@ai-agent-forge/agent-forge 0.85.0

package/docs/rpc.md

@@ -0,0 +1,1487 @@
+# RPC Mode
+
+RPC mode enables headless operation of the coding agent via a JSON protocol over stdin/stdout. This is useful for embedding the agent in other applications, IDEs, or custom UIs.
+
+**Note for Node.js/TypeScript users**: If you're building a Node.js application, consider using the stable SDK entry `createStableAgentSession()` from `@agent-forge/agent-forge` instead of spawning a subprocess. It returns the stable facade pair `{ facade, sessionAPI }` over `CodingAgentFacadeV1`, with no raw `AgentSession` on the surface; see [SDK documentation](sdk.md) for the API. For a subprocess-based TypeScript client, see [`src/modes/rpc/rpc-client.ts`](../src/modes/rpc/rpc-client.ts).
+
+## Starting RPC Mode
+
+```bash
+agent-forge --mode rpc [options]
+```
+
+Common options:
+- `--provider <name>`: Set the LLM provider (anthropic, openai, google, etc.)
+- `--model <pattern>`: Model pattern or ID (supports `provider/id` and optional `:<thinking>`)
+- `--name <name>` / `-n <name>`: Set the session display name at startup
+- `--no-session`: Disable session persistence
+- `--session-dir <path>`: Custom session storage directory
+
+## Protocol Overview
+
+- **Commands**: JSON objects sent to stdin, one per line
+- **Responses**: JSON objects with `type: "response"` indicating command success/failure
+- **Events**: Agent events streamed to stdout as JSON lines
+
+All commands support an optional `id` field for request/response correlation. If provided, the corresponding response will include the same `id`. `bash_execution_update` events also include the `id` of their originating `bash` command.
+
+### Framing
+
+RPC mode uses strict JSONL semantics with LF (`\n`) as the only record delimiter.
+
+This matters for clients:
+- Split records on `\n` only
+- Accept optional `\r\n` input by stripping a trailing `\r`
+- Do not use generic line readers that treat Unicode separators as newlines
+
+In particular, Node `readline` is not protocol-compliant for RPC mode because it also splits on `U+2028` and `U+2029`, which are valid inside JSON strings.
+
+## Commands
+
+### Prompting
+
+#### prompt
+
+Send a user prompt to the agent. The command response is emitted after the prompt is accepted, queued, or handled. Events continue streaming asynchronously after acceptance.
+
+```json
+{"id": "req-1", "type": "prompt", "message": "Hello, world!"}
+```
+
+With images:
+```json
+{"type": "prompt", "message": "What's in this image?", "images": [{"type": "image", "data": "base64-encoded-data", "mimeType": "image/png"}]}
+```
+
+**During streaming**: If the agent is already streaming, you must specify `streamingBehavior` to queue the message:
+
+```json
+{"type": "prompt", "message": "New instruction", "streamingBehavior": "steer"}
+```
+
+- `"steer"`: Queue the message while the agent is running. It is delivered after the current assistant turn finishes executing its tool calls, before the next LLM call.
+- `"followUp"`: Wait until the agent finishes. Message is delivered only when agent stops.
+
+If the agent is streaming and no `streamingBehavior` is specified, the command returns an error.
+
+**Plugin commands**: If the message is a plugin command (e.g., `/mycommand` from a capability plugin), it executes immediately even during streaming. Plugin commands manage their own execution and user interaction.
+
+**Input expansion**: Skill commands (`/skill:name`) and prompt templates (`/template`) are expanded before sending/queueing.
+
+Response:
+```json
+{"id": "req-1", "type": "response", "command": "prompt", "success": true}
+```
+
+`success: true` means the prompt was accepted, queued, or handled immediately. `success: false` means the prompt was rejected before acceptance. Failures after acceptance are reported through the normal event and message stream, not as a second `response` for the same request id.
+
+The `images` field is optional. Each image uses `ImageContent` format: `{"type": "image", "data": "base64-encoded-data", "mimeType": "image/png"}`.
+
+#### steer
+
+Queue a steering message while the agent is running. It is delivered after the current assistant turn finishes executing its tool calls, before the next LLM call. Skill commands and prompt templates are expanded. Plugin commands are not allowed (use `prompt` instead).
+
+```json
+{"type": "steer", "message": "Stop and do this instead"}
+```
+
+With images:
+```json
+{"type": "steer", "message": "Look at this instead", "images": [{"type": "image", "data": "base64-encoded-data", "mimeType": "image/png"}]}
+```
+
+The `images` field is optional. Each image uses `ImageContent` format (same as `prompt`).
+
+Response:
+```json
+{"type": "response", "command": "steer", "success": true}
+```
+
+See [set_steering_mode](#set_steering_mode) for controlling how steering messages are processed.
+
+#### follow_up
+
+Queue a follow-up message to be processed after the agent finishes. Delivered only when agent has no more tool calls or steering messages. Skill commands and prompt templates are expanded. Plugin commands are not allowed (use `prompt` instead).
+
+```json
+{"type": "follow_up", "message": "After you're done, also do this"}
+```
+
+With images:
+```json
+{"type": "follow_up", "message": "Also check this image", "images": [{"type": "image", "data": "base64-encoded-data", "mimeType": "image/png"}]}
+```
+
+The `images` field is optional. Each image uses `ImageContent` format (same as `prompt`).
+
+Response:
+```json
+{"type": "response", "command": "follow_up", "success": true}
+```
+
+See [set_follow_up_mode](#set_follow_up_mode) for controlling how follow-up messages are processed.
+
+#### abort
+
+Abort the current agent operation.
+
+```json
+{"type": "abort"}
+```
+
+Response:
+```json
+{"type": "response", "command": "abort", "success": true}
+```
+
+#### clear_queue
+
+Remove queued steering and follow-up messages and return their text.
+
+```json
+{"type": "clear_queue"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "clear_queue",
+  "success": true,
+  "data": {
+    "steering": ["Change direction"],
+    "followUp": ["Summarize when finished"]
+  }
+}
+```
+
+To implement interactive Esc behavior, send `clear_queue` before `abort`, then restore the returned text in the client editor. `abort` continues queued messages when they remain in the session.
+
+#### new_session
+
+Start a fresh session.
+
+```json
+{"type": "new_session"}
+```
+
+With optional parent session tracking:
+```json
+{"type": "new_session", "parentSession": "/path/to/parent-session.jsonl"}
+```
+
+Response:
+```json
+{"type": "response", "command": "new_session", "success": true, "data": {"cancelled": false}}
+```
+
+### State
+
+#### get_state
+
+Get current session state.
+
+```json
+{"type": "get_state"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "get_state",
+  "success": true,
+  "data": {
+    "model": {...},
+    "thinkingLevel": "medium",
+    "isStreaming": false,
+    "isCompacting": false,
+    "steeringMode": "all",
+    "followUpMode": "one-at-a-time",
+    "sessionFile": "/path/to/session.jsonl",
+    "sessionId": "abc123",
+    "sessionName": "my-feature-work",
+    "autoCompactionEnabled": true,
+    "messageCount": 5,
+    "pendingMessageCount": 0
+  }
+}
+```
+
+The `model` field is a full [Model](#model) object or `null`. The `sessionName` field is the display name set via `set_session_name`, or omitted if not set.
+
+#### get_messages
+
+Get all messages in the conversation.
+
+```json
+{"type": "get_messages"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "get_messages",
+  "success": true,
+  "data": {"messages": [...]}
+}
+```
+
+Messages are `AgentMessage` objects (see [Message Types](#message-types)).
+
+### Model
+
+#### set_model
+
+Switch to a specific model.
+
+```json
+{"type": "set_model", "provider": "anthropic", "modelId": "claude-sonnet-4-20250514"}
+```
+
+Response contains the full [Model](#model) object:
+```json
+{
+  "type": "response",
+  "command": "set_model",
+  "success": true,
+  "data": {...}
+}
+```
+
+#### cycle_model
+
+Cycle to the next available model. Returns `null` data if only one model available.
+
+```json
+{"type": "cycle_model"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "cycle_model",
+  "success": true,
+  "data": {
+    "model": {...},
+    "thinkingLevel": "medium",
+    "isScoped": false
+  }
+}
+```
+
+The `model` field is a full [Model](#model) object.
+
+#### get_available_models
+
+List all configured models.
+
+```json
+{"type": "get_available_models"}
+```
+
+Response contains an array of full [Model](#model) objects:
+```json
+{
+  "type": "response",
+  "command": "get_available_models",
+  "success": true,
+  "data": {
+    "models": [...]
+  }
+}
+```
+
+### Thinking
+
+#### set_thinking_level
+
+Set the reasoning/thinking level for models that support it.
+
+```json
+{"type": "set_thinking_level", "level": "high"}
+```
+
+Levels: `"off"`, `"minimal"`, `"low"`, `"medium"`, `"high"`, `"xhigh"`, `"max"`
+
+`"xhigh"` and `"max"` are exposed only when supported by the selected model. Some models, including GPT-5.6, expose both.
+
+Response:
+```json
+{"type": "response", "command": "set_thinking_level", "success": true}
+```
+
+#### cycle_thinking_level
+
+Cycle through available thinking levels. Returns `null` data if model doesn't support thinking.
+
+```json
+{"type": "cycle_thinking_level"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "cycle_thinking_level",
+  "success": true,
+  "data": {"level": "high"}
+}
+```
+
+#### get_available_thinking_levels
+
+List the thinking levels supported by the current model. Returns `["off"]` for a model without reasoning support.
+
+```json
+{"type": "get_available_thinking_levels"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "get_available_thinking_levels",
+  "success": true,
+  "data": {
+    "levels": ["off", "minimal", "low", "medium", "high"]
+  }
+}
+```
+
+### Queue Modes
+
+#### set_steering_mode
+
+Control how steering messages (from `steer`) are delivered.
+
+```json
+{"type": "set_steering_mode", "mode": "one-at-a-time"}
+```
+
+Modes:
+- `"all"`: Deliver all steering messages after the current assistant turn finishes executing its tool calls
+- `"one-at-a-time"`: Deliver one steering message per completed assistant turn (default)
+
+Response:
+```json
+{"type": "response", "command": "set_steering_mode", "success": true}
+```
+
+#### set_follow_up_mode
+
+Control how follow-up messages (from `follow_up`) are delivered.
+
+```json
+{"type": "set_follow_up_mode", "mode": "one-at-a-time"}
+```
+
+Modes:
+- `"all"`: Deliver all follow-up messages when agent finishes
+- `"one-at-a-time"`: Deliver one follow-up message per agent completion (default)
+
+Response:
+```json
+{"type": "response", "command": "set_follow_up_mode", "success": true}
+```
+
+### Compaction
+
+#### compact
+
+Manually compact conversation context to reduce token usage.
+
+```json
+{"type": "compact"}
+```
+
+With custom instructions:
+```json
+{"type": "compact", "customInstructions": "Focus on code changes"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "compact",
+  "success": true,
+  "data": {
+    "summary": "Summary of conversation...",
+    "firstKeptEntryId": "abc123",
+    "tokensBefore": 150000,
+    "estimatedTokensAfter": 32000,
+    "usage": {
+      "input": 32000,
+      "output": 1200,
+      "cacheRead": 0,
+      "cacheWrite": 0,
+      "totalTokens": 33200,
+      "cost": {"input": 0.01, "output": 0.02, "cacheRead": 0, "cacheWrite": 0, "total": 0.03}
+    },
+    "details": {}
+  }
+}
+```
+
+`estimatedTokensAfter` is a heuristic estimate over the runtime message context immediately after compaction, including any `ContextStrategy` replacement projection; it is not a provider-exact token count. `usage` reports the LLM call or calls that generated the summary and may be omitted by custom compaction handlers.
+
+#### set_auto_compaction
+
+Enable or disable automatic compaction when context is nearly full.
+
+```json
+{"type": "set_auto_compaction", "enabled": true}
+```
+
+Response:
+```json
+{"type": "response", "command": "set_auto_compaction", "success": true}
+```
+
+### Retry
+
+#### set_auto_retry
+
+Enable or disable automatic retry on transient errors (overloaded, rate limit, 5xx).
+
+```json
+{"type": "set_auto_retry", "enabled": true}
+```
+
+Response:
+```json
+{"type": "response", "command": "set_auto_retry", "success": true}
+```
+
+#### abort_retry
+
+Abort an in-progress retry (cancel the delay and stop retrying).
+
+```json
+{"type": "abort_retry"}
+```
+
+Response:
+```json
+{"type": "response", "command": "abort_retry", "success": true}
+```
+
+### Bash
+
+#### bash
+
+Execute a shell command and add output to conversation context. Output streams as `bash_execution_update` events while the command runs; the response contains the final result.
+
+```json
+{"id": "req-1", "type": "bash", "command": "ls -la"}
+```
+
+Include an `id` to associate streamed `bash_execution_update` events with this command.
+
+Response:
+```json
+{
+  "id": "req-1",
+  "type": "response",
+  "command": "bash",
+  "success": true,
+  "data": {
+    "output": "total 48\ndrwxr-xr-x ...",
+    "exitCode": 0,
+    "cancelled": false,
+    "truncated": false
+  }
+}
+```
+
+If output was truncated, includes `fullOutputPath`:
+```json
+{
+  "type": "response",
+  "command": "bash",
+  "success": true,
+  "data": {
+    "output": "truncated output...",
+    "exitCode": 0,
+    "cancelled": false,
+    "truncated": true,
+    "fullOutputPath": "/tmp/pi-bash-abc123.log"
+  }
+}
+```
+
+**How bash results reach the LLM:**
+
+The `bash` command executes immediately and returns a `BashResult`. Internally, a `BashExecutionMessage` is created and stored in the agent's message state.
+
+When the next `prompt` command is sent, all messages (including `BashExecutionMessage`) are transformed before being sent to the LLM. The `BashExecutionMessage` is converted to a `UserMessage` with this format:
+
+````
+Ran `ls -la`
+```
+total 48
+drwxr-xr-x ...
+```
+````
+
+This means:
+1. Bash output is included in the LLM context on the **next prompt**, not immediately
+2. Multiple bash commands can be executed before a prompt; all outputs will be included
+
+#### abort_bash
+
+Abort a running bash command.
+
+```json
+{"type": "abort_bash"}
+```
+
+Response:
+```json
+{"type": "response", "command": "abort_bash", "success": true}
+```
+
+### Session
+
+#### get_session_stats
+
+Get token usage, cost statistics, and current context window usage.
+
+```json
+{"type": "get_session_stats"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "get_session_stats",
+  "success": true,
+  "data": {
+    "sessionFile": "/path/to/session.jsonl",
+    "sessionId": "abc123",
+    "userMessages": 5,
+    "assistantMessages": 5,
+    "toolCalls": 12,
+    "toolResults": 12,
+    "totalMessages": 22,
+    "tokens": {
+      "input": 50000,
+      "output": 10000,
+      "cacheRead": 40000,
+      "cacheWrite": 5000,
+      "total": 105000
+    },
+    "cost": 0.45,
+    "contextUsage": {
+      "tokens": 60000,
+      "contextWindow": 200000,
+      "percent": 30
+    }
+  }
+}
+```
+
+`tokens` and `cost` include assistant messages, usage reported by tools, and compaction/branch-summary generation across the full session. `contextUsage` contains the actual current context-window estimate used for compaction and footer display.
+
+`contextUsage` is omitted when no model or context window is available. `contextUsage.tokens` and `contextUsage.percent` are `null` immediately after compaction until a fresh post-compaction assistant response provides valid usage data.
+
+#### export_html
+
+Export session to an HTML file.
+
+```json
+{"type": "export_html"}
+```
+
+With custom path:
+```json
+{"type": "export_html", "outputPath": "/tmp/session.html"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "export_html",
+  "success": true,
+  "data": {"path": "/tmp/session.html"}
+}
+```
+
+#### switch_session
+
+Load a different session file.
+
+```json
+{"type": "switch_session", "sessionPath": "/path/to/session.jsonl"}
+```
+
+Response:
+```json
+{"type": "response", "command": "switch_session", "success": true, "data": {"cancelled": false}}
+```
+
+#### fork
+
+Create a new fork from a previous user message on the active branch. Returns the text of the message being forked from.
+
+```json
+{"type": "fork", "entryId": "abc123"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "fork",
+  "success": true,
+  "data": {"text": "The original prompt text...", "cancelled": false}
+}
+```
+
+#### clone
+
+Duplicate the current active branch into a new session at the current position.
+
+```json
+{"type": "clone"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "clone",
+  "success": true,
+  "data": {"cancelled": false}
+}
+```
+
+#### get_fork_messages
+
+Get user messages available for forking.
+
+```json
+{"type": "get_fork_messages"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "get_fork_messages",
+  "success": true,
+  "data": {
+    "messages": [
+      {"entryId": "abc123", "text": "First prompt..."},
+      {"entryId": "def456", "text": "Second prompt..."}
+    ]
+  }
+}
+```
+
+#### get_entries
+
+Get all session entries in append order (excluding the session header). The session is an append-only tree of entries with stable ids, so an entry id works as a durable cursor: pass the last entry id you have seen as `since` to get only entries strictly after it, even across client restarts. Unlike `get_messages`, this includes pre-compaction history and abandoned branches.
+
+```json
+{"type": "get_entries"}
+```
+
+With a cursor:
+```json
+{"type": "get_entries", "since": "abc123"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "get_entries",
+  "success": true,
+  "data": {
+    "entries": [
+      {"type": "message", "id": "def456", "parentId": "abc123", "timestamp": "...", "message": {"role": "user", "...": "..."}}
+    ],
+    "leafId": "def456"
+  }
+}
+```
+
+`leafId` is the id of the current leaf entry (`null` for an empty session), so a client can tell in one round trip whether the active branch moved. If `since` does not match any entry id, the response is `success: false`.
+
+#### get_tree
+
+Get the session as a tree of entries. Each node is `{entry, children, label?, labelTimestamp?}`. A well-formed session has a single root; orphaned entries (broken parent chain) also appear as roots.
+
+```json
+{"type": "get_tree"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "get_tree",
+  "success": true,
+  "data": {
+    "tree": [
+      {
+        "entry": {"type": "message", "id": "abc123", "parentId": null, "...": "..."},
+        "children": [
+          {"entry": {"type": "message", "id": "def456", "parentId": "abc123", "...": "..."}, "children": []}
+        ]
+      }
+    ],
+    "leafId": "def456"
+  }
+}
+```
+
+#### get_last_assistant_text
+
+Get the text content of the last assistant message.
+
+```json
+{"type": "get_last_assistant_text"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "get_last_assistant_text",
+  "success": true,
+  "data": {"text": "The assistant's response..."}
+}
+```
+
+Returns `{"text": null}` if no assistant messages exist.
+
+#### set_session_name
+
+Set a display name for the current session. The name appears in session listings and helps identify sessions.
+
+```json
+{"type": "set_session_name", "name": "my-feature-work"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "set_session_name",
+  "success": true
+}
+```
+
+The current session name is available via `get_state` in the `sessionName` field. To set the initial name when starting RPC mode, pass `--name <name>` or `-n <name>` to the `agent-forge --mode rpc` process.
+
+### Commands
+
+#### get_commands
+
+Get available commands (plugin commands, prompt templates, and skills). These can be invoked via the `prompt` command by prefixing with `/`.
+
+```json
+{"type": "get_commands"}
+```
+
+Response:
+```json
+{
+  "type": "response",
+  "command": "get_commands",
+  "success": true,
+  "data": {
+    "commands": [
+      {"name": "session-name", "description": "Set or clear session name", "source": "capability", "sourceInfo": {"path": "<plugin:my-plugin:session-name>", "source": "my-plugin", "scope": "temporary", "origin": "top-level"}},
+      {"name": "fix-tests", "description": "Fix failing tests", "source": "prompt", "sourceInfo": {"path": "/home/user/myproject/.agent-forge/agent/prompts/fix-tests.md", "source": "local", "scope": "project", "origin": "top-level", "baseDir": "/home/user/myproject/.agent-forge/agent/prompts"}},
+      {"name": "skill:brave-search", "description": "Web search via Brave API", "source": "skill", "sourceInfo": {"path": "/home/user/.agent-forge/agent/skills/brave-search/SKILL.md", "source": "local", "scope": "user", "origin": "top-level", "baseDir": "/home/user/.agent-forge/agent/skills"}}
+    ]
+  }
+}
+```
+
+Each command has:
+- `name`: Command name (invoke with `/name`)
+- `description`: Human-readable description (optional for plugin commands)
+- `source`: What kind of command:
+  - `"capability"`: Registered by a capability plugin
+  - `"prompt"`: Loaded from a prompt template `.md` file
+  - `"skill"`: Loaded from a skill directory (name is prefixed with `skill:`)
+- `sourceInfo`: Where the command was loaded from:
+  - `path`: File path; capability plugins report a synthetic `<plugin:<plugin>:<command>>` label
+  - `source`: The providing source (`local` for directory discovery, the plugin id for plugin commands, or a package source string)
+  - `scope`: `"user"`, `"project"`, or `"temporary"`
+  - `origin`: `"top-level"` or `"package"`
+
+**Note**: Built-in TUI commands (`/settings`, `/hotkeys`, etc.) are not included. They are handled only in interactive mode and would not execute if sent via `prompt`.
+
+## Events
+
+Events are streamed to stdout as JSON lines during agent operation. Events do not generally include an `id` field; `bash_execution_update` includes the `id` of its originating `bash` command when one was provided.
+
+### Event Types
+
+| Event | Description |
+|-------|-------------|
+| `agent_start` | Agent begins processing |
+| `agent_end` | One low-level agent run completes (may still be followed by retry, compaction, or queued continuations) |
+| `agent_settled` | Agent run is fully settled; no automatic retry, compaction retry, or queued continuation remains |
+| `turn_start` | New turn begins |
+| `turn_end` | Turn completes (includes assistant message and tool results) |
+| `message_start` | Message begins |
+| `message_update` | Streaming update (text/thinking/toolcall deltas) |
+| `message_end` | Message completes |
+| `bash_execution_update` | Direct RPC bash command output chunk |
+| `tool_execution_start` | Tool begins execution |
+| `tool_execution_update` | Tool execution progress (streaming output) |
+| `tool_execution_end` | Tool completes |
+| `queue_update` | Pending steering/follow-up queue changed |
+| `compaction_start` | Compaction begins |
+| `compaction_end` | Compaction completes |
+| `auto_retry_start` | Auto-retry begins (after transient error) |
+| `auto_retry_end` | Auto-retry completes (success or final failure) |
+| `summarization_retry_scheduled` | Retry scheduled for a transient compaction or branch-summary summarization error |
+| `summarization_retry_attempt_start` | Retried summarization request starts |
+| `summarization_retry_finished` | Summarization retry loop completes |
+
+### agent_start
+
+Emitted when the agent begins processing a prompt.
+
+```json
+{"type": "agent_start"}
+```
+
+### agent_end
+
+Emitted when one low-level agent run completes. Contains all messages generated during this run. If `willRetry` is true, an automatic retry will follow.
+
+```json
+{
+  "type": "agent_end",
+  "messages": [...],
+  "willRetry": false
+}
+```
+
+### agent_settled
+
+Emitted after the full session-level run settles. At this point Pi will not continue automatically through retry, compaction retry, or queued follow-up messages.
+
+```json
+{"type": "agent_settled"}
+```
+
+### turn_start / turn_end
+
+A turn consists of one assistant response plus any resulting tool calls and results.
+
+```json
+{"type": "turn_start"}
+```
+
+```json
+{
+  "type": "turn_end",
+  "message": {...},
+  "toolResults": [...]
+}
+```
+
+### message_start / message_end
+
+Emitted when a message begins and completes. The `message` field contains an `AgentMessage`.
+
+```json
+{"type": "message_start", "message": {...}}
+{"type": "message_end", "message": {...}}
+```
+
+### message_update (Streaming)
+
+Emitted during streaming of assistant messages. Contains a delta event without a cumulative message snapshot.
+
+```json
+{
+  "type": "message_update",
+  "usage": {
+    "input": 100,
+    "output": 1,
+    "cacheRead": 0,
+    "cacheWrite": 0,
+    "totalTokens": 101,
+    "cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0, "total": 0}
+  },
+  "assistantMessageEvent": {
+    "type": "text_delta",
+    "contentIndex": 0,
+    "delta": "Hello "
+  }
+}
+```
+
+The `assistantMessageEvent` field contains one of these delta types:
+
+| Type | Description |
+|------|-------------|
+| `text_start` | Text content block started |
+| `text_delta` | Text content chunk |
+| `text_end` | Text content block ended |
+| `thinking_start` | Thinking block started |
+| `thinking_delta` | Thinking content chunk |
+| `thinking_end` | Thinking block ended |
+| `toolcall_start` | Tool call started (includes `id` and `toolName`) |
+| `toolcall_delta` | Tool call arguments chunk |
+| `toolcall_end` | Tool call ended (includes full `toolCall` object) |
+
+Example streaming a text response:
+```json
+{"type":"message_update","usage":{...},"assistantMessageEvent":{"type":"text_start","contentIndex":0}}
+{"type":"message_update","usage":{...},"assistantMessageEvent":{"type":"text_delta","contentIndex":0,"delta":"Hello"}}
+{"type":"message_update","usage":{...},"assistantMessageEvent":{"type":"text_delta","contentIndex":0,"delta":" world"}}
+{"type":"message_update","usage":{...},"assistantMessageEvent":{"type":"text_end","contentIndex":0,"content":"Hello world"}}
+```
+
+The top-level `usage` field contains the latest cumulative provider-reported usage. It may remain
+zero until completion when a provider does not report usage during streaming.
+
+Example starting a tool call:
+```json
+{"type":"message_update","usage":{...},"assistantMessageEvent":{"type":"toolcall_start","contentIndex":1,"id":"call_abc123","toolName":"write"}}
+```
+
+`message_update` intentionally omits the former cumulative `message` field and
+`assistantMessageEvent.partial`. Clients that need a live partial message must assemble it
+from `message_start` and subsequent events using `contentIndex`. Treat `message_end.message`
+as authoritative. For tool calls, `toolcall_start` provides the call `id` and `toolName`;
+buffer `toolcall_delta.delta` for arguments. `toolcall_end.toolCall` contains the completed
+call.
+
+### bash_execution_update
+
+Emitted once for each output chunk from a direct `bash` command. `id` matches the command's `id`, allowing clients to associate output with the correct command.
+
+Events stream all output while the command runs, even if the final `bash` response's `output` is truncated.
+
+```json
+{
+  "type": "bash_execution_update",
+  "id": "req-1",
+  "delta": "total 48\n"
+}
+```
+
+### tool_execution_start / tool_execution_update / tool_execution_end
+
+Emitted when a tool begins, streams progress, and completes execution.
+
+```json
+{
+  "type": "tool_execution_start",
+  "toolCallId": "call_abc123",
+  "toolName": "bash",
+  "args": {"command": "ls -la"}
+}
+```
+
+During execution, `tool_execution_update` events stream partial results (e.g., bash output as it arrives):
+
+```json
+{
+  "type": "tool_execution_update",
+  "toolCallId": "call_abc123",
+  "toolName": "bash",
+  "args": {"command": "ls -la"},
+  "partialResult": {
+    "content": [{"type": "text", "text": "partial output so far..."}],
+    "details": {"truncation": null, "fullOutputPath": null}
+  }
+}
+```
+
+When complete:
+
+```json
+{
+  "type": "tool_execution_end",
+  "toolCallId": "call_abc123",
+  "toolName": "bash",
+  "result": {
+    "content": [{"type": "text", "text": "total 48\n..."}],
+    "details": {...}
+  },
+  "isError": false
+}
+```
+
+Use `toolCallId` to correlate events. The `partialResult` in `tool_execution_update` contains the accumulated output so far (not just the delta), allowing clients to simply replace their display on each update.
+
+### queue_update
+
+Emitted whenever the pending steering or follow-up queue changes.
+
+```json
+{
+  "type": "queue_update",
+  "steering": ["Focus on error handling"],
+  "followUp": ["After that, summarize the result"]
+}
+```
+
+### compaction_start / compaction_end
+
+Emitted when compaction runs, whether manual or automatic.
+
+```json
+{"type": "compaction_start", "reason": "threshold"}
+```
+
+The `reason` field is `"manual"`, `"threshold"`, or `"overflow"`.
+
+```json
+{
+  "type": "compaction_end",
+  "reason": "threshold",
+  "result": {
+    "summary": "Summary of conversation...",
+    "firstKeptEntryId": "abc123",
+    "tokensBefore": 150000,
+    "estimatedTokensAfter": 32000,
+    "usage": {
+      "input": 32000,
+      "output": 1200,
+      "cacheRead": 0,
+      "cacheWrite": 0,
+      "totalTokens": 33200,
+      "cost": {"input": 0.01, "output": 0.02, "cacheRead": 0, "cacheWrite": 0, "total": 0.03}
+    },
+    "details": {}
+  },
+  "aborted": false,
+  "willRetry": false
+}
+```
+
+If `reason` was `"overflow"` and compaction succeeds, `willRetry` is `true` and the agent will automatically retry the prompt.
+
+If compaction was aborted, `result` is `null` and `aborted` is `true`.
+
+If compaction failed (e.g., API quota exceeded), `result` is `null`, `aborted` is `false`, and `errorMessage` contains the error description.
+
+### auto_retry_start / auto_retry_end
+
+Emitted when automatic retry is triggered after a transient error (overloaded, rate limit, 5xx).
+
+```json
+{
+  "type": "auto_retry_start",
+  "attempt": 1,
+  "maxAttempts": 3,
+  "delayMs": 2000,
+  "errorMessage": "529 {\"type\":\"error\",\"error\":{\"type\":\"overloaded_error\",\"message\":\"Overloaded\"}}"
+}
+```
+
+```json
+{
+  "type": "auto_retry_end",
+  "success": true,
+  "attempt": 2
+}
+```
+
+On final failure (max retries exceeded):
+```json
+{
+  "type": "auto_retry_end",
+  "success": false,
+  "attempt": 3,
+  "finalError": "529 overloaded_error: Overloaded"
+}
+```
+
+### summarization_retry_scheduled / summarization_retry_attempt_start / summarization_retry_finished
+
+Emitted when compaction or branch-summary summarization retries after a transient provider error. These events use the same retry settings as automatic assistant-turn retries.
+
+```json
+{
+  "type": "summarization_retry_scheduled",
+  "attempt": 1,
+  "maxAttempts": 3,
+  "delayMs": 2000,
+  "errorMessage": "terminated"
+}
+```
+
+```json
+{
+  "type": "summarization_retry_attempt_start",
+  "source": "compaction",
+  "reason": "threshold"
+}
+```
+
+For branch summaries, `source` is `"branchSummary"` and no `reason` is present.
+
+```json
+{
+  "type": "summarization_retry_finished"
+}
+```
+
+## Plugin Interaction Protocol
+
+Capability plugins request user interaction through the versioned interaction API (`notify`, `confirm`, `select`, `input`). In RPC mode, these are translated into a request/response sub-protocol on top of the base command/event flow. The wire type names (`extension_ui_request`/`extension_ui_response`) predate the capability surface and are kept as the frozen wire contract.
+
+There are two categories of interaction methods:
+
+- **Dialog methods** (`select`, `confirm`, `input`): emit an `extension_ui_request` on stdout and block until the client sends back an `extension_ui_response` on stdin with the matching `id`.
+- **Fire-and-forget methods** (`notify`): emit an `extension_ui_request` on stdout but do not expect a response. The client can display the information or ignore it.
+
+Dialog timeouts are owned by the agent side: if a dialog has a timeout configured, the agent auto-resolves with a default value when the timeout expires. The client does not need to track timeouts.
+
+The frozen wire contract also defines `editor`, `setStatus`, `setWidget`, `setTitle`, and `set_editor_text` request methods; the current capability interaction surface does not emit them, but clients should tolerate unknown methods.
+
+### Interaction Requests (stdout)
+
+All requests have `type: "extension_ui_request"`, a unique `id`, and a `method` field.
+
+#### select
+
+Prompt the user to choose from a list.
+
+```json
+{
+  "type": "extension_ui_request",
+  "id": "uuid-1",
+  "method": "select",
+  "title": "Allow dangerous command?",
+  "options": ["Allow", "Block"]
+}
+```
+
+Expected response: `extension_ui_response` with `value` (the selected option string) or `cancelled: true`.
+
+#### confirm
+
+Prompt the user for yes/no confirmation.
+
+```json
+{
+  "type": "extension_ui_request",
+  "id": "uuid-2",
+  "method": "confirm",
+  "title": "Clear session?",
+  "message": "All messages will be lost."
+}
+```
+
+Expected response: `extension_ui_response` with `confirmed: true/false` or `cancelled: true`.
+
+#### input
+
+Prompt the user for free-form text.
+
+```json
+{
+  "type": "extension_ui_request",
+  "id": "uuid-3",
+  "method": "input",
+  "title": "Enter a value",
+  "placeholder": "type something..."
+}
+```
+
+Expected response: `extension_ui_response` with `value` (the entered text) or `cancelled: true`.
+
+#### notify
+
+Display a notification. Fire-and-forget, no response expected.
+
+```json
+{
+  "type": "extension_ui_request",
+  "id": "uuid-4",
+  "method": "notify",
+  "message": "Command blocked by user",
+  "notifyType": "warning"
+}
+```
+
+The `notifyType` field is `"info"`, `"warning"`, or `"error"`. Defaults to `"info"` if omitted.
+
+### Interaction Responses (stdin)
+
+Responses are sent for dialog methods only (`select`, `confirm`, `input`). The `id` must match the request.
+
+#### Value response (select, input)
+
+```json
+{"type": "extension_ui_response", "id": "uuid-1", "value": "Allow"}
+```
+
+#### Confirmation response (confirm)
+
+```json
+{"type": "extension_ui_response", "id": "uuid-2", "confirmed": true}
+```
+
+#### Cancellation response (any dialog)
+
+Dismiss any dialog method. The plugin receives `undefined` (for select/input) or `false` (for confirm).
+
+```json
+{"type": "extension_ui_response", "id": "uuid-3", "cancelled": true}
+```
+
+## Error Handling
+
+Failed commands return a response with `success: false`:
+
+```json
+{
+  "type": "response",
+  "command": "set_model",
+  "success": false,
+  "error": "Model not found: invalid/model"
+}
+```
+
+Parse errors:
+
+```json
+{
+  "type": "response",
+  "command": "parse",
+  "success": false,
+  "error": "Failed to parse command: Unexpected token..."
+}
+```
+
+## Types
+
+Source files:
+- [`packages/ai/src/types.ts`](../../ai/src/types.ts) - `Model`, `UserMessage`, `AssistantMessage`, `ToolResultMessage`
+- [`packages/agent/src/types.ts`](../../agent/src/types.ts) - `AgentMessage`, `AgentEvent`
+- [`src/core/messages.ts`](../src/core/messages.ts) - `BashExecutionMessage`
+- [`src/modes/json-event.ts`](../src/modes/json-event.ts) - `JsonAgentSessionEvent`
+- [`src/modes/rpc/rpc-types.ts`](../src/modes/rpc/rpc-types.ts) - RPC command/response types, extension UI request/response types
+
+### Model
+
+```json
+{
+  "id": "claude-sonnet-4-20250514",
+  "name": "Claude Sonnet 4",
+  "api": "anthropic-messages",
+  "provider": "anthropic",
+  "baseUrl": "https://api.anthropic.com",
+  "reasoning": true,
+  "input": ["text", "image"],
+  "contextWindow": 200000,
+  "maxTokens": 16384,
+  "cost": {
+    "input": 3.0,
+    "output": 15.0,
+    "cacheRead": 0.3,
+    "cacheWrite": 3.75
+  }
+}
+```
+
+### UserMessage
+
+```json
+{
+  "role": "user",
+  "content": "Hello!",
+  "timestamp": 1733234567890,
+  "attachments": []
+}
+```
+
+The `content` field can be a string or an array of `TextContent`/`ImageContent` blocks.
+
+### AssistantMessage
+
+```json
+{
+  "role": "assistant",
+  "content": [
+    {"type": "text", "text": "Hello! How can I help?"},
+    {"type": "thinking", "thinking": "User is greeting me..."},
+    {"type": "toolCall", "id": "call_123", "name": "bash", "arguments": {"command": "ls"}}
+  ],
+  "api": "anthropic-messages",
+  "provider": "anthropic",
+  "model": "claude-sonnet-4-20250514",
+  "usage": {
+    "input": 100,
+    "output": 50,
+    "cacheRead": 0,
+    "cacheWrite": 0,
+    "cost": {"input": 0.0003, "output": 0.00075, "cacheRead": 0, "cacheWrite": 0, "total": 0.00105}
+  },
+  "stopReason": "stop",
+  "timestamp": 1733234567890
+}
+```
+
+Stop reasons: `"stop"`, `"length"`, `"toolUse"`, `"error"`, `"aborted"`
+
+### ToolResultMessage
+
+```json
+{
+  "role": "toolResult",
+  "toolCallId": "call_123",
+  "toolName": "bash",
+  "content": [{"type": "text", "text": "total 48\ndrwxr-xr-x ..."}],
+  "usage": {
+    "input": 100,
+    "output": 50,
+    "cacheRead": 0,
+    "cacheWrite": 0,
+    "totalTokens": 150,
+    "cost": {"input": 0.0003, "output": 0.00075, "cacheRead": 0, "cacheWrite": 0, "total": 0.00105}
+  },
+  "isError": false,
+  "timestamp": 1733234567890
+}
+```
+
+`usage` is optional and reports nested LLM work performed by the tool. When present, it contributes to session token and cost totals.
+
+### BashExecutionMessage
+
+Created by the `bash` RPC command (not by LLM tool calls):
+
+```json
+{
+  "role": "bashExecution",
+  "command": "ls -la",
+  "output": "total 48\ndrwxr-xr-x ...",
+  "exitCode": 0,
+  "cancelled": false,
+  "truncated": false,
+  "fullOutputPath": null,
+  "timestamp": 1733234567890
+}
+```
+
+### Attachment
+
+```json
+{
+  "id": "img1",
+  "type": "image",
+  "fileName": "photo.jpg",
+  "mimeType": "image/jpeg",
+  "size": 102400,
+  "content": "base64-encoded-data...",
+  "extractedText": null,
+  "preview": null
+}
+```
+
+## Example: Basic Client (Python)
+
+```python
+import subprocess
+import json
+
+proc = subprocess.Popen(
+    ["agent-forge", "--mode", "rpc", "--no-session"],
+    stdin=subprocess.PIPE,
+    stdout=subprocess.PIPE,
+    text=True
+)
+
+def send(cmd):
+    proc.stdin.write(json.dumps(cmd) + "\n")
+    proc.stdin.flush()
+
+def read_events():
+    for line in proc.stdout:
+        yield json.loads(line)
+
+# Send prompt
+send({"type": "prompt", "message": "Hello!"})
+
+# Process events
+for event in read_events():
+    if event.get("type") == "message_update":
+        delta = event.get("assistantMessageEvent", {})
+        if delta.get("type") == "text_delta":
+            print(delta["delta"], end="", flush=True)
+    
+    if event.get("type") == "agent_end":
+        print()
+        break
+```
+
+## Example: Interactive Client (Node.js)
+
+See [`test/rpc-example.ts`](../test/rpc-example.ts) for a complete interactive example, or [`src/modes/rpc/rpc-client.ts`](../src/modes/rpc/rpc-client.ts) for a typed client implementation.
+
+For a complete example of handling the extension UI protocol, see [`examples/rpc-extension-ui.ts`](../examples/rpc-extension-ui.ts).
+
+```javascript
+const { spawn } = require("child_process");
+const { StringDecoder } = require("string_decoder");
+
+const agent = spawn("agent-forge", ["--mode", "rpc", "--no-session"]);
+
+function attachJsonlReader(stream, onLine) {
+    const decoder = new StringDecoder("utf8");
+    let buffer = "";
+
+    stream.on("data", (chunk) => {
+        buffer += typeof chunk === "string" ? chunk : decoder.write(chunk);
+
+        while (true) {
+            const newlineIndex = buffer.indexOf("\n");
+            if (newlineIndex === -1) break;
+
+            let line = buffer.slice(0, newlineIndex);
+            buffer = buffer.slice(newlineIndex + 1);
+            if (line.endsWith("\r")) line = line.slice(0, -1);
+            onLine(line);
+        }
+    });
+
+    stream.on("end", () => {
+        buffer += decoder.end();
+        if (buffer.length > 0) {
+            onLine(buffer.endsWith("\r") ? buffer.slice(0, -1) : buffer);
+        }
+    });
+}
+
+attachJsonlReader(agent.stdout, (line) => {
+    const event = JSON.parse(line);
+
+    if (event.type === "message_update") {
+        const { assistantMessageEvent } = event;
+        if (assistantMessageEvent.type === "text_delta") {
+            process.stdout.write(assistantMessageEvent.delta);
+        }
+    }
+});
+
+// Send prompt
+agent.stdin.write(JSON.stringify({ type: "prompt", message: "Hello" }) + "\n");
+
+// Abort on Ctrl+C
+process.on("SIGINT", () => {
+    agent.stdin.write(JSON.stringify({ type: "abort" }) + "\n");
+});
+```