refactor(agent): remove agent gateway instead of twilight sdk (#264)

* refactor(agent): replace TypeScript agent gateway with in-process Go agent using twilight-ai SDK

- Remove apps/agent (Bun/Elysia gateway), packages/agent (@memoh/agent),
  internal/bun runtime manager, and all embedded agent/bun assets
- Add internal/agent package powered by twilight-ai SDK for LLM calls,
  tool execution, streaming, sential logic, tag extraction, and prompts
- Integrate ToolGatewayService in-process for both built-in and user MCP
  tools, eliminating HTTP round-trips to the old gateway
- Update resolver to convert between sdk.Message and ModelMessage at the
  boundary (resolver_messages.go), keeping agent package free of
  persistence concerns
- Prepend user message before storeRound since SDK only returns output
  messages (assistant + tool)
- Clean up all Docker configs, TOML configs, nginx proxy, Dockerfile.agent,
  and Go config structs related to the removed agent gateway
- Update cmd/agent and cmd/memoh entry points with setter-based
  ToolGateway injection to avoid FX dependency cycles

* fix(web): move form declaration before computed properties that reference it

The `form` reactive object was declared after computed properties like
`selectedMemoryProvider` and `isSelectedMemoryProviderPersisted` that
reference it, causing a TDZ ReferenceError during setup.

* fix: prevent UTF-8 character corruption in streaming text output

StreamTagExtractor.Push() used byte-level string slicing to hold back
buffer tails for tag detection, which could split multi-byte UTF-8
characters. After json.Marshal replaced invalid bytes with U+FFFD,
the corruption became permanent — causing garbled CJK characters (�)
in agent responses.

Add safeUTF8SplitIndex() to back up split points to valid character
boundaries. Also fix byte-level truncation in command/formatter.go
and command/fs.go to use rune-aware slicing.

* fix: add agent error logging and fix Gemini tool schema validation

- Log agent stream errors in both SSE and WebSocket paths with bot/model context
- Fix send tool `attachments` parameter: empty `items` schema rejected by
  Google Gemini API (INVALID_ARGUMENT), now specifies `{"type": "string"}`
- Upgrade twilight-ai to d898f0b (includes raw body in API error messages)

* chore(ci): remove agent gateway from Docker build and release pipelines

Agent gateway has been replaced by in-process Go agent; remove the
obsolete Docker image matrix entry, Bun/UPX CI steps, and agent-binary
build logic from the release script.

* fix: preserve attachment filename, metadata, and container path through persistence

- Add `name` column to `bot_history_message_assets` (migration 0034) to
  persist original filenames across page refreshes.
- Add `metadata` JSONB column (migration 0035) to store source_path,
  source_url, and other context alongside each asset.
- Update SQL queries, sqlc-generated code, and all Go types (MessageAsset,
  AssetRef, OutboundAssetRef, FileAttachment) to carry name and metadata
  through the full lifecycle.
- Extract filenames from path/URL in AttachmentsResolver before clearing
  raw paths; enrich streaming event metadata with name, source_path, and
  source_url in both the WebSocket and channel inbound ingestion paths.
- Implement `LinkAssets` on message service and `LinkOutboundAssets` on
  flow resolver so WebSocket-streamed bot attachments are persisted to the
  correct assistant message after streaming completes.
- Frontend: update MessageAsset type with metadata field, pass metadata
  through to attachment items, and reorder attachment-block.vue template
  so container files (identified by metadata.source_path) open in the
  sidebar file manager instead of triggering a download.

* refactor(agent): decouple built-in tools from MCP, load via ToolProvider interface

Migrate all 13 built-in tool providers from internal/mcp/providers/ to
internal/agent/tools/ using the twilight-ai sdk.Tool structure. The agent
now loads tools through a ToolProvider interface instead of the MCP
ToolGatewayService, which is simplified to only manage external federation
sources. This enables selective tool loading and removes the coupling
between business tools and the MCP protocol layer.

* refactor(flow): split monolithic resolver.go into focused modules

Break the 1959-line resolver.go into 12 files organized by concern:
- resolver.go: core orchestration (Resolver struct, resolve, Chat, prepareRunConfig)
- resolver_stream.go: streaming (StreamChat, StreamChatWS, tryStoreStream)
- resolver_trigger.go: schedule/heartbeat triggers
- resolver_attachments.go: attachment routing, inlining, encoding
- resolver_history.go: message loading, deduplication, token trimming
- resolver_store.go: persistence (storeRound, storeMessages, asset linking)
- resolver_memory.go: memory provider integration
- resolver_model_selection.go: model selection and candidate matching
- resolver_identity.go: display name and channel identity resolution
- resolver_settings.go: bot settings, loop detection, inbox
- user_header.go: YAML front-matter formatting
- resolver_util.go: shared utilities (sanitize, normalize, dedup, UUID)

* fix(agent): enable Anthropic extended thinking by passing ReasoningConfig to provider

Anthropic's thinking requires WithThinking() at provider creation time,
unlike OpenAI which uses per-request ReasoningEffort. The config was
never wired through, so Claude models could not trigger thinking.

* refactor(agent): extract prompts into embedded markdown templates

Move inline prompt strings from prompt.go into separate .md files under
internal/agent/prompts/, using {{key}} placeholders and a simple render
engine. Remove obsolete SystemPromptParams fields (Language,
MaxContextLoadTime, Channels, CurrentChannel) and their call-site usage.

* fix: lint

.agents/skills/twilight-ai/SKILL.md

@@ -22,6 +22,7 @@ Twilight AI is a lightweight Go AI SDK with a provider-agnostic core API.
 - Text generation: `sdk.GenerateText`, `sdk.GenerateTextResult`, `sdk.StreamText`
 - Embeddings: `sdk.Embed`, `sdk.EmbedMany`
 - Tool calling: `sdk.Tool`, `sdk.NewTool[T]`, `WithMaxSteps`, approval flow
+- MCP tool integration: `sdk.CreateMCPClient`, `sdk.MCPClient`, `sdk.MCPClientConfig`
 - Streaming: typed `StreamPart` events over Go channels
 - Current providers:
   - `provider/openai/completions`
@@ -38,6 +39,7 @@ Prefer the high-level SDK API first, then drop to provider details only when nee
 - `sdk.Model` binds a chat model to a `sdk.Provider`
 - `sdk.EmbeddingModel` binds an embedding model to an `sdk.EmbeddingProvider`
 - The client orchestrates tool loops, callbacks, approvals, and streaming lifecycle
+- MCP clients can load remote MCP tools and turn them into ordinary `sdk.Tool` values
 - Providers handle backend-specific HTTP, request mapping, response parsing, and SSE translation
 
 ## Core API Guidance
@@ -114,6 +116,26 @@ When streaming with tools, ensure the implementation can emit:
 - progress updates
 - denial/error events when applicable
 
+### MCP Tool Calling
+
+Use MCP when the task needs remote tools exposed by an MCP server rather than locally implemented `Execute` handlers.
+
+Default guidance:
+
+- use `sdk.CreateMCPClient(ctx, &sdk.MCPClientConfig{...})`
+- use `sdk.MCPTransportHTTP` for streamable HTTP MCP servers
+- use `sdk.MCPTransportSSE` only when the server exposes legacy SSE transport
+- for stdio, build the transport with the official MCP Go SDK and pass `Transport: ...`
+- call `mcpClient.Tools(ctx)` and pass the result into `sdk.WithTools(...)`
+- call `defer mcpClient.Close()` after successful creation
+
+Important behavior:
+
+- MCP tools become ordinary `sdk.Tool` values from the caller's perspective
+- Twilight AI converts MCP `InputSchema` into `*jsonschema.Schema`
+- MCP tool execution is delegated to `tools/call` on the remote server
+- remote MCP text output becomes the tool result visible to the model
+
 ### Streaming
 
 Twilight AI streaming is channel-first and type-safe. Prefer type switches over loosely typed event parsing.
@@ -198,6 +220,7 @@ Before finishing work in this repo, verify:
 - public examples use top-level `sdk` APIs unless lower-level behavior is the point
 - streaming logic uses typed `StreamPart` handling
 - tool-calling changes cover both inspection mode and multi-step mode when relevant
+- MCP examples show both transport setup and normal `WithTools(...)` usage when relevant
 - provider work includes health checks or model discovery behavior if the backend supports them
 
 ## Additional Resources

.agents/skills/twilight-ai/reference.md

@@ -309,6 +309,42 @@ type ToolResult struct {
 }
 ```
 
+### MCP
+
+```go
+type MCPTransportType string
+
+const (
+    MCPTransportHTTP MCPTransportType = "http"
+    MCPTransportSSE  MCPTransportType = "sse"
+)
+
+type MCPClientConfig struct {
+    Type       MCPTransportType
+    URL        string
+    Headers    map[string]string
+    Transport  mcp.Transport
+    HTTPClient *http.Client
+    Name       string
+    Version    string
+}
+
+type MCPClient struct { /* unexported fields */ }
+
+func CreateMCPClient(ctx context.Context, config *MCPClientConfig) (*MCPClient, error)
+func (c *MCPClient) Tools(ctx context.Context) ([]Tool, error)
+func (c *MCPClient) Close() error
+```
+
+Usage notes:
+
+- `MCPTransportHTTP` is the default built-in transport and uses the official MCP Go SDK's streamable HTTP client transport.
+- `MCPTransportSSE` uses the official MCP Go SDK's SSE client transport.
+- For stdio or other custom transports, create the transport with `github.com/modelcontextprotocol/go-sdk/mcp` and pass it through `Transport`.
+- `Tools(ctx)` converts remote MCP tools into ordinary `sdk.Tool` values suitable for `WithTools(...)`.
+- MCP tool schemas are converted from MCP `InputSchema` into `*jsonschema.Schema`.
+- MCP execution wrappers call `tools/call` and return concatenated text content to the model.
+
 ### Streaming
 
 ```go