[mirror] fix: persist per-turn model on chats and queued messages

[yashwant86]

Mirror of upstream coder#24688 for benchmark. Do not merge.

---

Summary by MergeMonkey

• Enhancements:
  • Persist per-turn model configuration on queued messages to preserve user-selected model overrides across chat turns.
  • Add model config validation when sending messages with explicit model overrides.
  • Implement fallback resolution for queued messages with missing or invalid model configs.
• Fixes & Patches:
  • Fix stale field snapshots from watch events clobbering fresher cached metadata by scoping updates by event kind.
• Maintenance:
  • Add database migration to persist model_config_id on chat_queued_messages table.
  • Refactor chat cache merge logic into reusable helper functions for consistency.

[bot-mergemonkey]

⚡ Risk Assessment — NEEDS-TESTING · ~45 min review

Focus areas: Model config resolution fallback chain (explicit → queued → last → default) · Timestamp comparison logic for cache merge stale guards · Queued message model persistence across promotion and auto-promotion · Watch event propagation of updated model config to clients

Assessment: Adds per-turn model persistence with fallback resolution logic requiring comprehensive integration testing.

Walkthrough

User sends a chat message with an optional model override. The API validates the model config exists, then the chat daemon resolves the effective model config (explicit override → queued message config → chat's last config → default). If the chat is busy, the message queues with its model config persisted. When promoted or auto-promoted, the queued message's stored model config is used, updating the chat's last_model_config_id. Watch events propagate the updated model config to clients, which merge it into caches using event-kind-scoped logic to prevent stale snapshots from clobbering fresher metadata.

Changes

Files	Summary
Database Schema & Querying coderd/database/dump.sql coderd/database/migrations/000477_chat_queued_message_model_config.up.sql coderd/database/migrations/000477_chat_queued_message_model_config.down.sql coderd/database/queries/chats.sql coderd/database/queries.sql.go coderd/database/models.go	Add model_config_id column to chat_queued_messages table with migration that backfills from chat.last_model_config_id. Update all related query functions and models to include the new field.
Chat Daemon Model Resolution coderd/x/chatd/chatd.go	Implement model config resolution logic with three-tier fallback: explicit override → queued message config → chat's last model config → default. Validate requested model configs exist and reject invalid IDs with ErrInvalidModelConfigID.
Chat Daemon Tests coderd/x/chatd/chatd_test.go	Add comprehensive tests for model config persistence: queued message model capture, promotion with stored configs, fallback for legacy/invalid configs, and auto-promotion order preservation across multiple queued messages.
API Layer & SDK coderd/exp_chats.go codersdk/chats.go site/src/api/typesGenerated.ts	Add model_config_id field to ChatQueuedMessage SDK type and API response. Validate model config IDs in CreateChatMessage endpoint and return ErrInvalidModelConfigID for invalid overrides.
Integration Tests coderd/exp_chats_test.go	Add end-to-end tests verifying model override persistence: direct sends, queued messages, enqueue-time capture, subsequent sends reusing persisted model, and watch event propagation of updated model config.
Frontend Cache Merge Logic site/src/api/queries/chats.ts site/src/api/queries/chats.test.ts site/src/pages/AgentsPage/AgentsPage.tsx	Extract event-scoped cache merge logic into mergeWatchedChatSummary and mergeWatchedChatIntoCaches helpers. Implement precise timestamp comparison to prevent stale watch payloads from clobbering fresher cached metadata. Update AgentsPage to use new merge functions.
Database SDK Conversion coderd/database/db2sdk/db2sdk.go	Update ChatQueuedMessage conversion to include model_config_id field mapping from database to SDK type.

Sequence Diagram

sequenceDiagram
    participant Client
    participant API
    participant ChatDaemon
    participant Database
    Client->>API: CreateChatMessage(modelConfigID?)
    API->>ChatDaemon: SendMessage(modelConfigID?)
    ChatDaemon->>Database: GetChatModelConfigByID(requested)
    alt Explicit Override
        Database-->>ChatDaemon: config found
        ChatDaemon->>ChatDaemon: use requested config
    else No Override
        ChatDaemon->>ChatDaemon: use chat.LastModelConfigID
    end
    alt Chat Not Busy
        ChatDaemon->>Database: InsertChatMessage(modelConfigID)
        Database-->>ChatDaemon: message
        ChatDaemon-->>API: direct send result
    else Chat Busy
        ChatDaemon->>Database: InsertChatQueuedMessage(modelConfigID)
        Database-->>ChatDaemon: queued message
        ChatDaemon-->>API: queued result
    end
    API-->>Client: response with model config
    alt Queued Message Promotion
        Client->>API: PromoteQueuedMessage(queuedMessageID)
        API->>ChatDaemon: PromoteQueued(queuedMessageID)
        ChatDaemon->>Database: GetChatQueuedMessage(queuedMessageID)
        Database-->>ChatDaemon: queued message with stored modelConfigID
        ChatDaemon->>ChatDaemon: resolveQueuedMessageModelConfigID
        ChatDaemon->>Database: InsertChatMessage(resolvedModelConfigID)
        ChatDaemon->>Database: UpdateChatLastModelConfig(resolvedModelConfigID)
        ChatDaemon->>Database: PublishChatWatchEvent(statusChange)
        Database-->>ChatDaemon: event published
        ChatDaemon-->>API: promoted result
        API-->>Client: response
        Client->>Client: mergeWatchedChatIntoCaches(eventKind)
    end

Loading

Dig Deeper With Commands

• /review <file-path> <function-optional>
• /chat <file-path> "<question>"
• /roast <file-path>

Runs only when explicitly triggered.