a selector. The profile ID is generated automatically and stays the same during a save retry. Saving selects that new profile and opens its sign-in flow. Choose the server with Up/Down and press Enter; web and desktop offer a server selector. Claude shows an official browser link, while other agents use their existing official CLI login.
Once the host confirms sign-in, the controller requests the available usage for that account and host immediately. The existing periodic refresh remains a fallback. Quota collection uses the existing read-only collector and cannot invalidate a successful login if telemetry is unavailable.
Select any account in Accounts and press Enter (or m), or click its row in the browser/desktop interface. This opens its model settings; Antigravity also offers the ordered quota fallback chain. Enter on a child session still attaches that session. The detailed/compact toggle (v) controls account details. Choose a host, refresh its installed-agent catalog and select the account's default model. Names come from that installed CLI. The catalog records its source workspace and CLI version; a pinned model is checked again in the target container after CLI preparation, since containers on one host can have different versions.
Accounts opens in compact view. The terminal keeps a vertical list; browser and native WebView clients keep account cards. Token meters use a green-to-red gradient; time meters use a separate blue gradient. Fill represents consumed tokens or elapsed time, while the accompanying text reports the remainder. Five-hour and weekly windows stay separate, and unavailable reset times do not produce an invented time meter.
Agent default (current CLI) is stored as null: devctrl omits the override and lets that version's native configuration select its default. It never guesses a fixed model. Explicit session/project settings take precedence. Saving an account preference preserves working processes. Missing/stale discovery stays explicit; clearing a preference does not require a catalog.
Claude uses its native SDK initialize response; Codex uses app-server model/list, without a user prompt, thread or turn. These are installed client catalogs, not proof of a paid inference request's entitlement. Native access errors remain authoritative. Antigravity uses models plus zero-turn /model and /usage JSON (read-only slash commands require agy 1.1.11+). Other agent packages need a verified discovery adapter; unsupported discovery never becomes an invented list.
Ordered Antigravity fallback
Enable quota fallback and order the models to try. The configured default is first; later entries are reserves in the same account/conversation. Prefer reserves in independent native quota groups: two Claude models can share every limit. The chain has no duplicates and moves forward after quota exhaustion; a reset does not send a session back to an earlier model.
The execution host checks the policy every 30 seconds, independently of the console and expensive container inventory scans. It refreshes missing/stale model and quota evidence through native read-only commands, with a per-account backoff. Every reported window in a candidate's shared native group must have known positive quota. A reported zero excludes all models sharing that group. A fresh current root Stop that carries the reporter’s verified native refusal marker can also trigger selection while cached usage remains positive. The strict marker currently recognizes the known RESOURCE_EXHAUSTED (code 429): Individual quota reached. refusal and its reported overage/reset suffix, as observed in the official CLI tracker. The complete native upgrade-and-reset refusal observed by the operator is also recognized without an RPC prefix; a bare prefixless quota phrase is insufficient. Generic quota mentions, tool errors, quoted messages, plain 429/rate-limit errors and old boolean-only records cannot override positive usage. The Stop must match the owned native process/conversation/generation, failed idle state and exact model/lifecycle timestamps; a 90-second freshness bound rejects old episodes. Every sibling in that failed native quota group is skipped. Missing or unknown candidate windows still block fallback. Small positive balances are not rounded to zero. Missing, stale, past-reset and unknown quota stay unknown.
Native refusal also records a separate quotaBlock on the account's affected model-group buckets, independently of whether automatic fallback is enabled. A backend can report a tiny positive remainder while refusing further requests: a reported balance such as 0.008% remains unchanged, and both terminal and browser show >99% used · <1% remaining · EXHAUSTED. The model picker and launch selection treat this group as exhausted, while independent groups in the same account remain usable. The block contains only the verified refusal source, model, occurrence time, first group reset and reported balance; it never fabricates zero or refreshes the quota sample's age. Matching bucket/group/reset polls retain the block while the balance is unchanged or decreases. A reset, increased official balance or block expiry clears it. A durable per-account/group receipt prevents replay of the same refusal after replenishment. Stale quota samples remain unavailable rather than proving fresh availability.
The owned reporter keeps up to 32 sanitized root refusals in devctl-agent-quota-refusals.json, so an ordinary subsequent Stop or model change cannot erase a refusal before the host reads it. The host publishes only fresh exact root/conversation/generation evidence; no raw error, prompt or transcript leaves the account store. For existing agy 1.4.0 conversations whose Stop was already overwritten, a narrow read-only backfill recognizes the observed root SQLite error layout, immediately preceding successful Argon backend model with the same native model enum, matching root provenance and no later successful retry. The controller additionally requires the current account catalog to be 1.4.0, a unique Argon group/window and its future official reset within two seconds of the error's native reset. Unknown schemas, other model families, child records, replaced processes and expired windows are skipped. Completed history scans run at most once per root every 15 minutes, retry snapshot failures after a minute, retain sanitized observations for delivery retries during the scan cooldown, read at most 64 error rows and 2 MiB of generation metadata, and have a 1.5 second bound. SQLite queries use a disposable private snapshot (main database capped at 128 MiB and WAL at 64 MiB), with owned read-only source descriptors and file-stability checks before and after the query; they create no native database/WAL/SHM files. Future refusals use the hook journal.
New sessions can use a fallback when their selected model's group is exhausted. Existing idle conversations also converge to an enabled account policy: an old model outside the chain selects the first available configured choice. This is a policy update, not a fabricated quota error or Stop event. A last invocation can have used a one-turn override; devctrl establishes a persistent selector once and retains its receipt across later invocation hooks.
The host verifies the same UID, process start, conversation and generation, unchanged native idle evidence, an empty composer and no attached terminal. Working, background-busy, compacting, question/permission, draft and attached sessions wait. A controller reservation and native input lock serialize input; account policy, Run leases and Overnight locks are rechecked before execution. Long-idle sessions can supply a new exact host observation without requiring a new model turn or Stop event.
Public read-only snapshot viewers can remain connected during native model selection: they cannot send terminal input and do not attach tmux. The broker keeps input exclusively reserved and still blocks every PTY/control client, pending attachment and competing operation. Ordinary background status overlays retain their stronger viewer exclusion so a viewer cannot be moved into a side conversation.
The guarded path uses the native persistent /model <slug> command, separate Enter and a new exact CLI confirmation. It preserves the process. Only an actual model change for a correlated native individual-quota failure can then submit Continue the current task. with a separate Enter. A normal idle policy/default update and same-model selector convergence never revive a completed task. This continuation requires the controller’s explicit quota-v1 capability and eligible reservation; old controllers retain selector-only behavior. A failed task may have been waiting longer than 90 seconds: an older exact Stop can authorize continuation only when a fresh periodic host read still proves that same failed idle episode and the controller selects using current official exhausted quota. Older failure history never overrides positive or unknown usage and never authorizes ordinary policy convergence. The 90-second refusal bound remains unchanged for overriding a positive cached quota sample. Native selection provenance remains separate from native invocation hooks. A later model observation is never overwritten by a delayed selection callback. A started or uncertain operation is durable and never automatically replayed, including after a policy edit or management restart; an exact native selection receipt can resolve a delayed callback. Safe skipped attempts have a cooldown. The guarded implementation requires agy 1.2.6+ and the host capability native-model-policy-v1.
Native receipt phases distinguish prepared, pasted, Enter intent, submitted and confirmed. The helper waits for two matching composer observations within a bounded window instead of assuming a paste has rendered after 100 ms. Enter intent is persisted before sending Enter; sanitized phase/failure codes explain uncertain outcomes without exposing terminal text. Pasted or submitted uncertainty is not automatically replayed. After native selection, continuation uses a new-model fence and separate durable prepared, pasted, submit-intent and submitted phases in the same operation receipt. UID, process start, pane, conversation/generation, freshly read unchanged failed-idle evidence, empty composer, drafts, questions and attachments are checked again; the exact pasted prompt must be stable before its sole Enter. A skipped or uncertain continuation leaves the model selection confirmed. Continuation submitted means the prompt was sent, not that native task execution or completion was confirmed. Neither a paste nor an uncertain Enter is retried, including after restart or repeated callbacks.
A bounded per-account recovery history records policy convergence, quota fallback and confirmed/skipped/uncertain outcomes, without transcripts or credentials. An optional nested continuation receipt reports skipped, submitted or uncertain, its input phase and a controlled failure code; model selection and task submission remain separate outcomes. The normal session recovery status explains waiting/blocked states. Enabling fallback refreshes only owned managed integration assets and preserves unrelated configuration.
Catalog/usage refresh and fallback run on the execution host with no connected console. Failed probes retain prior choices with an unavailable/stale label; pending discovery expires after two minutes. APIs enforce account ownership and host/workspace scope. Catalog metadata contains no credentials, raw errors or transcripts.
APIs and MCP
GET /v1/accounts/:accountId/models?hostId=...POST /v1/accounts/:accountId/models/refreshwith{ "hostId": "..." }GET /v1/accounts/:accountId/model-recovery-history?limit=25&after=...&sessionId=...returns bounded transition metadata andnextCursor; at most 50 rows per page, account-owner scoped.GET /v1/accounts/:accountId/model-preferencesPUT /v1/accounts/:accountId/model-preferenceswithhostId, nullabledefaultModel, orderedfallbackModelsandquotaFallbackEnabled.PUT /v1/agents/:id/sessions/:sessionId/model-quota-refusalaccepts owned native root Stop or verified 1.4.0 root-history observations independently of fallback policy and without taking session input.
Operator MCP discovers these operations with controller_operations, then calls their exact routes with controller_request. Session MCP get_session_settings reports its own defaults, applied launch choice and recovery state. Project sessions cannot edit account-wide settings or another user's account.
Official references: model discovery, quota groups, Stop hooks, persistent and one-turn model command semantics, Codex app-server.