ios: preserve chat scroll position on resume

adds gemini support
ci: restore Ruby PATH export for fastlane step
2026-07-12 13:57:45 -07:00 · 2026-07-11 14:16:21 -07:00 · 2026-07-11 13:04:47 -07:00 · 2026-07-11 13:01:42 -07:00 · 2026-07-11 12:01:27 -07:00 · 2026-07-11 11:55:04 -07:00
33 changed files with 897 additions and 179 deletions
@@ -4,20 +4,15 @@ on:
  workflow_dispatch:
  push:
    tags:
-      - "v*"
+      - "release/ios/v*"

 jobs:
  testflight:
-    runs-on: xcode
-    defaults:
-      run:
-        shell: bash
+    runs-on: macos-arm64

    steps:
      - name: Checkout
        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0

      - name: Setup Ruby
        uses: ruby/setup-ruby@v1
@@ -27,35 +22,11 @@ jobs:
          working-directory: ios

      - name: Install XcodeGen
-        run: |
-          set -euo pipefail
-          if ! command -v xcodegen >/dev/null 2>&1; then
-            brew install xcodegen
-          fi
-
-      - name: Prepare Runner Keychain
-        env:
-          HOME: /var/lib/act_runner
-        run: |
-          set -euo pipefail
-          mkdir -p "${HOME}/Library/Keychains"
-
-          login_keychain="${HOME}/Library/Keychains/login.keychain"
-          if [ ! -f "${login_keychain}-db" ]; then
-            security create-keychain -p "" "${login_keychain}"
-          fi
-
-          security unlock-keychain -p "" "${login_keychain}" 2>/dev/null || \
-            security unlock-keychain -p "sybil-ci-keychain-password" "${login_keychain}" 2>/dev/null || true
-          security default-keychain -d user -s "${login_keychain}"
-          security list-keychains -d user -s "${login_keychain}-db"
-          security delete-keychain "${HOME}/Library/Keychains/sybil_ci_keychain" >/dev/null 2>&1 || true
-          rm -f "${HOME}/Library/Keychains/sybil_ci_keychain" "${HOME}/Library/Keychains/sybil_ci_keychain-db"
+        run: command -v xcodegen >/dev/null 2>&1 || brew install xcodegen

      - name: Upload to TestFlight
        working-directory: ios
        env:
-          HOME: /var/lib/act_runner
          APP_STORE_CONNECT_KEY_ID: ${{ secrets.APP_STORE_CONNECT_KEY_ID }}
          APP_STORE_CONNECT_ISSUER_ID: ${{ secrets.APP_STORE_CONNECT_ISSUER_ID }}
          APP_STORE_CONNECT_KEY_CONTENT: ${{ secrets.APP_STORE_CONNECT_KEY_CONTENT }}
@@ -65,7 +36,9 @@ jobs:
          SYBIL_BUILD_NUMBER: ${{ github.run_number }}
          FASTLANE_SKIP_UPDATE_CHECK: "1"
          FASTLANE_XCODEBUILD_SETTINGS_TIMEOUT: "120"
+        # act_runner does not propagate setup-ruby's PATH changes into later
+        # steps, so put the selected Ruby back on PATH or `bundle` resolves to
+        # the toolcache default and misses the gems installed above.
        run: |
          export PATH="/Users/runner/hostedtoolcache/Ruby/3.1.7/arm64/bin:${PATH}"
-          ruby --version
          bundle exec fastlane ios beta
@@ -12,6 +12,7 @@ services:
      OPENAI_API_KEY: ${OPENAI_API_KEY:-}
      ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-}
      XAI_API_KEY: ${XAI_API_KEY:-}
+      GEMINI_API_KEY: ${GEMINI_API_KEY:-}
      HERMES_AGENT_API_BASE_URL: ${HERMES_AGENT_API_BASE_URL:-http://127.0.0.1:8642/v1}
      HERMES_AGENT_API_KEY: ${HERMES_AGENT_API_KEY:-}
      HERMES_AGENT_MODEL: ${HERMES_AGENT_MODEL:-}
@@ -34,11 +34,13 @@ Chat upload limits:
    "openai": { "models": ["gpt-4.1-mini"], "loadedAt": "2026-02-14T00:00:00.000Z", "error": null },
    "anthropic": { "models": ["claude-3-5-sonnet-latest"], "loadedAt": null, "error": null },
    "xai": { "models": ["grok-3-mini"], "loadedAt": null, "error": null },
+    "gemini": { "models": ["gemini-3.5-flash"], "loadedAt": null, "error": null },
    "hermes-agent": { "models": ["hermes-agent"], "loadedAt": null, "error": null }
  }
 }
 ```
 - OpenAI model lists are filtered to models that are expected to work with the backend's Responses API implementation.
+- Gemini model lists are loaded from Google's native Models API and filtered to Gemini `generateContent` model ids.
 - `hermes-agent` is included only when `HERMES_AGENT_API_KEY` is configured. Set it to Hermes `API_SERVER_KEY`, or any non-empty value if that local server does not require auth. `HERMES_AGENT_API_BASE_URL` defaults to `http://127.0.0.1:8642/v1`; set `HERMES_AGENT_MODEL` only when you need an additional fallback/override model id.
 - The backend loads provider model lists at startup and refreshes them about once every 24 hours. If a later provider refresh fails, the response keeps the last loaded model list for that provider and sets `error` to the latest failure message.

@@ -56,7 +58,7 @@ Chat upload limits:
 ```

 Behavior notes:
- Lists Sybil-managed chat tools that can be enabled for `openai`, `anthropic`, and `xai` chat completions.
+- Lists Sybil-managed chat tools that can be enabled for `openai`, `anthropic`, `xai`, and `gemini` chat completions.
 - Optional tools such as `codex_exec` and `shell_exec` appear only when enabled by server environment configuration.

 ## Active Runs
@@ -128,7 +130,7 @@ Behavior notes:
 ```json
 {
  "title": "optional title",
-  "provider": "optional openai|anthropic|xai|hermes-agent",
+  "provider": "optional openai|anthropic|xai|gemini|hermes-agent",
  "model": "optional model id",
  "additionalSystemPrompt": "optional stored system prompt",
  "enabledTools": ["web_search", "fetch_url"],
@@ -234,7 +236,7 @@ Notes:
 ```json
 {
  "chatId": "optional-chat-id",
-  "provider": "openai|anthropic|xai|hermes-agent",
+  "provider": "openai|anthropic|xai|gemini|hermes-agent",
  "model": "string",
  "messages": [
    {
@@ -285,6 +287,7 @@ Behavior notes:
 - For `chatId` calls, server stores only *new* non-assistant messages from provided history to avoid duplicates.
 - `additionalSystemPrompt`, when present directly or loaded from stored chat settings, is prepended to the provider request as a `system` message and is not inserted into the persisted chat transcript by this endpoint.
 - `enabledTools` limits Sybil-managed tools for this request. When omitted for a saved chat, the stored chat setting is used; otherwise all available tools are enabled by default. An empty array disables Sybil-managed tools.
+- `maxTokens` is optional. For `anthropic`, when omitted the backend requests the selected model's maximum output token limit from Anthropic's Models API and uses that as `max_tokens`; if the model limit cannot be loaded, the fallback is 128000. For other providers, omitted `maxTokens` is not sent as an explicit cap.
 - Server persists final assistant output and call metadata (`LlmCall`) in DB.
 - Server updates chat-level model metadata on each call: `lastUsedProvider`/`lastUsedModel`; first successful/failed call also initializes `initiatedProvider`/`initiatedModel` if unset.
 - Attachments are optional and currently apply to `user` messages. Persisted chat history stores them under `message.metadata.attachments`.
@@ -293,12 +296,14 @@ Behavior notes:
 - For `openai`, backend calls OpenAI's Responses API and enables internal tool use with an internal system instruction.
 - For `anthropic`, backend calls Anthropic's Messages API and enables internal tool use with Anthropic `tool_use`/`tool_result` content blocks.
 - For `xai`, backend calls xAI's OpenAI-compatible Chat Completions API and enables internal tool use with the same internal system instruction.
+- For `gemini`, backend calls Google's native Gemini `generateContent` API and enables internal tool use with Gemini function calling.
 - For `hermes-agent`, backend calls the configured Hermes Agent OpenAI-compatible Chat Completions API without adding Sybil-managed tool definitions; Hermes Agent handles its own tools server-side.
 - For `openai`, image attachments are sent as Responses `input_image` items and text attachments are sent as `input_text` items.
+- For `gemini`, image attachments are sent as native Gemini `inlineData` parts and text attachments are sent as text parts.
 - For `xai` and `hermes-agent`, image attachments are sent as Chat Completions content parts alongside text.
 - For `openai`, Responses calls that can enter the server-managed tool loop use `store: true` so reasoning and function-call items can be passed between tool rounds.
 - For `anthropic`, image attachments are sent as Messages API `image` blocks using base64 source data; text attachments are added as `text` blocks.
- Available Sybil-managed tool calls for `openai`, `anthropic`, and `xai`: `web_search` and `fetch_url`. When `CHAT_CODEX_TOOL_ENABLED=true`, `codex_exec` is also available. When `CHAT_SHELL_TOOL_ENABLED=true`, `shell_exec` is also available.
+- Available Sybil-managed tool calls for `openai`, `anthropic`, `xai`, and `gemini`: `web_search` and `fetch_url`. When `CHAT_CODEX_TOOL_ENABLED=true`, `codex_exec` is also available. When `CHAT_SHELL_TOOL_ENABLED=true`, `shell_exec` is also available.
 - `web_search` returns ranked results with per-result summaries/snippets. Its backend engine is selected by `CHAT_WEB_SEARCH_ENGINE` (`exa` default, or `searxng` with `SEARXNG_BASE_URL` set). SearXNG mode requires the instance to allow `format=json`.
 - `fetch_url` fetches a URL with browser-like navigation headers and returns plaintext page content (HTML converted to text server-side).
 - `codex_exec` delegates coding, shell, repository inspection, and other complex software tasks to a persistent remote Codex CLI workspace over SSH. The server runs `codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check <non-interactive wrapped prompt>` on the configured devbox inside `CHAT_CODEX_REMOTE_WORKDIR`, with SSH stdin closed.
@@ -416,9 +421,9 @@ Behavior notes:
  "updatedAt": "...",
  "starred": false,
  "starredAt": null,
-  "initiatedProvider": "openai|anthropic|xai|hermes-agent|null",
+  "initiatedProvider": "openai|anthropic|xai|gemini|hermes-agent|null",
  "initiatedModel": "string|null",
-  "lastUsedProvider": "openai|anthropic|xai|hermes-agent|null",
+  "lastUsedProvider": "openai|anthropic|xai|gemini|hermes-agent|null",
  "lastUsedModel": "string|null",
  "additionalSystemPrompt": null,
  "enabledTools": ["web_search", "fetch_url"]
@@ -468,9 +473,9 @@ Behavior notes:
  "updatedAt": "...",
  "starred": false,
  "starredAt": null,
-  "initiatedProvider": "openai|anthropic|xai|hermes-agent|null",
+  "initiatedProvider": "openai|anthropic|xai|gemini|hermes-agent|null",
  "initiatedModel": "string|null",
-  "lastUsedProvider": "openai|anthropic|xai|hermes-agent|null",
+  "lastUsedProvider": "openai|anthropic|xai|gemini|hermes-agent|null",
  "lastUsedModel": "string|null",
  "additionalSystemPrompt": null,
  "enabledTools": ["web_search", "fetch_url"],
@@ -21,7 +21,7 @@ Authentication:
 {
  "chatId": "optional-chat-id",
  "persist": true,
-  "provider": "openai|anthropic|xai|hermes-agent",
+  "provider": "openai|anthropic|xai|gemini|hermes-agent",
  "model": "string",
  "messages": [
    {
@@ -64,6 +64,7 @@ Notes:
 - For persisted streams, backend stores only new non-assistant input history rows to avoid duplicates.
 - `additionalSystemPrompt`, when present directly or loaded from stored chat settings, is prepended to the provider request as a `system` message and is not inserted into the persisted chat transcript by this endpoint.
 - `enabledTools` limits Sybil-managed tools for this request. When omitted for a saved chat, the stored chat setting is used; otherwise all available tools are enabled by default. An empty array disables Sybil-managed tools.
+- `maxTokens` is optional. For `anthropic`, when omitted the backend requests the selected model's maximum output token limit from Anthropic's Models API and uses that as `max_tokens`; if the model limit cannot be loaded, the fallback is 128000. For other providers, omitted `maxTokens` is not sent as an explicit cap.
 - Attachments are optional and are persisted under `message.metadata.attachments` on stored user messages when `persist` is `true`.

 Persisted chat streams with a `chatId` are backend-owned active runs:
@@ -173,9 +174,11 @@ Terminal tool-call event:
 - `openai`: backend uses OpenAI's Responses API and may execute internal function tool calls (`web_search`, `fetch_url`, optional `codex_exec`, and optional `shell_exec`) before producing final text.
 - `anthropic`: backend uses Anthropic's Messages API and may execute the same internal tools with `tool_use`/`tool_result` content blocks before producing final text.
 - `xai`: backend uses xAI's OpenAI-compatible Chat Completions API and may execute the same internal tool calls before producing final text.
+- `gemini`: backend uses Google's native Gemini `streamGenerateContent` API and may execute the same internal tool calls before producing final text.
 - `fetch_url` sends browser-like navigation headers for outbound URL requests to reduce false 403s from sites that reject generic server clients.
 - `hermes-agent`: backend uses the configured Hermes Agent OpenAI-compatible Chat Completions API. Sybil does not add its own tool definitions for this provider; Hermes Agent handles its own tools server-side. Custom Hermes stream events are normalized away unless they produce text deltas in this SSE contract.
 - `openai`: image attachments are sent as Responses `input_image` items; text attachments are sent as `input_text` items.
+- `gemini`: image attachments are sent as native Gemini `inlineData` parts; text attachments are inlined as text parts.
 - `xai` and `hermes-agent`: image attachments are sent as Chat Completions content parts; text attachments are inlined as text parts.
 - `openai`: Responses calls that can enter the server-managed tool loop use `store: true` so reasoning and function-call items can be passed between tool rounds.
 - `anthropic`: streamed via event stream; emits `delta` from `content_block_delta` with `text_delta`, and emits normalized `tool_call` SSE events when Anthropic `tool_use` blocks are executed. Image attachments are sent as base64 `image` blocks and text attachments are appended as `text` blocks.
@@ -184,7 +187,7 @@ Terminal tool-call event:
 - `shell_exec` is available only when `CHAT_SHELL_TOOL_ENABLED=true`. It uses the same devbox SSH configuration, starts in `CHAT_CODEX_REMOTE_WORKDIR`, and runs non-interactive shell commands there with SSH stdin closed, not inside the Sybil server container.
 - `CHAT_MAX_TOOL_ROUNDS` controls how many model/tool result cycles may occur before the backend returns a tool-call limit message; default is 100.

-Tool-enabled streaming notes (`openai`/`anthropic`/`xai`):
+Tool-enabled streaming notes (`openai`/`anthropic`/`xai`/`gemini`):
 - Stream still emits standard `meta`, `delta`, `done|error` events.
 - Stream may emit `tool_call` events while tool calls are executed.
 - `delta` events carry assistant text and are emitted incrementally for normal text rounds. The backend may buffer model-native text briefly while determining whether a provider round contains tool calls.
@@ -21,6 +21,12 @@ Instructions for work under `/Users/buzzert/src/sybil-2/ios`.
 - To choose a screenshot path, run `just screenshot path=build/name.png`.
 - The underlying screenshot command is `xcrun simctl io booted screenshot <path>` and requires a booted simulator.

+## Release Workflow
+- iOS release tags use the annotated tag namespace `release/ios/vX.Y.Z`; increment from the latest existing `release/ios/v*` tag.
+- Tag message convention is `ios: X.Y.Z`, for example `git tag -a release/ios/v1.13.5 -m "ios: 1.13.5"`.
+- Push the release commit and tag together with `git push origin <branch> release/ios/vX.Y.Z`.
+- Fastlane derives the marketing version from the release tag and stamps `ios/Apps/Sybil/project.yml` during CI, so do not manually bump `MARKETING_VERSION` for normal tagged releases unless explicitly requested.
+
 ## App Structure
 - App target entry: `/Users/buzzert/src/sybil-2/ios/Apps/Sybil/Sources/SybilApp.swift`
 - Shared iOS app code lives in Swift package:
@@ -51,4 +57,5 @@ Instructions for work under `/Users/buzzert/src/sybil-2/ios`.
  - OpenAI: `gpt-4.1-mini`
  - Anthropic: `claude-3-5-sonnet-latest`
  - xAI: `grok-3-mini`
+  - Gemini: `gemini-3.5-flash`
  - Hermes Agent: `hermes-agent`
@@ -24,7 +24,7 @@ targets:
        GENERATE_INFOPLIST_FILE: YES
        INFOPLIST_FILE: Apps/Sybil/Info.plist
        ASSETCATALOG_COMPILER_APPICON_NAME: AppIcon
-        MARKETING_VERSION: "1.10"
+        MARKETING_VERSION: "1.13.2"
        CURRENT_PROJECT_VERSION: 11
        INFOPLIST_KEY_CFBundleDisplayName: Sybil
        INFOPLIST_KEY_ITSAppUsesNonExemptEncryption: NO
@@ -58,7 +58,8 @@ struct SybilChatTranscriptView: View {
                        .frame(height: 18 + bottomContentInset)
                        .id(bottomAnchorID)
                }
-                .frame(maxWidth: .infinity, alignment: .leading)
+                .frame(maxWidth: SybilLayout.webContentMaxWidth, alignment: .leading)
+                .frame(maxWidth: .infinity, alignment: .center)
                .padding(.horizontal, 14)
                .padding(.top, 18 + topContentInset)
            }
@@ -66,7 +67,6 @@ struct SybilChatTranscriptView: View {
            .scrollDismissesKeyboard(.interactively)
            .onAppear {
                syncKnownToolCallMessageIDs()
-                scrollToBottom(with: proxy, animated: false)
            }
            .onChange(of: toolCallMessageIDSignature) { _, _ in
                syncKnownToolCallMessageIDs()
@@ -4,6 +4,7 @@ public enum Provider: String, Codable, CaseIterable, Hashable, Sendable {
    case openai
    case anthropic
    case xai
+    case gemini
    case hermesAgent = "hermes-agent"

    public var displayName: String {
@@ -11,6 +12,7 @@ public enum Provider: String, Codable, CaseIterable, Hashable, Sendable {
        case .openai: return "OpenAI"
        case .anthropic: return "Anthropic"
        case .xai: return "xAI"
+        case .gemini: return "Gemini"
        case .hermesAgent: return "Hermes Agent"
        }
    }
@@ -98,7 +98,8 @@ struct SybilSearchResultsView: View {
                        .foregroundStyle(SybilTheme.danger)
                }
            }
-            .frame(maxWidth: .infinity, alignment: .leading)
+            .frame(maxWidth: SybilLayout.webContentMaxWidth, alignment: .leading)
+            .frame(maxWidth: .infinity, alignment: .center)
            .padding(.horizontal, 14)
            .padding(.top, 20 + topContentInset)
            .padding(.bottom, 20 + bottomContentInset)
@@ -11,11 +11,13 @@ final class SybilSettingsStore {
        static let preferredOpenAIModel = "sybil.ios.preferredOpenAIModel"
        static let preferredAnthropicModel = "sybil.ios.preferredAnthropicModel"
        static let preferredXAIModel = "sybil.ios.preferredXAIModel"
+        static let preferredGeminiModel = "sybil.ios.preferredGeminiModel"
        static let preferredHermesAgentModel = "sybil.ios.preferredHermesAgentModel"
        static let quickQuestionPreferredProvider = "sybil.ios.quickQuestionPreferredProvider"
        static let quickQuestionPreferredOpenAIModel = "sybil.ios.quickQuestionPreferredOpenAIModel"
        static let quickQuestionPreferredAnthropicModel = "sybil.ios.quickQuestionPreferredAnthropicModel"
        static let quickQuestionPreferredXAIModel = "sybil.ios.quickQuestionPreferredXAIModel"
+        static let quickQuestionPreferredGeminiModel = "sybil.ios.quickQuestionPreferredGeminiModel"
        static let quickQuestionPreferredHermesAgentModel = "sybil.ios.quickQuestionPreferredHermesAgentModel"
    }

@@ -44,6 +46,7 @@ final class SybilSettingsStore {
            .openai: defaults.string(forKey: Keys.preferredOpenAIModel) ?? "gpt-4.1-mini",
            .anthropic: defaults.string(forKey: Keys.preferredAnthropicModel) ?? "claude-3-5-sonnet-latest",
            .xai: defaults.string(forKey: Keys.preferredXAIModel) ?? "grok-3-mini",
+            .gemini: defaults.string(forKey: Keys.preferredGeminiModel) ?? "gemini-3.5-flash",
            .hermesAgent: defaults.string(forKey: Keys.preferredHermesAgentModel) ?? "hermes-agent"
        ]
        self.preferredModelByProvider = preferredModels
@@ -54,6 +57,7 @@ final class SybilSettingsStore {
            .openai: defaults.string(forKey: Keys.quickQuestionPreferredOpenAIModel) ?? preferredModels[.openai] ?? "gpt-4.1-mini",
            .anthropic: defaults.string(forKey: Keys.quickQuestionPreferredAnthropicModel) ?? preferredModels[.anthropic] ?? "claude-3-5-sonnet-latest",
            .xai: defaults.string(forKey: Keys.quickQuestionPreferredXAIModel) ?? preferredModels[.xai] ?? "grok-3-mini",
+            .gemini: defaults.string(forKey: Keys.quickQuestionPreferredGeminiModel) ?? preferredModels[.gemini] ?? "gemini-3.5-flash",
            .hermesAgent: defaults.string(forKey: Keys.quickQuestionPreferredHermesAgentModel) ?? preferredModels[.hermesAgent] ?? "hermes-agent"
        ]
    }
@@ -72,12 +76,14 @@ final class SybilSettingsStore {
        defaults.set(preferredModelByProvider[.openai], forKey: Keys.preferredOpenAIModel)
        defaults.set(preferredModelByProvider[.anthropic], forKey: Keys.preferredAnthropicModel)
        defaults.set(preferredModelByProvider[.xai], forKey: Keys.preferredXAIModel)
+        defaults.set(preferredModelByProvider[.gemini], forKey: Keys.preferredGeminiModel)
        defaults.set(preferredModelByProvider[.hermesAgent], forKey: Keys.preferredHermesAgentModel)

        defaults.set(quickQuestionPreferredProvider.rawValue, forKey: Keys.quickQuestionPreferredProvider)
        defaults.set(quickQuestionPreferredModelByProvider[.openai], forKey: Keys.quickQuestionPreferredOpenAIModel)
        defaults.set(quickQuestionPreferredModelByProvider[.anthropic], forKey: Keys.quickQuestionPreferredAnthropicModel)
        defaults.set(quickQuestionPreferredModelByProvider[.xai], forKey: Keys.quickQuestionPreferredXAIModel)
+        defaults.set(quickQuestionPreferredModelByProvider[.gemini], forKey: Keys.quickQuestionPreferredGeminiModel)
        defaults.set(quickQuestionPreferredModelByProvider[.hermesAgent], forKey: Keys.quickQuestionPreferredHermesAgentModel)
    }

@@ -64,6 +64,10 @@ extension Font {
    }
 }

+enum SybilLayout {
+    static let webContentMaxWidth: CGFloat = 896
+}
+
 enum SybilTheme {
    static let background = Color(red: 0.02, green: 0.02, blue: 0.05)
    static let surface = Color(red: 0.05, green: 0.04, blue: 0.10)
@@ -160,6 +160,7 @@ final class SybilViewModel {
        .openai: ["gpt-4.1-mini"],
        .anthropic: ["claude-3-5-sonnet-latest"],
        .xai: ["grok-3-mini"],
+        .gemini: ["gemini-3.5-flash", "gemini-flash-latest"],
        .hermesAgent: ["hermes-agent"]
    ]

@@ -1751,13 +1752,16 @@ final class SybilViewModel {
                switch target {
                case let .chat(chatID):
                    SybilLog.debug(SybilLog.app, "Refreshing chat \(chatID)")
+                    let isSelectingDifferentChat = selectedChat?.id != chatID
                    let chat = try await client.getChat(chatID: chatID)
                    guard selectedItem == target, draftKind == nil else {
                        return
                    }
                    selectedChat = chat
                    selectedSearch = nil
+                    if isSelectingDifferentChat {
                        requestChatBottomPin()
+                    }

                    if let provider = chat.lastUsedProvider,
                       let model = chat.lastUsedModel,
@@ -544,12 +544,14 @@ private func makeToolCallMessage(id: String, date: Date, summary: String = "Ran
@MainActor
@Test func foregroundChatRefreshReloadsSelectedTranscript() async throws {
    let date = Date(timeIntervalSince1970: 1_700_000_100)
+    let staleDetail = makeChatDetail(id: "chat-2", date: date, body: "stale transcript")
    let detail = makeChatDetail(id: "chat-2", date: date, body: "refreshed transcript")
    let client = MockSybilClient(chatDetails: ["chat-2": detail])
    let viewModel = SybilViewModel(settings: testSettings(named: #function)) { _ in client }
    viewModel.isAuthenticated = true
    viewModel.isCheckingSession = false
    viewModel.selectedItem = .chat("chat-2")
+    viewModel.selectedChat = staleDetail

    await viewModel.refreshVisibleContent(refreshCollections: false, refreshSelection: true)

@@ -559,7 +561,7 @@ private func makeToolCallMessage(id: String, date: Date, summary: String = "Ran
    #expect(snapshot.listSearches == 0)
    #expect(snapshot.getChat == 1)
    #expect(viewModel.selectedChat?.messages.first?.content == "refreshed transcript")
-    #expect(viewModel.chatBottomPinRequestID == 1)
+    #expect(viewModel.chatBottomPinRequestID == 0)
 }

@MainActor
@@ -675,6 +677,7 @@ private func makeToolCallMessage(id: String, date: Date, summary: String = "Ran

    #expect(viewModel.displayedMessages.first?.content == "fresh transcript")
    #expect(!viewModel.isLoadingSelection)
+    #expect(viewModel.chatBottomPinRequestID == 1)
 }

@MainActor
@@ -1,4 +1,3 @@
-require "fileutils"
 require "shellwords"

 default_platform(:ios)
@@ -7,56 +6,59 @@ APP_IDENTIFIER = "net.buzzert.sybil2"
 SCHEME = "Sybil"
 TEAM_ID = "DQQH5H6GBD"
 PROFILE_NAME = "Sybil AppStore CI"
-SIGNING_IDENTITY = "Apple Distribution: James Magahern (DQQH5H6GBD)"
 CI_KEYCHAIN_NAME = "sybil_ci_keychain"
 CI_KEYCHAIN_PASSWORD = "sybil-ci-keychain-password"
+CI_KEYCHAIN_DB_PATH = File.expand_path("~/Library/Keychains/#{CI_KEYCHAIN_NAME}-db")
 IOS_ROOT = File.expand_path("..", __dir__)
 PROJECT_FILE = File.join(IOS_ROOT, "Sybil.xcodeproj")
 PROJECT_SPEC = File.join(IOS_ROOT, "project.yml")
-CI_KEYCHAIN_PATH = File.join(File.expand_path("~/Library/Keychains"), CI_KEYCHAIN_NAME)
-CI_KEYCHAIN_DB_PATH = "#{CI_KEYCHAIN_PATH}-db"
-LOGIN_KEYCHAIN_PATH = File.expand_path("~/Library/Keychains/login.keychain")
-LOGIN_KEYCHAIN_DB_PATH = "#{LOGIN_KEYCHAIN_PATH}-db"
+APP_PROJECT_SPEC = File.join(IOS_ROOT, "Apps/Sybil/project.yml")

 def present?(value)
  !value.to_s.strip.empty?
 end

-def release_version
-  tag = ENV["SYBIL_VERSION_TAG"].to_s
-  tag = ENV["GITHUB_REF_NAME"].to_s if !present?(tag)
-  tag = ENV["GITHUB_REF"].to_s.sub(%r{\Arefs/tags/}, "") if !present?(tag)
-  tag = sh("git describe --tags --abbrev=0").strip if !present?(tag)
-  version = tag.sub(%r{\Arelease/}, "").sub(/\Av/, "")
-
-  unless version.match?(/\A\d+\.\d+\.\d+\z/)
-    UI.user_error!("Release tag must look like v1.2.3; got #{tag.inspect}")
-  end
-
-  version
-end
-
 def ci?
  present?(ENV["CI"])
 end

+def release_version
+  tag = ENV["SYBIL_VERSION_TAG"]
+  tag = ENV["GITHUB_REF_NAME"] if !present?(tag)
+  tag = ENV["GITHUB_REF"].to_s.sub(%r{\Arefs/tags/}, "") if !present?(tag)
+  tag = sh("git describe --tags --abbrev=0").strip if !present?(tag)
+  match = tag.to_s.match(%r{\Arelease/ios/v(\d+\.\d+\.\d+)\z})
+
+  unless match
+    UI.user_error!("Release tag must look like release/ios/v1.2.3; got #{tag.inspect}")
+  end
+
+  match[1]
+end
+
 # App Store Connect requires CFBundleVersion to be unique and strictly
 # increasing app-wide (not just per marketing version), so we derive it from
 # the monotonic CI run number rather than querying TestFlight (that query can
 # lag behind builds still processing and hand back a colliding value).
 def build_number
-  value = ENV["SYBIL_BUILD_NUMBER"].to_s
-  value = ENV["GITHUB_RUN_NUMBER"].to_s if !present?(value)
+  value = present?(ENV["SYBIL_BUILD_NUMBER"]) ? ENV["SYBIL_BUILD_NUMBER"] : ENV["GITHUB_RUN_NUMBER"]

-  unless value.match?(/\A\d+\z/)
+  unless value.to_s.match?(/\A\d+\z/)
    UI.user_error!("Build number must come from SYBIL_BUILD_NUMBER/GITHUB_RUN_NUMBER; got #{value.inspect}")
  end

  value.to_i
 end

-def ci_keychain_path
-  File.file?(CI_KEYCHAIN_DB_PATH) ? CI_KEYCHAIN_DB_PATH : CI_KEYCHAIN_PATH
+def stamp_marketing_version(version)
+  contents = File.read(APP_PROJECT_SPEC)
+  updated = contents.sub(/^(\s*MARKETING_VERSION:\s*).*/, "\\1\"#{version}\"")
+
+  if updated == contents
+    UI.user_error!("Could not find MARKETING_VERSION in #{APP_PROJECT_SPEC}")
+  end
+
+  File.write(APP_PROJECT_SPEC, updated)
 end

 platform :ios do
@@ -69,54 +71,37 @@ platform :ios do
    )
  end

-  private_lane :setup_ci_signing do
+  # CI signs headlessly, so match needs a fresh unlocked keychain to import
+  # into. codesign resolves identities through the user keychain *search list*
+  # (first match wins; the --keychain flag does not restrict the lookup), and
+  # other projects' keychains on this runner hold the same identity but are
+  # usually locked — so ours must come first. delete_keychain in the beta
+  # lane's ensure removes both the keychain and its search-list entry, which
+  # also keeps our (later locked) copy from shadowing those other projects.
+  private_lane :prepare_ci_keychain do
    next unless ci?

-    FileUtils.mkdir_p(File.dirname(CI_KEYCHAIN_PATH))
-    sh("security delete-keychain #{CI_KEYCHAIN_PATH.shellescape} || true", log: false)
-    FileUtils.rm_f(CI_KEYCHAIN_PATH)
-    FileUtils.rm_f(CI_KEYCHAIN_DB_PATH)
-
+    delete_keychain(name: CI_KEYCHAIN_NAME) if File.file?(CI_KEYCHAIN_DB_PATH)
    create_keychain(
-      path: CI_KEYCHAIN_PATH,
+      name: CI_KEYCHAIN_NAME,
      password: CI_KEYCHAIN_PASSWORD,
-      default_keychain: false,
      unlock: true,
      timeout: 3600,
-      lock_when_sleeps: true,
      add_to_search_list: false
    )

-    sh("security default-keychain -d user -s #{CI_KEYCHAIN_PATH.shellescape}", log: false)
-    sh("security list-keychains -d user -s #{ci_keychain_path.shellescape}", log: false)
-    sh("security list-keychains -d dynamic -s #{ci_keychain_path.shellescape} || true", log: false)
-    sh("security list-keychains -d common -s #{ci_keychain_path.shellescape} || true", log: false)
+    others = sh("security list-keychains -d user", log: false)
+      .scan(/"([^"]+)"/)
+      .flatten
+      .reject { |path| path.include?(CI_KEYCHAIN_NAME) }
+    sh("security list-keychains -d user -s #{([CI_KEYCHAIN_DB_PATH] + others).shelljoin}")

-    ENV["MATCH_KEYCHAIN_NAME"] = CI_KEYCHAIN_PATH
+    ENV["MATCH_KEYCHAIN_NAME"] = CI_KEYCHAIN_NAME
    ENV["MATCH_KEYCHAIN_PASSWORD"] = CI_KEYCHAIN_PASSWORD
-    ENV["MATCH_READONLY"] = "true"
-  end
-
-  private_lane :cleanup_ci_signing do
-    next unless ci?
-
-    if File.file?(LOGIN_KEYCHAIN_DB_PATH) || File.file?(LOGIN_KEYCHAIN_PATH)
-      sh("security default-keychain -d user -s #{LOGIN_KEYCHAIN_PATH.shellescape} || true", log: false)
-      sh("security list-keychains -d user -s #{LOGIN_KEYCHAIN_DB_PATH.shellescape} || true", log: false)
-    end
-    sh("security delete-keychain #{ci_keychain_path.shellescape} || true", log: false)
-    FileUtils.rm_f(CI_KEYCHAIN_PATH)
-    FileUtils.rm_f(CI_KEYCHAIN_DB_PATH)
-  rescue => error
-    UI.message("Unable to delete temporary CI keychain: #{error.message}")
-  ensure
-    ENV.delete("MATCH_KEYCHAIN_NAME")
-    ENV.delete("MATCH_KEYCHAIN_PASSWORD")
-    ENV.delete("MATCH_READONLY")
  end

  private_lane :sync_signing do |options|
-    match_options = {
+    match(
      type: "appstore",
      readonly: options.fetch(:readonly),
      app_identifier: APP_IDENTIFIER,
@@ -127,28 +112,7 @@ platform :ios do
      git_full_name: "Sybil Release Bot",
      git_user_email: "james.magahern@me.com",
      api_key: options.fetch(:api_key)
-    }
-    match_options[:keychain_name] = ENV["MATCH_KEYCHAIN_NAME"] if present?(ENV["MATCH_KEYCHAIN_NAME"])
-    match_options[:keychain_password] = ENV["MATCH_KEYCHAIN_PASSWORD"] if ENV.key?("MATCH_KEYCHAIN_PASSWORD")
-
-    match(match_options)
-  end
-
-  private_lane :verify_ci_signing do
-    next unless ci?
-
-    if File.file?(ci_keychain_path)
-      password = ENV.fetch("MATCH_KEYCHAIN_PASSWORD", "")
-      sh("security unlock-keychain -p #{password.shellescape} #{ci_keychain_path.shellescape}", log: false)
-      sh("security set-key-partition-list -S apple-tool:,apple:,codesign: -s -k #{password.shellescape} #{ci_keychain_path.shellescape}", log: false)
-    end
-
-    identities = sh("security find-identity -v -p codesigning #{ci_keychain_path.shellescape}", log: false)
-    UI.message(identities)
-
-    unless identities.include?(SIGNING_IDENTITY)
-      UI.user_error!("The CI keychain search list does not contain #{SIGNING_IDENTITY}")
-    end
+    )
  end

  desc "Create or update match signing assets"
@@ -158,43 +122,29 @@ platform :ios do

  desc "Build and upload to TestFlight"
  lane :beta do
-    setup_ci_signing
+    prepare_ci_keychain

    api_key = app_store_api_key

-    sh("xcodegen --spec #{PROJECT_SPEC.shellescape}")
+    version = release_version
+    stamp_marketing_version(version)
+    sh("xcodegen", "--spec", PROJECT_SPEC)

-    increment_version_number(
-      version_number: release_version,
-      xcodeproj: PROJECT_FILE
-    )
-
-    increment_build_number(
-      build_number: build_number,
-      xcodeproj: PROJECT_FILE
-    )
+    increment_version_number(version_number: version, xcodeproj: PROJECT_FILE)
+    increment_build_number(build_number: build_number, xcodeproj: PROJECT_FILE)

    sync_signing(api_key: api_key, readonly: true)
-    verify_ci_signing
-
-    xcargs = [
-      "DEVELOPMENT_TEAM=#{TEAM_ID.shellescape}",
-      "CODE_SIGN_STYLE=Manual",
-      "CODE_SIGN_IDENTITY=Apple\\ Distribution",
-      "PROVISIONING_PROFILE_SPECIFIER=#{PROFILE_NAME.shellescape}"
-    ]
-
-    if ci?
-      xcargs << "CODE_SIGN_KEYCHAIN=#{ci_keychain_path.shellescape}"
-      xcargs << "OTHER_CODE_SIGN_FLAGS=#{("--keychain #{ci_keychain_path}").shellescape}"
-    end

    build_app(
      project: PROJECT_FILE,
      scheme: SCHEME,
      export_method: "app-store",
-      codesigning_identity: "Apple Distribution",
-      xcargs: xcargs.join(" "),
+      xcargs: [
+        "DEVELOPMENT_TEAM=#{TEAM_ID.shellescape}",
+        "CODE_SIGN_STYLE=Manual",
+        "CODE_SIGN_IDENTITY=Apple\\ Distribution",
+        "PROVISIONING_PROFILE_SPECIFIER=#{PROFILE_NAME.shellescape}"
+      ].join(" "),
      export_options: {
        signingStyle: "manual",
        teamID: TEAM_ID,
@@ -209,6 +159,6 @@ platform :ios do
      skip_waiting_for_build_processing: true
    )
  ensure
-    cleanup_ci_signing
+    delete_keychain(name: CI_KEYCHAIN_NAME) if ci? && File.file?(CI_KEYCHAIN_DB_PATH)
  end
 end
@@ -1,7 +1,7 @@
 # Sybil Server

 Backend API for:
- LLM multiplexer (OpenAI Responses / Anthropic / xAI Chat Completions-compatible Grok / Hermes Agent)
+- LLM multiplexer (OpenAI Responses / Anthropic / xAI Chat Completions-compatible Grok / Gemini / Hermes Agent)
 - Personal chat database (chats/messages + LLM call log)

 ## Stack
@@ -43,6 +43,7 @@ If `ADMIN_TOKEN` is not set, the server runs in open mode (dev).
 - `OPENAI_API_KEY`
 - `ANTHROPIC_API_KEY`
 - `XAI_API_KEY`
+- `GEMINI_API_KEY`
 - `HERMES_AGENT_API_BASE_URL` (`http://127.0.0.1:8642/v1` by default; include the `/v1` suffix)
 - `HERMES_AGENT_API_KEY` (enables the Hermes Agent provider; set to Hermes `API_SERVER_KEY`, or any non-empty value if that local server does not require auth)
 - `HERMES_AGENT_MODEL` (optional fallback/override model id; defaults client-side to `hermes-agent`)
@@ -50,7 +51,7 @@ If `ADMIN_TOKEN` is not set, the server runs in open mode (dev).
 - `CHAT_WEB_SEARCH_ENGINE` (`exa` by default, or `searxng` for chat tool calls only)
 - `SEARXNG_BASE_URL` (required when `CHAT_WEB_SEARCH_ENGINE=searxng`; instance must allow `format=json`)
 - `CHAT_MAX_TOOL_ROUNDS` (`100` by default; maximum model/tool result cycles per chat completion)
- `CHAT_CODEX_TOOL_ENABLED` (`false` by default; enables the `codex_exec` chat tool for OpenAI/xAI)
+- `CHAT_CODEX_TOOL_ENABLED` (`false` by default; enables the `codex_exec` chat tool for managed-tool providers)
 - `CHAT_CODEX_REMOTE_HOST` (required when Codex tool is enabled; SSH host/IP or `user@host`)
 - `CHAT_CODEX_REMOTE_USER` (optional SSH user when host does not include one)
 - `CHAT_CODEX_REMOTE_PORT` (`22` by default)
@@ -58,7 +59,7 @@ If `ADMIN_TOKEN` is not set, the server runs in open mode (dev).
 - `CHAT_CODEX_SSH_KEY_PATH` (recommended: path to a read-only mounted private key)
 - `CHAT_CODEX_SSH_PRIVATE_KEY_B64` (optional fallback private key delivery)
 - `CHAT_CODEX_EXEC_TIMEOUT_MS` (`600000` by default)
- `CHAT_SHELL_TOOL_ENABLED` (`false` by default; enables the `shell_exec` chat tool for OpenAI/xAI on the same devbox)
+- `CHAT_SHELL_TOOL_ENABLED` (`false` by default; enables the `shell_exec` chat tool for managed-tool providers on the same devbox)
 - `CHAT_SHELL_EXEC_TIMEOUT_MS` (`120000` by default)

 ## API
@@ -13,6 +13,7 @@ enum Provider {
  openai
  anthropic
  xai
+  gemini
  hermes_agent @map("hermes-agent")
 }

@@ -66,6 +66,7 @@ const EnvSchema = z.object({
  OPENAI_API_KEY: z.string().optional(),
  ANTHROPIC_API_KEY: z.string().optional(),
  XAI_API_KEY: z.string().optional(),
+  GEMINI_API_KEY: z.string().optional(),
  HERMES_AGENT_API_BASE_URL: HermesAgentApiBaseUrlSchema,
  HERMES_AGENT_API_KEY: OptionalTrimmedStringSchema,
  HERMES_AGENT_MODEL: OptionalTrimmedStringSchema,
@@ -0,0 +1,501 @@
+import {
+  buildChatToolSystemPrompt,
+  executeToolCallAndBuildEvent,
+  getEnabledChatTools,
+  getUnstreamedText,
+  looksLikeDanglingToolIntent,
+  MAX_DANGLING_TOOL_INTENT_RETRIES,
+  MAX_TOOL_ROUNDS,
+  prepareToolCallExecution,
+  type NormalizedToolCall,
+  type ToolAwareCompletionParams,
+  type ToolAwareCompletionResult,
+  type ToolAwareStreamingEvent,
+  type ToolAwareUsage,
+  type ToolExecutionEvent,
+} from "../chat-tools.js";
+import {
+  buildImageSummaryText,
+  buildTextAttachmentPrompt,
+  buildTopLevelSystemPrompt,
+  getImageAttachments,
+  getTextAttachments,
+  parseImageDataUrl,
+} from "../message-content.js";
+import type { ChatMessage } from "../types.js";
+
+type GeminiClient = {
+  apiKey: string;
+  baseURL: string;
+};
+
+const INTERNAL_CORRECTION =
+  "Internal correction: the previous assistant message claimed it would run a tool, but no tool call was made. If the task needs an available tool, call it now. Otherwise provide the final answer directly without saying you will run a tool.";
+
+function normalizeModelResourceName(model: string) {
+  const trimmed = model.trim().replace(/^\/+/, "");
+  return trimmed.startsWith("models/") || trimmed.startsWith("tunedModels/") ? trimmed : `models/${trimmed}`;
+}
+
+function geminiUrl(client: GeminiClient, model: string, method: "generateContent" | "streamGenerateContent", extraParams: Record<string, string> = {}) {
+  const url = new URL(`${client.baseURL.replace(/\/+$/, "")}/${normalizeModelResourceName(model)}:${method}`);
+  url.searchParams.set("key", client.apiKey);
+  for (const [key, value] of Object.entries(extraParams)) {
+    url.searchParams.set(key, value);
+  }
+  return url;
+}
+
+function generationConfig(params: Pick<ToolAwareCompletionParams, "temperature" | "maxTokens">) {
+  const config: Record<string, unknown> = {};
+  if (params.temperature !== undefined) config.temperature = params.temperature;
+  if (params.maxTokens !== undefined) config.maxOutputTokens = params.maxTokens;
+  return Object.keys(config).length ? config : undefined;
+}
+
+function toGeminiJsonSchema(schema: unknown): Record<string, unknown> | undefined {
+  if (!schema || typeof schema !== "object" || Array.isArray(schema)) return undefined;
+  const input = schema as Record<string, unknown>;
+  const output: Record<string, unknown> = {};
+
+  if (typeof input.type === "string") output.type = input.type;
+  if (typeof input.description === "string") output.description = input.description;
+  if (typeof input.format === "string") output.format = input.format;
+  if (typeof input.nullable === "boolean") output.nullable = input.nullable;
+  if (Array.isArray(input.enum)) output.enum = input.enum.filter((value) => typeof value === "string");
+  if (Array.isArray(input.required)) output.required = input.required.filter((value) => typeof value === "string");
+
+  const items = toGeminiJsonSchema(input.items);
+  if (items) output.items = items;
+
+  if (input.properties && typeof input.properties === "object" && !Array.isArray(input.properties)) {
+    const properties: Record<string, unknown> = {};
+    for (const [key, value] of Object.entries(input.properties)) {
+      const propertySchema = toGeminiJsonSchema(value);
+      if (propertySchema) properties[key] = propertySchema;
+    }
+    if (Object.keys(properties).length) output.properties = properties;
+  }
+
+  return Object.keys(output).length ? output : undefined;
+}
+
+function toGeminiTools(tools: any[]) {
+  const functionDeclarations = tools
+    .map((tool) => {
+      if (tool?.type !== "function") return null;
+      const declaration: Record<string, unknown> = {
+        name: tool.function.name,
+        description: tool.function.description,
+      };
+      const parameters = toGeminiJsonSchema(tool.function.parameters);
+      if (parameters) declaration.parameters = parameters;
+      return declaration;
+    })
+    .filter(Boolean);
+
+  return functionDeclarations.length ? [{ functionDeclarations }] : undefined;
+}
+
+function toContentParts(message: ChatMessage) {
+  const imageAttachments = getImageAttachments(message);
+  const textAttachments = getTextAttachments(message);
+  const parts: Array<Record<string, unknown>> = [];
+
+  for (const attachment of imageAttachments) {
+    const source = parseImageDataUrl(attachment);
+    parts.push({
+      inlineData: {
+        mimeType: source.mediaType,
+        data: source.data,
+      },
+    });
+  }
+
+  const imageSummary = buildImageSummaryText(imageAttachments);
+  if (imageSummary) {
+    parts.push({ text: imageSummary });
+  }
+
+  for (const attachment of textAttachments) {
+    parts.push({ text: buildTextAttachmentPrompt(attachment) });
+  }
+
+  if (message.content.trim()) {
+    parts.push({ text: message.content });
+  }
+
+  return parts.length ? parts : [{ text: "" }];
+}
+
+function buildConversationContent(message: ChatMessage) {
+  if (message.role === "system") {
+    throw new Error("System messages must be handled separately for Gemini.");
+  }
+
+  if (message.role === "tool") {
+    const name = message.name?.trim() || "tool";
+    return {
+      role: "user",
+      parts: [{ text: `Tool output (${name}):\n${message.content}` }],
+    };
+  }
+
+  return {
+    role: message.role === "assistant" ? "model" : "user",
+    parts: toContentParts(message),
+  };
+}
+
+function buildBaseContents(messages: ChatMessage[]) {
+  return messages.filter((message) => message.role !== "system").map((message) => buildConversationContent(message));
+}
+
+function buildSystemInstruction(params: ToolAwareCompletionParams, toolSystemPrompt?: string) {
+  const text = buildTopLevelSystemPrompt(params.messages, params.userLocation, toolSystemPrompt);
+  return text ? { parts: [{ text }] } : undefined;
+}
+
+function mergeUsage(acc: Required<ToolAwareUsage>, usage: any) {
+  const normalized = normalizeUsage(usage);
+  if (!normalized) return false;
+  acc.inputTokens += normalized.inputTokens;
+  acc.outputTokens += normalized.outputTokens;
+  acc.totalTokens += normalized.totalTokens;
+  return true;
+}
+
+function normalizeUsage(usage: any) {
+  if (!usage) return null;
+  const inputTokens = usage.promptTokenCount ?? 0;
+  const outputTokens = usage.candidatesTokenCount ?? 0;
+  const totalTokens = usage.totalTokenCount ?? inputTokens + outputTokens;
+  return { inputTokens, outputTokens, totalTokens };
+}
+
+function getCandidate(response: any) {
+  return Array.isArray(response?.candidates) ? response.candidates[0] : null;
+}
+
+function getParts(response: any) {
+  const parts = getCandidate(response)?.content?.parts;
+  return Array.isArray(parts) ? parts : [];
+}
+
+function extractText(response: any) {
+  return getParts(response)
+    .map((part: any) => (typeof part?.text === "string" ? part.text : ""))
+    .join("");
+}
+
+function stringifyToolArgs(args: unknown) {
+  try {
+    return JSON.stringify(args ?? {});
+  } catch {
+    return "{}";
+  }
+}
+
+function normalizeToolCallsFromParts(parts: any[], round: number): NormalizedToolCall[] {
+  return parts
+    .filter((part) => part?.functionCall)
+    .map((part, index) => ({
+      id: part.functionCall.id ?? `tool_call_${round}_${index}`,
+      name: part.functionCall.name ?? "unknown_tool",
+      arguments: stringifyToolArgs(part.functionCall.args),
+    }));
+}
+
+function buildFunctionResponsePart(call: NormalizedToolCall, toolResult: unknown) {
+  return {
+    functionResponse: {
+      id: call.id,
+      name: call.name,
+      response: toolResult,
+    },
+  };
+}
+
+function appendCorrection(conversation: any[], text: string) {
+  conversation.push({ role: "model", parts: [{ text }] });
+  conversation.push({ role: "user", parts: [{ text: INTERNAL_CORRECTION }] });
+}
+
+async function parseGeminiResponse(response: Response) {
+  const bodyText = await response.text();
+  let body: any = null;
+  try {
+    body = bodyText ? JSON.parse(bodyText) : null;
+  } catch {
+    body = { raw: bodyText };
+  }
+
+  if (!response.ok) {
+    throw new Error(body?.error?.message ?? `Gemini API request failed with status ${response.status}.`);
+  }
+
+  return body;
+}
+
+async function generateContent(params: ToolAwareCompletionParams, body: Record<string, unknown>) {
+  const response = await fetch(geminiUrl(params.client, params.model, "generateContent"), {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(body),
+  });
+  return parseGeminiResponse(response);
+}
+
+function getFailureMessage(response: any, text: string, toolCallCount: number) {
+  const promptBlockReason = response?.promptFeedback?.blockReason;
+  if (promptBlockReason) return `Gemini prompt blocked: ${promptBlockReason}.`;
+
+  const candidate = getCandidate(response);
+  const finishReason = candidate?.finishReason;
+  if (!finishReason || finishReason === "STOP" || finishReason === "MAX_TOKENS") return null;
+  if (text || toolCallCount > 0) return null;
+  return candidate?.finishMessage ?? `Gemini response stopped: ${finishReason}.`;
+}
+
+function buildRequest(params: ToolAwareCompletionParams, conversation: any[], enabledTools: any[] = []) {
+  const tools = toGeminiTools(enabledTools);
+  return {
+    contents: conversation,
+    systemInstruction: buildSystemInstruction(params, enabledTools.length ? buildChatToolSystemPrompt(params) : undefined),
+    generationConfig: generationConfig(params),
+    tools,
+    toolConfig: tools ? { functionCallingConfig: { mode: "AUTO" } } : undefined,
+  };
+}
+
+export async function completeWithGeminiApi(params: ToolAwareCompletionParams): Promise<ToolAwareCompletionResult> {
+  const enabledTools = getEnabledChatTools(params);
+  const conversation = buildBaseContents(params.messages);
+  const rawResponses: unknown[] = [];
+  const toolEvents: ToolExecutionEvent[] = [];
+  const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
+  let sawUsage = false;
+  let totalToolCalls = 0;
+  let danglingToolIntentRetries = 0;
+
+  for (let round = 0; round < MAX_TOOL_ROUNDS; round += 1) {
+    const response = await generateContent(params, buildRequest(params, conversation, enabledTools));
+    rawResponses.push(response);
+    sawUsage = mergeUsage(usageAcc, response?.usageMetadata) || sawUsage;
+
+    const parts = getParts(response);
+    const text = extractText(response);
+    const normalizedToolCalls = normalizeToolCallsFromParts(parts, round);
+    const failureMessage = getFailureMessage(response, text, normalizedToolCalls.length);
+    if (failureMessage) throw new Error(failureMessage);
+
+    if (!normalizedToolCalls.length) {
+      if (danglingToolIntentRetries < MAX_DANGLING_TOOL_INTENT_RETRIES && looksLikeDanglingToolIntent(text)) {
+        danglingToolIntentRetries += 1;
+        appendCorrection(conversation, text);
+        continue;
+      }
+      return {
+        text,
+        usage: sawUsage ? usageAcc : undefined,
+        raw: { responses: rawResponses, toolCallsUsed: totalToolCalls, api: "gemini.generateContent" },
+        toolEvents,
+      };
+    }
+
+    totalToolCalls += normalizedToolCalls.length;
+    conversation.push({ role: "model", parts });
+
+    const toolResultParts: any[] = [];
+    for (const call of normalizedToolCalls) {
+      const { execution } = prepareToolCallExecution(call);
+      const { event, toolResult } = await executeToolCallAndBuildEvent(call, execution, params);
+      toolEvents.push(event);
+      toolResultParts.push(buildFunctionResponsePart(call, toolResult));
+    }
+
+    conversation.push({ role: "user", parts: toolResultParts });
+  }
+
+  return {
+    text: "I reached the tool-call limit while gathering information. Please narrow the request and try again.",
+    usage: sawUsage ? usageAcc : undefined,
+    raw: { responses: rawResponses, toolCallsUsed: totalToolCalls, toolCallLimitReached: true, api: "gemini.generateContent" },
+    toolEvents,
+  };
+}
+
+function findSseBoundary(buffer: string) {
+  const crlf = buffer.indexOf("\r\n\r\n");
+  const lf = buffer.indexOf("\n\n");
+  if (crlf === -1) return lf === -1 ? null : { index: lf, length: 2 };
+  if (lf === -1) return { index: crlf, length: 4 };
+  return crlf < lf ? { index: crlf, length: 4 } : { index: lf, length: 2 };
+}
+
+function parseSseEvent(rawEvent: string) {
+  const data = rawEvent
+    .split(/\r?\n/)
+    .filter((line) => line.startsWith("data:"))
+    .map((line) => line.slice("data:".length).trimStart())
+    .join("\n")
+    .trim();
+  if (!data || data === "[DONE]") return null;
+  return JSON.parse(data);
+}
+
+async function* streamGeminiResponses(params: ToolAwareCompletionParams, body: Record<string, unknown>) {
+  const response = await fetch(geminiUrl(params.client, params.model, "streamGenerateContent", { alt: "sse" }), {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(body),
+  });
+
+  if (!response.ok) {
+    await parseGeminiResponse(response);
+    return;
+  }
+
+  if (!response.body) {
+    throw new Error("Gemini stream response did not include a body.");
+  }
+
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+  let buffer = "";
+
+  while (true) {
+    const { value, done } = await reader.read();
+    if (done) break;
+    buffer += decoder.decode(value, { stream: true });
+    let boundary = findSseBoundary(buffer);
+    while (boundary) {
+      const rawEvent = buffer.slice(0, boundary.index);
+      buffer = buffer.slice(boundary.index + boundary.length);
+      const event = parseSseEvent(rawEvent);
+      if (event) yield event;
+      boundary = findSseBoundary(buffer);
+    }
+  }
+
+  buffer += decoder.decode();
+  const tail = buffer.trim();
+  if (tail) {
+    const event = parseSseEvent(tail);
+    if (event) yield event;
+  }
+}
+
+export async function* streamWithGeminiApi(params: ToolAwareCompletionParams): AsyncGenerator<ToolAwareStreamingEvent> {
+  const enabledTools = getEnabledChatTools(params);
+  const conversation = buildBaseContents(params.messages);
+  const rawResponses: unknown[] = [];
+  const toolEvents: ToolExecutionEvent[] = [];
+  const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
+  let sawUsage = false;
+  let totalToolCalls = 0;
+  let danglingToolIntentRetries = 0;
+
+  if (!enabledTools.length) {
+    let text = "";
+    let latestUsage: any = null;
+    for await (const response of streamGeminiResponses(params, buildRequest(params, conversation))) {
+      rawResponses.push(response);
+      if (response?.usageMetadata) latestUsage = response.usageMetadata;
+      const failureMessage = getFailureMessage(response, extractText(response), 0);
+      if (failureMessage) throw new Error(failureMessage);
+      const delta = extractText(response);
+      if (delta) {
+        text += delta;
+        yield { type: "delta", text: delta };
+      }
+    }
+
+    sawUsage = mergeUsage(usageAcc, latestUsage) || sawUsage;
+
+    yield {
+      type: "done",
+      result: {
+        text,
+        usage: sawUsage ? usageAcc : undefined,
+        raw: { streamed: true, responses: rawResponses, toolCallsUsed: 0, api: "gemini.streamGenerateContent" },
+        toolEvents: [],
+      },
+    };
+    return;
+  }
+
+  for (let round = 0; round < MAX_TOOL_ROUNDS; round += 1) {
+    const roundParts: any[] = [];
+    let roundText = "";
+    let latestRoundResponse: any = null;
+    let latestRoundUsage: any = null;
+
+    for await (const response of streamGeminiResponses(params, buildRequest(params, conversation, enabledTools))) {
+      rawResponses.push(response);
+      latestRoundResponse = response;
+      if (response?.usageMetadata) latestRoundUsage = response.usageMetadata;
+      roundParts.push(...getParts(response));
+      roundText += extractText(response);
+    }
+
+    sawUsage = mergeUsage(usageAcc, latestRoundUsage) || sawUsage;
+
+    const normalizedToolCalls = normalizeToolCallsFromParts(roundParts, round);
+    const failureMessage = getFailureMessage(latestRoundResponse ?? { candidates: [{ content: { parts: roundParts } }] }, roundText, normalizedToolCalls.length);
+    if (failureMessage) throw new Error(failureMessage);
+
+    if (!normalizedToolCalls.length) {
+      if (danglingToolIntentRetries < MAX_DANGLING_TOOL_INTENT_RETRIES && looksLikeDanglingToolIntent(roundText)) {
+        danglingToolIntentRetries += 1;
+        appendCorrection(conversation, roundText);
+        continue;
+      }
+      const unstreamedText = getUnstreamedText(roundText, "");
+      if (unstreamedText) {
+        yield { type: "delta", text: unstreamedText };
+      }
+      yield {
+        type: "done",
+        result: {
+          text: roundText,
+          usage: sawUsage ? usageAcc : undefined,
+          raw: { streamed: true, responses: rawResponses, toolCallsUsed: totalToolCalls, api: "gemini.streamGenerateContent" },
+          toolEvents,
+        },
+      };
+      return;
+    }
+
+    totalToolCalls += normalizedToolCalls.length;
+    conversation.push({ role: "model", parts: roundParts });
+
+    const toolResultParts: any[] = [];
+    for (const call of normalizedToolCalls) {
+      const { event: initiatedEvent, execution } = prepareToolCallExecution(call);
+      yield { type: "tool_call", event: initiatedEvent };
+      const { event, toolResult } = await executeToolCallAndBuildEvent(call, execution, params);
+      toolEvents.push(event);
+      yield { type: "tool_call", event };
+      toolResultParts.push(buildFunctionResponsePart(call, toolResult));
+    }
+
+    conversation.push({ role: "user", parts: toolResultParts });
+  }
+
+  yield {
+    type: "done",
+    result: {
+      text: "I reached the tool-call limit while gathering information. Please narrow the request and try again.",
+      usage: sawUsage ? usageAcc : undefined,
+      raw: {
+        streamed: true,
+        responses: rawResponses,
+        toolCallsUsed: totalToolCalls,
+        toolCallLimitReached: true,
+        api: "gemini.streamGenerateContent",
+      },
+      toolEvents,
+    },
+  };
+}
@@ -28,6 +28,45 @@ import type { ChatMessage } from "../types.js";
 const INTERNAL_CORRECTION =
  "Internal correction: the previous assistant message claimed it would run a tool, but no tool call was made. If the task needs an available tool, call it now. Otherwise provide the final answer directly without saying you will run a tool.";

+const DEFAULT_ANTHROPIC_MAX_TOKENS = 128_000;
+const MODEL_MAX_TOKENS_CACHE_MS = 24 * 60 * 60 * 1000;
+
+const modelMaxTokensCache = new Map<string, { maxTokens: number; expiresAt: number }>();
+
+function readMaxTokens(value: unknown) {
+  return Number.isSafeInteger(value) && (value as number) > 0 ? (value as number) : undefined;
+}
+
+function getModelInfoMaxTokens(modelInfo: any) {
+  return readMaxTokens(modelInfo?.max_tokens) ?? readMaxTokens(modelInfo?.maxTokens);
+}
+
+async function getMessagesMaxTokens(params: ToolAwareCompletionParams) {
+  if (params.maxTokens) return params.maxTokens;
+
+  const cached = modelMaxTokensCache.get(params.model);
+  if (cached && cached.expiresAt > Date.now()) return cached.maxTokens;
+
+  try {
+    const retrieve = params.client?.models?.retrieve;
+    if (typeof retrieve === "function") {
+      const modelInfo = await retrieve.call(params.client.models, params.model);
+      const maxTokens = getModelInfoMaxTokens(modelInfo);
+      if (maxTokens) {
+        modelMaxTokensCache.set(params.model, {
+          maxTokens,
+          expiresAt: Date.now() + MODEL_MAX_TOKENS_CACHE_MS,
+        });
+        return maxTokens;
+      }
+    }
+  } catch {
+    // Fall back to the documented max for Claude Opus 4.8 and related high-output models.
+  }
+
+  return DEFAULT_ANTHROPIC_MAX_TOKENS;
+}
+
 function toTools(tools: any[]) {
  return tools
    .map((tool) => {
@@ -160,11 +199,12 @@ function mergeUsage(acc: Required<ToolAwareUsage>, usage: any) {

 export async function completeWithMessagesApi(params: ToolAwareCompletionParams): Promise<ToolAwareCompletionResult> {
  const enabledTools = getEnabledChatTools(params);
+  const maxTokens = await getMessagesMaxTokens(params);
  if (!enabledTools.length) {
    const response = await params.client.messages.create({
      model: params.model,
      system: buildTopLevelSystemPrompt(params.messages, params.userLocation),
-      max_tokens: params.maxTokens ?? 1024,
+      max_tokens: maxTokens,
      temperature: params.temperature,
      messages: buildBaseMessages(params),
    } as any);
@@ -192,7 +232,7 @@ export async function completeWithMessagesApi(params: ToolAwareCompletionParams)
    const response = await params.client.messages.create({
      model: params.model,
      system: buildTopLevelSystemPrompt(params.messages, params.userLocation, buildChatToolSystemPrompt(params)),
-      max_tokens: params.maxTokens ?? 1024,
+      max_tokens: maxTokens,
      temperature: params.temperature,
      messages: conversation,
      tools: toTools(enabledTools),
@@ -248,6 +288,7 @@ export async function completeWithMessagesApi(params: ToolAwareCompletionParams)

 export async function* streamWithMessagesApi(params: ToolAwareCompletionParams): AsyncGenerator<ToolAwareStreamingEvent> {
  const enabledTools = getEnabledChatTools(params);
+  const maxTokens = await getMessagesMaxTokens(params);
  if (!enabledTools.length) {
    const rawResponses: unknown[] = [];
    const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
@@ -259,7 +300,7 @@ export async function* streamWithMessagesApi(params: ToolAwareCompletionParams):
    const stream = await params.client.messages.create({
      model: params.model,
      system: buildTopLevelSystemPrompt(params.messages, params.userLocation),
-      max_tokens: params.maxTokens ?? 1024,
+      max_tokens: maxTokens,
      temperature: params.temperature,
      messages: buildBaseMessages(params),
      stream: true,
@@ -315,7 +356,7 @@ export async function* streamWithMessagesApi(params: ToolAwareCompletionParams):
    const stream = await params.client.messages.create({
      model: params.model,
      system: buildTopLevelSystemPrompt(params.messages, params.userLocation, buildChatToolSystemPrompt(params)),
-      max_tokens: params.maxTokens ?? 1024,
+      max_tokens: maxTokens,
      temperature: params.temperature,
      messages: conversation,
      tools: toTools(enabledTools),
@@ -5,10 +5,11 @@ import {
  type ToolAwareStreamingEvent,
 } from "./chat-tools.js";
 import { completeWithChatCompletionsApi, streamWithChatCompletionsApi } from "./protocols/chat-completions-api.js";
+import { completeWithGeminiApi, streamWithGeminiApi } from "./protocols/gemini-api.js";
 import { completeWithMessagesApi, streamWithMessagesApi } from "./protocols/messages-api.js";
 import { completeWithResponsesApi, streamWithResponsesApi } from "./protocols/responses-api.js";
 import { env } from "../env.js";
-import { anthropicClient, hermesAgentClient, isHermesAgentConfigured, openaiClient, xaiClient } from "./providers.js";
+import { anthropicClient, geminiClient, hermesAgentClient, isHermesAgentConfigured, openaiClient, xaiClient } from "./providers.js";
 import type { ChatMessage, Provider } from "./types.js";

 type ProviderAdapterParams = {
@@ -27,7 +28,7 @@ export type ProviderChatAdapter = {
  stream(params: ProviderAdapterParams): AsyncGenerator<ToolAwareStreamingEvent>;
 };

-type ChatProtocolId = "chat-completions" | "messages" | "responses";
+type ChatProtocolId = "chat-completions" | "gemini" | "messages" | "responses";

 type ChatProtocol = {
  id: ChatProtocolId;
@@ -39,6 +40,7 @@ type ModelCatalogSpec = {
  enabled?: () => boolean;
  fetchModels(client: any): Promise<string[]>;
  fallbackModels?: () => string[];
+  sortModels?: (models: string[]) => string[];
 };

 type ProviderBackendSpec = {
@@ -61,6 +63,12 @@ const messagesProtocol: ChatProtocol = {
  stream: streamWithMessagesApi,
 };

+const geminiProtocol: ChatProtocol = {
+  id: "gemini",
+  complete: completeWithGeminiApi,
+  stream: streamWithGeminiApi,
+};
+
 const responsesProtocol: ChatProtocol = {
  id: "responses",
  complete: completeWithResponsesApi,
@@ -77,6 +85,10 @@ function modelIdsFromListResponse(page: any) {
    : [];
 }

+function stripModelResourcePrefix(model: string) {
+  return model.startsWith("models/") ? model.slice("models/".length) : model;
+}
+
 function isLikelyResponsesApiModel(model: string) {
  const id = model.toLowerCase();
  if (id.includes("embedding") || id.includes("moderation")) return false;
@@ -86,6 +98,37 @@ function isLikelyResponsesApiModel(model: string) {
  return /^(gpt-|o\d|chatgpt-)/.test(id);
 }

+function isLikelyGeminiChatModel(model: string) {
+  const id = model.toLowerCase();
+  if (!id.startsWith("gemini-")) return false;
+  if (id.includes("embedding") || id.includes("embed")) return false;
+  if (id.includes("image") || id.includes("imagen") || id.includes("veo")) return false;
+  if (id.includes("audio") || id.includes("tts") || id.includes("live")) return false;
+  if (id.includes("computer-use") || id.includes("robotics")) return false;
+  return true;
+}
+
+function preferGeminiModels(models: string[]) {
+  const preferred = [
+    "gemini-3.5-flash",
+    "gemini-flash-latest",
+    "gemini-3.1-flash-lite",
+    "gemini-3-flash-preview",
+    "gemini-pro-latest",
+  ];
+  const modelSet = new Set(models);
+  return [...preferred.filter((model) => modelSet.delete(model)), ...[...modelSet].sort((a, b) => a.localeCompare(b))];
+}
+
+async function fetchJson(url: URL): Promise<any> {
+  const response = await fetch(url);
+  const body: any = await response.json().catch(() => null);
+  if (!response.ok) {
+    throw new Error(body?.error?.message ?? `Gemini model fetch failed with status ${response.status}.`);
+  }
+  return body;
+}
+
 function withClient(params: ProviderAdapterParams, client: any, enabledTools?: string[]): ToolAwareCompletionParams {
  return {
    client,
@@ -160,6 +203,29 @@ const backendSpecs: Record<Provider, ProviderBackendSpec> = {
      },
    },
  },
+  gemini: {
+    createClient: geminiClient,
+    plainProtocol: geminiProtocol,
+    toolProtocol: geminiProtocol,
+    managedTools: true,
+    modelCatalog: {
+      async fetchModels(client) {
+        const url = new URL(`${client.baseURL.replace(/\/+$/, "")}/models`);
+        url.searchParams.set("key", client.apiKey);
+        url.searchParams.set("pageSize", "1000");
+        const page = await fetchJson(url);
+        return Array.isArray(page?.models)
+          ? page.models
+              .filter((model: any) => Array.isArray(model?.supportedGenerationMethods) && model.supportedGenerationMethods.includes("generateContent"))
+              .map((model: any) => model?.name)
+              .filter((id: unknown): id is string => typeof id === "string")
+              .map(stripModelResourcePrefix)
+              .filter(isLikelyGeminiChatModel)
+          : [];
+      },
+      sortModels: preferGeminiModels,
+    },
+  },
  "hermes-agent": {
    createClient: hermesAgentClient,
    plainProtocol: chatCompletionsProtocol,
@@ -209,7 +275,8 @@ export function listModelCatalogProviders(): Provider[] {
 export async function fetchProviderCatalogModels(provider: Provider) {
  const spec = backendSpecs[provider].modelCatalog;
  if (!spec) return [];
-  return uniqSorted(await spec.fetchModels(backendSpecs[provider].createClient()));
+  const models = uniqSorted(await spec.fetchModels(backendSpecs[provider].createClient()));
+  return spec.sortModels ? spec.sortModels(models) : models;
 }

 export function getProviderCatalogFallbackModels(provider: Provider) {
@@ -6,6 +6,7 @@ const apiToPrismaProvider = {
  openai: "openai",
  anthropic: "anthropic",
  xai: "xai",
+  gemini: "gemini",
  "hermes-agent": "hermes_agent",
 } as const satisfies Record<Provider, PrismaProvider>;

@@ -13,6 +14,7 @@ const prismaToApiProvider = {
  openai: "openai",
  anthropic: "anthropic",
  xai: "xai",
+  gemini: "gemini",
  hermes_agent: "hermes-agent",
  "hermes-agent": "hermes-agent",
 } as const satisfies Record<PrismaProvider | "hermes-agent", Provider>;
@@ -1,5 +1,5 @@
-import OpenAI from "openai";
 import Anthropic from "@anthropic-ai/sdk";
+import OpenAI from "openai";
 import { env } from "../env.js";

 export function openaiClient() {
@@ -13,6 +13,14 @@ export function xaiClient() {
  return new OpenAI({ apiKey: env.XAI_API_KEY, baseURL: "https://api.x.ai/v1" });
 }

+export function geminiClient() {
+  if (!env.GEMINI_API_KEY) throw new Error("GEMINI_API_KEY not set");
+  return {
+    apiKey: env.GEMINI_API_KEY,
+    baseURL: "https://generativelanguage.googleapis.com/v1beta",
+  };
+}
+
 export function isHermesAgentConfigured() {
  return Boolean(env.HERMES_AGENT_API_KEY);
 }
@@ -1,4 +1,4 @@
-export const PROVIDERS = ["openai", "anthropic", "xai", "hermes-agent"] as const;
+export const PROVIDERS = ["openai", "anthropic", "xai", "gemini", "hermes-agent"] as const;

 export type Provider = (typeof PROVIDERS)[number];

@@ -16,7 +16,7 @@ import { exaClient } from "./search/exa.js";
 import { isFreshSearchCacheHit, normalizeSearchQuery } from "./search-cache.js";
 import type { ChatAttachment } from "./llm/types.js";

-const ProviderSchema = z.enum(["openai", "anthropic", "xai", "hermes-agent"]);
+const ProviderSchema = z.enum(["openai", "anthropic", "xai", "gemini", "hermes-agent"]);
 const MAX_ADDITIONAL_SYSTEM_PROMPT_CHARS = 12_000;
 const EnabledToolsSchema = z.array(z.string().trim().min(1).max(80)).max(20).transform((value) => normalizeEnabledChatTools(value));

@@ -140,6 +140,94 @@ test("plain Chat Completions stream does not send Sybil-managed tools", async ()
  assert.equal(events.at(-1)?.type === "done" ? events.at(-1)?.result.text : null, "Hi");
 });

+test("Messages API defaults max_tokens to the Anthropic model maximum", async () => {
+  let requestBody: any = null;
+  let retrievedModel: string | null = null;
+  const client = {
+    models: {
+      retrieve: async (model: string) => {
+        retrievedModel = model;
+        return { id: model, max_tokens: 128000 };
+      },
+    },
+    messages: {
+      create: async (body: any) => {
+        requestBody = body;
+        return {
+          content: [{ type: "text", text: "Done" }],
+          usage: { input_tokens: 1, output_tokens: 1 },
+        };
+      },
+    },
+  };
+
+  const result = await completeWithMessagesApi({
+    client: client as any,
+    model: "claude-max-default-test",
+    messages: [{ role: "user", content: "Say done" }],
+  });
+
+  assert.equal(retrievedModel, "claude-max-default-test");
+  assert.equal(requestBody?.max_tokens, 128000);
+  assert.equal(result.text, "Done");
+});
+
+test("Messages API preserves explicit maxTokens", async () => {
+  let requestBody: any = null;
+  let didRetrieveModel = false;
+  const client = {
+    models: {
+      retrieve: async () => {
+        didRetrieveModel = true;
+        return { max_tokens: 128000 };
+      },
+    },
+    messages: {
+      create: async (body: any) => {
+        requestBody = body;
+        return streamFrom([
+          {
+            type: "message_start",
+            message: {
+              usage: { input_tokens: 1, output_tokens: 0 },
+            },
+          },
+          {
+            type: "content_block_start",
+            index: 0,
+            content_block: { type: "text", text: "" },
+          },
+          {
+            type: "content_block_delta",
+            index: 0,
+            delta: { type: "text_delta", text: "Done" },
+          },
+          { type: "content_block_stop", index: 0 },
+          {
+            type: "message_delta",
+            delta: { stop_reason: "end_turn", stop_sequence: null },
+            usage: { output_tokens: 1 },
+          },
+          { type: "message_stop" },
+        ]);
+      },
+    },
+  };
+
+  const events = await collectEvents(
+    streamWithMessagesApi({
+      client: client as any,
+      model: "claude-explicit-max-test",
+      messages: [{ role: "user", content: "Say done" }],
+      maxTokens: 4096,
+    })
+  );
+
+  assert.equal(didRetrieveModel, false);
+  assert.equal(requestBody?.max_tokens, 4096);
+  assert.equal(events.at(-1)?.type === "done" ? events.at(-1)?.result.text : null, "Done");
+});
+
 test("fetch_url sends browser-like navigation headers", async () => {
  const originalFetch = globalThis.fetch;
  const fetchCalls: Array<{ input: RequestInfo | URL; init?: RequestInit }> = [];
@@ -27,6 +27,12 @@ test("provider backend registry selects chat protocol and managed-tool mode", ()
    managedTools: true,
    enabledTools: ["web_search"],
  });
+  assert.deepEqual(describeProviderChatBackend("gemini", ["web_search"]), {
+    provider: "gemini",
+    protocol: "gemini",
+    managedTools: true,
+    enabledTools: ["web_search"],
+  });
  assert.deepEqual(describeProviderChatBackend("hermes-agent", ["web_search"]), {
    provider: "hermes-agent",
    protocol: "chat-completions",
@@ -5,8 +5,10 @@ import { fromPrismaProvider, serializeProviderFields, toPrismaProvider } from ".
 test("Hermes Agent provider id maps between API and Prisma enum forms", () => {
  assert.equal(toPrismaProvider("hermes-agent"), "hermes_agent");
  assert.equal(fromPrismaProvider("hermes_agent"), "hermes-agent");
-  assert.deepEqual(serializeProviderFields({ initiatedProvider: "hermes_agent", lastUsedProvider: "xai" }), {
+  assert.equal(toPrismaProvider("gemini"), "gemini");
+  assert.equal(fromPrismaProvider("gemini"), "gemini");
+  assert.deepEqual(serializeProviderFields({ initiatedProvider: "hermes_agent", lastUsedProvider: "gemini" }), {
    initiatedProvider: "hermes-agent",
-    lastUsedProvider: "xai",
+    lastUsedProvider: "gemini",
  });
 });
@@ -1,6 +1,6 @@
 import type { Provider } from "./types.js";

-const PROVIDERS: Provider[] = ["openai", "anthropic", "xai", "hermes-agent"];
+const PROVIDERS: Provider[] = ["openai", "anthropic", "xai", "gemini", "hermes-agent"];

 function normalizeBaseUrl(value: string) {
  const trimmed = value.trim();
@@ -42,12 +42,13 @@ type ToolLogMetadata = {
  resultPreview?: string | null;
 };

-const BASE_PROVIDERS: Provider[] = ["openai", "anthropic", "xai"];
+const BASE_PROVIDERS: Provider[] = ["openai", "anthropic", "xai", "gemini"];
 const PROVIDERS: Provider[] = [...BASE_PROVIDERS, "hermes-agent"];
 const PROVIDER_FALLBACK_MODELS: Record<Provider, string[]> = {
  openai: ["gpt-4.1-mini"],
  anthropic: ["claude-3-5-sonnet-latest"],
  xai: ["grok-3-mini"],
+  gemini: ["gemini-3.5-flash", "gemini-flash-latest"],
  "hermes-agent": ["hermes-agent"],
 };

@@ -55,6 +56,7 @@ const EMPTY_MODEL_CATALOG: ModelCatalogResponse["providers"] = {
  openai: { models: [], loadedAt: null, error: null },
  anthropic: { models: [], loadedAt: null, error: null },
  xai: { models: [], loadedAt: null, error: null },
+  gemini: { models: [], loadedAt: null, error: null },
 };

 function escapeTags(value: string) {
@@ -79,6 +81,7 @@ function getProviderLabel(provider: Provider | null | undefined) {
  if (provider === "openai") return "OpenAI";
  if (provider === "anthropic") return "Anthropic";
  if (provider === "xai") return "xAI";
+  if (provider === "gemini") return "Gemini";
  if (provider === "hermes-agent") return "Hermes Agent";
  return "";
 }
@@ -266,6 +269,7 @@ async function main() {
    openai: null,
    anthropic: null,
    xai: null,
+    gemini: null,
    "hermes-agent": null,
  };
  let model: string = config.defaultModel ?? pickProviderModel(getModelOptions(modelCatalog, provider), null);
@@ -1,4 +1,4 @@
-export type Provider = "openai" | "anthropic" | "xai" | "hermes-agent";
+export type Provider = "openai" | "anthropic" | "xai" | "gemini" | "hermes-agent";

 export type ProviderModelInfo = {
  models: string[];
@@ -123,6 +123,7 @@ const PROVIDER_FALLBACK_MODELS: Record<Provider, string[]> = {
  openai: ["gpt-4.1-mini"],
  anthropic: ["claude-3-5-sonnet-latest"],
  xai: ["grok-3-mini"],
+  gemini: ["gemini-3.5-flash", "gemini-flash-latest"],
  "hermes-agent": ["hermes-agent"],
 };

@@ -130,9 +131,10 @@ const EMPTY_MODEL_CATALOG: ModelCatalogResponse["providers"] = {
  openai: { models: [], loadedAt: null, error: null },
  anthropic: { models: [], loadedAt: null, error: null },
  xai: { models: [], loadedAt: null, error: null },
+  gemini: { models: [], loadedAt: null, error: null },
 };

-const BASE_PROVIDERS: Provider[] = ["openai", "anthropic", "xai"];
+const BASE_PROVIDERS: Provider[] = ["openai", "anthropic", "xai", "gemini"];
 const ALL_PROVIDERS: Provider[] = [...BASE_PROVIDERS, "hermes-agent"];

 const MODEL_PREFERENCES_STORAGE_KEY = "sybil:modelPreferencesByProvider";
@@ -149,6 +151,7 @@ const EMPTY_MODEL_PREFERENCES: ProviderModelPreferences = {
  openai: null,
  anthropic: null,
  xai: null,
+  gemini: null,
  "hermes-agent": null,
 };
 const EMPTY_ACTIVE_RUNS: ActiveRunsState = {
@@ -345,6 +348,7 @@ function loadStoredModelPreferences() {
      openai: typeof parsed.openai === "string" && parsed.openai.trim() ? parsed.openai.trim() : null,
      anthropic: typeof parsed.anthropic === "string" && parsed.anthropic.trim() ? parsed.anthropic.trim() : null,
      xai: typeof parsed.xai === "string" && parsed.xai.trim() ? parsed.xai.trim() : null,
+      gemini: typeof parsed.gemini === "string" && parsed.gemini.trim() ? parsed.gemini.trim() : null,
      "hermes-agent":
        typeof parsed["hermes-agent"] === "string" && parsed["hermes-agent"].trim() ? parsed["hermes-agent"].trim() : null,
    };
@@ -354,7 +358,9 @@ function loadStoredModelPreferences() {
 }

 function normalizeStoredProvider(value: unknown): Provider {
-  return value === "anthropic" || value === "xai" || value === "openai" || value === "hermes-agent" ? value : "openai";
+  return value === "anthropic" || value === "xai" || value === "gemini" || value === "openai" || value === "hermes-agent"
+    ? value
+    : "openai";
 }

 function normalizeStoredModelPreferences(value: unknown): ProviderModelPreferences {
@@ -364,6 +370,7 @@ function normalizeStoredModelPreferences(value: unknown): ProviderModelPreferenc
      openai: typeof parsed.openai === "string" && parsed.openai.trim() ? parsed.openai.trim() : null,
      anthropic: typeof parsed.anthropic === "string" && parsed.anthropic.trim() ? parsed.anthropic.trim() : null,
      xai: typeof parsed.xai === "string" && parsed.xai.trim() ? parsed.xai.trim() : null,
+      gemini: typeof parsed.gemini === "string" && parsed.gemini.trim() ? parsed.gemini.trim() : null,
      "hermes-agent":
        typeof parsed["hermes-agent"] === "string" && parsed["hermes-agent"].trim() ? parsed["hermes-agent"].trim() : null,
    };
@@ -395,6 +402,7 @@ function getProviderLabel(provider: Provider | null | undefined) {
  if (provider === "openai") return "OpenAI";
  if (provider === "anthropic") return "Anthropic";
  if (provider === "xai") return "xAI";
+  if (provider === "gemini") return "Gemini";
  if (provider === "hermes-agent") return "Hermes Agent";
  return "";
 }
@@ -286,6 +286,14 @@ textarea {
  word-break: break-word;
 }

+.md-content > :first-child {
+  margin-top: 0;
+}
+
+.md-content > :last-child {
+  margin-bottom: 0;
+}
+
 .md-table-scroll {
  max-width: 100%;
  margin: 0.35rem 0 1rem;
@@ -384,7 +392,8 @@ textarea {

 .md-content ul,
 .md-content ol {
-  margin-top: 0.65rem;
+  margin-top: 0.85rem;
+  margin-bottom: 0.85rem;
  margin-left: 0;
  padding-left: 0;
  list-style: none;
@@ -396,6 +405,26 @@ textarea {
  padding-left: 1.35rem;
 }

+.md-content ul > li {
+  position: relative;
+  padding-left: 1.1rem;
+}
+
+.md-content ul > li::before {
+  content: "";
+  position: absolute;
+  left: 0;
+  top: 0.76em;
+  width: 0.36rem;
+  height: 0.36rem;
+  border-radius: 9999px;
+  background: hsl(188 86% 62%);
+  box-shadow:
+    0 0 0 2px hsl(188 86% 62% / 0.12),
+    0 0 10px hsl(188 86% 62% / 0.42);
+  transform: translateY(-50%);
+}
+
 .md-content li + li {
  margin-top: 0.3rem;
 }
@@ -149,7 +149,7 @@ export type CompletionRequestMessage = {
  attachments?: ChatAttachment[];
 };

-export type Provider = "openai" | "anthropic" | "xai" | "hermes-agent";
+export type Provider = "openai" | "anthropic" | "xai" | "gemini" | "hermes-agent";

 export type ProviderModelInfo = {
  models: string[];
Author	SHA1	Message	Date
buzzert	622659f6ca	ios: preserve chat scroll position on resume TestFlight / testflight (push) Successful in 1m46s Details	2026-07-12 13:57:45 -07:00
buzzert	93ca8a76c3	adds gemini support TestFlight / testflight (push) Successful in 1m53s Details	2026-07-11 14:16:21 -07:00
buzzertandClaude Fable 5	69f50064a3	ci: restore Ruby PATH export for fastlane step TestFlight / testflight (push) Successful in 1m58s Details act_runner does not carry setup-ruby's PATH changes into later steps, so without this the step runs the toolcache default Ruby (3.3.11) and bundler cannot find the gems installed for 3.1.7. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-11 13:04:47 -07:00
buzzertandClaude Fable 5	e7b81f24bd	ci: put CI keychain first in codesign search list TestFlight / testflight (push) Failing after 19s Details codesign resolves signing identities through the user keychain search list (first match wins) and ignores --keychain for the lookup. This runner hosts another project whose keychain holds the same Apple Distribution identity, so when that keychain is locked, codesign fails with errSecInternalComponent no matter how correctly our own keychain is set up. Prepend the fresh CI keychain to the search list for the build and always delete it afterward, which restores the original list. Also drop the runner-diagnostics step, the GitHub-hosted Ruby PATH export, and the duplicate CODE_SIGN_IDENTITY param. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-11 13:01:42 -07:00
buzzert	20a310a4b1	ci: use Ruby 3.1 for TestFlight upload TestFlight / testflight (push) Failing after 46s Details	2026-07-11 12:01:27 -07:00
buzzert	b6859706db	ios: simplify CI signing keychain setup TestFlight / testflight (push) Failing after 25s Details	2026-07-11 11:55:04 -07:00
buzzert	ea148839d3	Revert "ios: pass match keychain to codesign" This reverts commit `a0e410155d`.	2026-07-11 11:32:00 -07:00
buzzert	a0e410155d	ios: pass match keychain to codesign TestFlight / testflight (push) Failing after 55s Details	2026-07-11 11:29:10 -07:00
buzzert	b5a5d21767	docs: document iOS release workflow	2026-07-11 11:17:05 -07:00
buzzert	0ad96f4f99	ios: center transcript content on iPad TestFlight / testflight (push) Failing after 52s Details	2026-07-11 11:15:01 -07:00
buzzert	b0c2b74c16	remove temporary keychain (unneeded) TestFlight / testflight (push) Successful in 1m46s Details	2026-07-06 09:25:24 -07:00
buzzert	14d2d8cb28	fix spacing between lists and paragraphs	2026-07-05 19:54:34 -07:00
buzzert	03836bbc4c	Raise Anthropic default max tokens	2026-07-05 11:53:17 -07:00
buzzert	76ce62025a	bullets in ul	2026-07-04 20:22:13 -07:00
buzzert	b15473d24e	ios: stamp release version into XcodeGen spec TestFlight / testflight (push) Successful in 1m43s Details	2026-06-26 01:29:54 -07:00
buzzert	a512a65844	ios: accept release tag namespace in Fastfile TestFlight / testflight (push) Successful in 1m55s Details	2026-06-26 01:18:02 -07:00
buzzert	207b44f67f	ios: ci: actually `release/ios/v*`. TestFlight / testflight (push) Failing after 19s Details	2026-06-26 01:15:08 -07:00
buzzert	c5ccd212c9	ios: ci: testflight only on release/v* TestFlight / testflight (push) Successful in 1m55s Details	2026-06-26 01:08:47 -07:00
buzzertandClaude Opus 4.8	ee990dde5d	ios: simplify Fastfile signing now that the runner has a real session With SessionCreate on the runner's launchd job, standard fastlane keychain handling works, so drop the debugging-era workarounds: the manual default-keychain / list-keychains search-list juggling, the login-keychain restoration in cleanup, the verify_ci_signing re-unlock/partition/find-identity step (match already imports the cert and sets the key partition list), and the CODE_SIGN_KEYCHAIN / OTHER_CODE_SIGN_FLAGS xcargs. CI signing is now a single create_keychain + match. No behavior change; validated end-to-end on TestFlight. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-26 00:52:36 -07:00