Blocking — preserve the released provider identity. This removes the migration-aware persisted-provider schema and reinterprets the existing unversioned gemini value as the new native Gemini provider. In released Scient data, that same byte value meant Antigravity. I reproduced this with the same saved settings/model/handoff bytes: the base decodes them as Antigravity, while this head decodes them as native Gemini. That can resume a draft or handoff against the wrong provider/account. Please introduce a distinct persisted discriminator or a versioned migration and apply it consistently to app settings, composer drafts, server model selection, orchestration, and thread handoffs. Add A/B fixtures using actual released payloads. Do not change the historical meaning of gemini in place, and do not update tests to bless the reinterpretation. This migration should be its own PR.

Blocking — Plan mode must never auto-approve mutations. This branch checks only runtimeMode === "full-access"; it does not check the active interaction mode recorded on the adapter context. A Gemini turn in Plan mode can therefore auto-approve filesystem or command permissions when full-access runtime is selected. Scient's existing invariant is that Plan rejects mutation permissions even under full access. Fail closed on Plan before this auto-approval path and add a real adapter permission test equivalent to the existing Droid negative control.

Blocking — Gemini effort selections do not survive the wire contract. Gemini is routed through the generic effort-descriptor path, which produces reasoningEffort, but the Gemini contract accepts only thinkingLevel (Gemini 3) or thinkingBudget (Gemini 2.5). In exact probes, a selected Gemini 3 LOW was displayed as HIGH after round-trip and a Gemini 2.5 budget of 512 was displayed as -1; the codec stripped both selections to empty options. Add Gemini-specific descriptors keyed by model family and end-to-end tests covering picker selection → encoded contract → decoded selection → adapter dispatch for both Gemini 3 and 2.5. Keep this in the adapter/model-options PR.

Blocking — remove this OAuth-token usage fetcher. It reads the Gemini CLI OAuth credential and calls private cloudcode-pa.googleapis.com/v1internal endpoints. Google's official Gemini CLI FAQ says third-party software must not harvest/piggyback Gemini CLI OAuth credentials to access backend services and warns this can result in account suspension or termination: https://github.com/google-gemini/gemini-cli/blob/main/docs/resources/faq.md#why-cant-i-use-third-party-software-like-claude-code-openclaw-or-opencode-with-gemini-cli. Launching the official CLI/ACP is a separate supported boundary; reusing its bearer token from Scient is not. Remove/disable this provider-usage implementation. If usage is reintroduced, do it in a separate PR through a documented Google-supported API and credential flow, with explicit product/security review and failure/backoff tests.

Blocking — delegate Codex authentication to the official Codex flow. Scient already invokes codex login; this new 382-line OAuth implementation bypasses Codex's credential-store selection (file/keyring/auto/ephemeral), auth restrictions, configured route/proxy behavior, evolving schema, and concurrency handling. In a disposable full-flow probe, writing into a seeded auth.json removed the existing API key, auth_mode, and unrelated marker and replaced them with raw OAuth tokens. Please remove the native implementation and have one-click connect install/resolve Codex and then invoke the official login/app-server flow. Official implementation reference: https://github.com/openai/codex/blob/main/codex-rs/login/src/server.rs. Keep connection orchestration separate from credential ownership.

Blocking product disclosure — this statement is false for this head. The new OpenAI path itself receives and writes raw access, refresh, and ID tokens into CODEX_HOME/auth.json, so credentials do not simply 'stay with the provider.' Gemini usage code also reads the CLI OAuth token. Please remove the native credential handling and Gemini OAuth piggyback as requested above, then use truthful, provider-specific copy describing who stores credentials and where. The UI must not promise a security boundary the implementation does not maintain.

Required — bind the connection chain to the exact install operation. This is one global {provider, token} slot and is not associated with the managed install's operationId. Starting provider B can overwrite provider A; a stale completion can trigger or clear the wrong chain; and closing/reopening the dialog deliberately leaves the chain alive without an explicit cancellation owner. Represent the chain with provider + exact operationId + lifecycle state, define replacement/cancellation semantics, and test concurrent providers, stale completion, failure/retry, and dialog close/reopen. This belongs in the small connection-orchestration PR.

Required — the custom Gemini settings path is only partially wired. This makes Gemini appear in the generic custom-model editor, but the settings route's onValueChange provider allowlist omits gemini, so selecting the visible Gemini entry is rejected. The reset/dirty-state paths also omit Gemini fields (including customGeminiModels and geminiBinaryPath in relevant aggregates). Please use one exhaustive provider-keyed source rather than duplicated allowlists, cover save/reset/dirty behavior, and add a browser test that actually edits and persists a Gemini model. This can be a separate settings-integration PR.

Required — this query cannot make Gemini selectable in the library. Gemini is added to the capability map but omitted from PROVIDER_DISCOVERY_ORDER, which drives provider fallback/selection. The integration is therefore internally inconsistent and future Gemini discovery support remains invisible. Either omit this inactive integration from the adapter PR or add Gemini through an exhaustive provider list and test provider selection/fallback. Please keep plugin-library work separate from core ACP transport.

Required — make SVG definitions instance-safe and document provenance. These fixed mask/filter IDs are duplicated whenever more than one Gemini icon renders, so references can collide across instances. Use useId (as other reusable icons do) or remove the filter-ID dependency. This artwork also matches the donor implementation; before landing, record its source/lineage and verify that its license/brand-use terms are compatible with this repository, including the two copied public SVGs. Icon/provenance cleanup does not need to ride with provider behavior.

Required — do not silently swallow browser-launch failure. If openExternal rejects, the operation is marked as already auto-opened and the error disappears. Record an observable failure/warning on the connection operation and keep a clear manual Open browser again action. Ensure a failed automatic launch can be retried and add a browser test for openExternal rejection. This matters even after switching Codex back to official codex login, because every browser-auth provider needs a recoverable launch failure.

Required — do not clear the chain before the login handoff succeeds. The chain is deleted before startProviderConnection is accepted. If that call fails, retry ownership is lost; when the dialog is closed, setActionError is also not visible to the user. Keep the exact operation active through the handoff, transition it to a durable terminal failure/retry state on rejection, and clear it only after the server acknowledges the new login operation or the user explicitly cancels it. Cover closed-dialog failure, retry, stale completion, and concurrent providers.

Required scope/compatibility fix — split this from the connection PR. This introduces a raw call to the explicitly unstable session/set_model method when no config ID exists. It bypasses the typed client surface, is not capability-gated, and has no direct test for unsupported agents, protocol failure, or successful model application. Move it to a focused ACP compatibility PR; use the typed setSessionModel surface where available, gate it on advertised support, define the unsupported fallback, and add direct protocol tests before shipping it.

Rebase blocker — preserve current main’s AppSnap removal and arbitrate startup overlays. This branch now conflicts with main here. Main intentionally removed AppSnapWelcomeDialog; resolving by taking the branch block would restore the startup snapshot notification. Keep that dialog removed. Also avoid simply stacking onboarding beside GlobalWhatsNewSurface and other root dialogs: route startup surfaces through explicit modal arbitration so only one can open at a time, with a test for competing eligible dialogs.