Roo-Code/apps/cli/docs/AGENT_LOOP.md
Daniel d52b6834e3
Add back post-revert bug fixes and features (Step 2) (#11463)
* fix: cancel backend auto-approval timeout when auto-approve is toggled off mid-countdown (#11439)

Co-authored-by: Sannidhya <sann@Sannidhyas-MacBook-Pro.local>

* fix: prevent chat history loss during cloud/settings navigation (#11371) (#11372)

Co-authored-by: Sannidhya <sann@Sannidhyas-MacBook-Pro.local>

* fix: preserve pasted images in chatbox during chat activity (#11375)

Co-authored-by: Roo Code <roomote@roocode.com>

* fix: resolve chat scroll anchoring and task-switch scroll race condit… (#11385)

* fix: avoid zsh process-substitution false positives in assignments (#11365)

* fix(editor): make tab close best-effort in DiffViewProvider.open (#11363)

* fix(checkpoints): canonicalize core.worktree comparison to prevent Windows path mismatch failures (#11346)

* fix: prevent double notification sound playback (#11283)

* fix: prevent false unsaved changes prompt with OpenAI Compatible headers (#8230) (#11334)

fix: prevent false unsaved changes prompt with OpenAI Compatible headers

Mark automatic header syncs in ApiOptions and OpenAICompatible as
non-user actions (isUserAction: false) and enhance SettingsView change
detection to skip automatic syncs with semantically equal values.

Root cause: two components (ApiOptions and OpenAICompatible) manage
openAiHeaders state and automatically sync it back on mount/remount.
These syncs were treated as user changes, triggering a false dirty state.

Co-authored-by: Robert McIntyre <robertjmcintyre@users.noreply.github.com>

* fix: remove noisy console.warn logs from NativeToolCallParser (#11264)

Remove two console.warn messages that fire excessively when loading tasks
from history:
- 'Attempting to finalize unknown tool call' in finalizeStreamingToolCall()
- 'Received chunk for unknown tool call' in processStreamingChunk()

The defensive null-return behavior is preserved; only the log output is removed.

* refactor: remove footgun prompting (file-based system prompt override) (#11387)

* refactor: delete orphaned per-provider caching transform files (#11388)

* feat: add disabledTools setting to globally disable native tools (#11277)

* feat: add disabledTools setting to globally disable native tools

Add a disabledTools field to GlobalSettings that allows disabling specific
native tools by name. This enables cloud agents to be configured with
restricted tool access.

Schema:
- Add disabledTools: z.array(toolNamesSchema).optional() to globalSettingsSchema
- Add disabledTools to organizationDefaultSettingsSchema.pick()
- Add disabledTools to ExtensionState Pick type

Prompt generation (tool filtering):
- Add disabledTools to BuildToolsOptions interface
- Pass disabledTools through filterSettings to filterNativeToolsForMode()
- Remove disabled tools from allowedToolNames set in filterNativeToolsForMode()

Execution-time validation (safety net):
- Extract disabledTools from state in presentAssistantMessage
- Convert disabledTools to toolRequirements format for validateToolUse()

Wiring:
- Add disabledTools to ClineProvider getState() and getStateToPostToWebview()
- Pass disabledTools to all buildNativeToolsArrayWithRestrictions() call sites

EXT-778

* fix: check toolRequirements before ALWAYS_AVAILABLE_TOOLS

Moves the toolRequirements check before the ALWAYS_AVAILABLE_TOOLS
early-return in isToolAllowedForMode(). This ensures disabledTools
can block always-available tools (switch_mode, new_task, etc.) at
execution time, making the validation layer consistent with the
filtering layer.

* feat: add support for .agents/skills directory (#11181)

* feat: add support for .agents/skills directory

This change adds support for discovering skills from the .agents/skills
directory, following the Agent Skills convention for sharing skills
across different AI coding tools.

Priority order (later entries override earlier ones):
1. Global ~/.agents/skills (shared across AI coding tools, lowest priority)
2. Project .agents/skills
3. Global ~/.roo/skills (Roo-specific)
4. Project .roo/skills (highest priority)

Changes:
- Add getGlobalAgentsDirectory() and getProjectAgentsDirectoryForCwd()
  functions to roo-config
- Update SkillsManager.getSkillsDirectories() to include .agents/skills
- Update SkillsManager.setupFileWatchers() to watch .agents/skills
- Add tests for new functionality

* fix: clarify skill priority comment to match actual behavior

* fix: clarify skill priority comment to explain Map.set replacement mechanism

---------

Co-authored-by: Roo Code <roomote@roocode.com>

* feat(history): render nested subtasks as recursive tree (#11299)

* feat(history): render nested subtasks as recursive tree

* fix(lockfile): resolve missing ai-sdk provider entry

* fix: address review feedback — dedupe countAll, increase SubtaskRow max-h

- HistoryView: replace local countAll with imported countAllSubtasks from types.ts
- SubtaskRow: increase nested children max-h from 500px to 2000px to match TaskGroupItem

* perf(refactor): consolidate getState calls in resolveWebviewView (#11320)

* perf(refactor): consolidate getState calls in resolveWebviewView

Replace three separate this.getState().then() calls with a single
await this.getState() and destructuring. This avoids running the
full getState() method (CloudService calls, ContextProxy reads, etc.)
three times during webview view resolution.

* fix: keep getState consolidation non-blocking to avoid delaying webview render

---------

Co-authored-by: daniel-lxs <ricciodaniel98@gmail.com>

* fix: harden command auto-approval against inline JS false positives (#11382)

* feat: rename search_and_replace tool to edit and unify edit-family UI (#11296)

* Revert "refactor: delete orphaned per-provider caching transform files (#11388)"

This reverts commit 13a45b0361.

* chore: regenerate built-in-skills.ts with updated formatting

* fix: add missing maxReadFileLine property to test baseState

The ExtensionState type now requires maxReadFileLine property (added in commit 63e3f769a).
Update the test to include this property with the default value of -1 (unlimited reading).

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

* feat: add pnpm serve command for code-server development (#10964)

Co-authored-by: Roo Code <roomote@roocode.com>

* chore: remove Feature Request from issue template options (#11141)

Co-authored-by: Roo Code <roomote@roocode.com>

* refactor(docs-extractor): simplify mode to focus on raw fact extraction (#11129)

* Add cli support for linux (#11167)

* fix: replace heredocs with echo statements in cli-release workflow (#11168)

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>

* Drop MacOS-13 cli support (#11169)

* fix(cli): correct example in install script (#11170)

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>

* feat: add Kimi K2.5 model to Fireworks provider (#11177)

* feat(cli): improve dev experience and roo provider API key support (#11203)

- Allow --api-key and ROO_API_KEY env var for the roo provider instead of
  requiring cloud auth token
- Switch dev/start scripts to use tsx for running directly from source
  without building first
- Fix path resolution (version.ts, extension.ts, extension-host.ts) to
  work from both source and bundled locations
- Disable debug log file (~/.roo/cli-debug.log) unless --debug is passed
- Update README with complete env var table and dev workflow docs

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>

* Roo Code CLI v0.0.50 (#11204)

* Roo Code CLI v0.0.50

* docs(cli): add --exit-on-error to changelog

---------

Co-authored-by: Roo Code <roomote@roocode.com>

* feat(cli): update default model from Opus 4.5 to Opus 4.6 (#11273)

Co-authored-by: Roo Code <roomote@roocode.com>

* feat(web): replace Roomote Control with Linear Integration in cloud features grid (#11280)

Co-authored-by: Roo Code <roomote@roocode.com>

* Add linux-arm64 for the roo cli (#11314)

* chore: clean up repo-facing mode rules (#11410)

* Make CLI auto-approve by default with require-approval opt-in (#11424)

Co-authored-by: Roo Code <roomote@roocode.com>

* Add new code owners to CODEOWNERS file

* Update next.js (#11108)

* feat(web): Replace bespoke navigation menu with shadcn navigation menu (#11117)

Co-authored-by: Roo Code <roomote@roocode.com>

---------

Co-authored-by: SannidhyaSah <sah_sannidhya@outlook.com>
Co-authored-by: Sannidhya <sann@Sannidhyas-MacBook-Pro.local>
Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com>
Co-authored-by: Roo Code <roomote@roocode.com>
Co-authored-by: Hannes Rudolph <hrudolph@gmail.com>
Co-authored-by: 0xMink <dennis@dennismink.com>
Co-authored-by: Robert McIntyre <robertjmcintyre@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
Co-authored-by: Matt Rubens <mrubens@users.noreply.github.com>
Co-authored-by: Chris Estreich <cestreich@gmail.com>
2026-02-13 18:40:28 -05:00

356 lines
13 KiB
Markdown

# CLI Agent Loop
This document explains how the Roo Code CLI detects and tracks the agent loop state.
## Overview
The CLI needs to know when the agent is:
- **Running** (actively processing)
- **Streaming** (receiving content from the API)
- **Waiting for input** (needs user approval or answer)
- **Idle** (task completed or failed)
This is accomplished by analyzing the messages the extension sends to the client.
## The Message Model
All agent activity is communicated through **ClineMessages** - a stream of timestamped messages that represent everything the agent does.
### Message Structure
```typescript
interface ClineMessage {
ts: number // Unique timestamp identifier
type: "ask" | "say" // Message category
ask?: ClineAsk // Specific ask type (when type="ask")
say?: ClineSay // Specific say type (when type="say")
text?: string // Message content
partial?: boolean // Is this message still streaming?
}
```
### Two Types of Messages
| Type | Purpose | Blocks Agent? |
| ------- | ---------------------------------------------- | ------------- |
| **say** | Informational - agent is telling you something | No |
| **ask** | Interactive - agent needs something from you | Usually yes |
## The Key Insight
> **The agent loop stops whenever the last message is an `ask` type (with `partial: false`).**
The specific `ask` value tells you exactly what the agent needs.
## Ask Categories
The CLI categorizes asks into four groups:
### 1. Interactive Asks → `WAITING_FOR_INPUT` state
These require user action to continue:
| Ask Type | What It Means | Required Response |
| ----------------------- | --------------------------------- | ----------------- |
| `tool` | Wants to edit/create/delete files | Approve or Reject |
| `command` | Wants to run a terminal command | Approve or Reject |
| `followup` | Asking a question | Text answer |
| `browser_action_launch` | Wants to use the browser | Approve or Reject |
| `use_mcp_server` | Wants to use an MCP server | Approve or Reject |
### 2. Idle Asks → `IDLE` state
These indicate the task has stopped:
| Ask Type | What It Means | Response Options |
| ------------------------------- | --------------------------- | --------------------------- |
| `completion_result` | Task completed successfully | New task or feedback |
| `api_req_failed` | API request failed | Retry or new task |
| `mistake_limit_reached` | Too many errors | Continue anyway or new task |
| `auto_approval_max_req_reached` | Auto-approval limit hit | Continue manually or stop |
| `resume_completed_task` | Viewing completed task | New task |
### 3. Resumable Asks → `RESUMABLE` state
| Ask Type | What It Means | Response Options |
| ------------- | ------------------------- | ----------------- |
| `resume_task` | Task paused mid-execution | Resume or abandon |
### 4. Non-Blocking Asks → `RUNNING` state
| Ask Type | What It Means | Response Options |
| ---------------- | ------------------ | ----------------- |
| `command_output` | Command is running | Continue or abort |
## Streaming Detection
The agent is **streaming** when:
1. **`partial: true`** on the last message, OR
2. **An `api_req_started` message exists** with `cost: undefined` in its text field
```typescript
// Streaming detection pseudocode
function isStreaming(messages) {
const lastMessage = messages.at(-1)
// Check partial flag (primary indicator)
if (lastMessage?.partial === true) {
return true
}
// Check for in-progress API request
const apiReq = messages.findLast((m) => m.say === "api_req_started")
if (apiReq?.text) {
const data = JSON.parse(apiReq.text)
if (data.cost === undefined) {
return true // API request not yet complete
}
}
return false
}
```
## State Machine
```
┌─────────────────┐
│ NO_TASK │ (no messages)
└────────┬────────┘
│ newTask
┌─────────────────────────────┐
┌───▶│ RUNNING │◀───┐
│ └──────────┬──────────────────┘ │
│ │ │
│ ┌──────────┼──────────────┐ │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ ┌──────┐ ┌─────────┐ ┌──────────┐ │
│ │STREAM│ │WAITING_ │ │ IDLE │ │
│ │ ING │ │FOR_INPUT│ │ │ │
│ └──┬───┘ └────┬────┘ └────┬─────┘ │
│ │ │ │ │
│ │ done │ approved │ newTask │
└────┴───────────┴────────────┘ │
┌──────────────┐ │
│ RESUMABLE │────────────────────────┘
└──────────────┘ resumed
```
## Architecture
```
┌─────────────────────────────────────────────────────────────────┐
│ ExtensionHost │
│ │
│ ┌──────────────────┐ │
│ │ Extension │──── extensionWebviewMessage ─────┐ │
│ │ (Task.ts) │ │ │
│ └──────────────────┘ │ │
│ ▼ │
│ ┌───────────────────────────────────────────────────────────┐ │
│ │ ExtensionClient │ │
│ │ (Single Source of Truth) │ │
│ │ │ │
│ │ ┌─────────────────┐ ┌────────────────────┐ │ │
│ │ │ MessageProcessor │───▶│ StateStore │ │ │
│ │ │ │ │ (clineMessages) │ │ │
│ │ └─────────────────┘ └────────┬───────────┘ │ │
│ │ │ │ │
│ │ ▼ │ │
│ │ detectAgentState() │ │
│ │ │ │ │
│ │ ▼ │ │
│ │ Events: stateChange, message, waitingForInput, etc. │ │
│ └───────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌────────────────┐ ┌────────────────┐ ┌────────────────┐ │
│ │ OutputManager │ │ AskDispatcher │ │ PromptManager │ │
│ │ (stdout) │ │ (ask routing) │ │ (user input) │ │
│ └────────────────┘ └────────────────┘ └────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
```
## Component Responsibilities
### ExtensionClient
The **single source of truth** for agent state, including the current mode. It:
- Receives all messages from the extension
- Stores them in the `StateStore`
- Tracks the current mode from state messages
- Computes the current state via `detectAgentState()`
- Emits events when state changes (including mode changes)
```typescript
const client = new ExtensionClient({
sendMessage: (msg) => extensionHost.sendToExtension(msg),
debug: true, // Writes to ~/.roo/cli-debug.log
})
// Query state at any time
const state = client.getAgentState()
if (state.isWaitingForInput) {
console.log(`Agent needs: ${state.currentAsk}`)
}
// Query current mode
const mode = client.getCurrentMode()
console.log(`Current mode: ${mode}`) // e.g., "code", "architect", "ask"
// Subscribe to events
client.on("waitingForInput", (event) => {
console.log(`Waiting for: ${event.ask}`)
})
// Subscribe to mode changes
client.on("modeChanged", (event) => {
console.log(`Mode changed: ${event.previousMode} -> ${event.currentMode}`)
})
```
### StateStore
Holds the `clineMessages` array, computed state, and current mode:
```typescript
interface StoreState {
messages: ClineMessage[] // The raw message array
agentState: AgentStateInfo // Computed state
isInitialized: boolean // Have we received any state?
currentMode: string | undefined // Current mode (e.g., "code", "architect")
}
```
### MessageProcessor
Handles incoming messages from the extension:
- `"state"` messages → Update `clineMessages` array and track mode
- `"messageUpdated"` messages → Update single message in array
- Emits events for state transitions and mode changes
### AskDispatcher
Routes asks to appropriate handlers:
- Uses type guards: `isIdleAsk()`, `isInteractiveAsk()`, etc.
- Coordinates between `OutputManager` and `PromptManager`
- By default, the CLI auto-approves tool/command/browser/MCP actions
- In `--require-approval` mode, those actions prompt for manual approval
### OutputManager
Handles all CLI output:
- Streams partial content with delta computation
- Tracks what's been displayed to avoid duplicates
- Writes directly to `process.stdout` (bypasses quiet mode)
### PromptManager
Handles user input:
- Yes/no prompts
- Text input prompts
- Timed prompts with auto-defaults
## Response Messages
When the agent is waiting, send these responses:
```typescript
// Approve an action (tool, command, browser, MCP)
client.sendMessage({
type: "askResponse",
askResponse: "yesButtonClicked",
})
// Reject an action
client.sendMessage({
type: "askResponse",
askResponse: "noButtonClicked",
})
// Answer a question
client.sendMessage({
type: "askResponse",
askResponse: "messageResponse",
text: "My answer here",
})
// Start a new task
client.sendMessage({
type: "newTask",
text: "Build a web app",
})
// Cancel current task
client.sendMessage({
type: "cancelTask",
})
```
## Type Guards
The CLI uses type guards from `@roo-code/types` for categorization:
```typescript
import { isIdleAsk, isInteractiveAsk, isResumableAsk, isNonBlockingAsk } from "@roo-code/types"
const ask = message.ask
if (isInteractiveAsk(ask)) {
// Needs approval: tool, command, followup, etc.
} else if (isIdleAsk(ask)) {
// Task stopped: completion_result, api_req_failed, etc.
} else if (isResumableAsk(ask)) {
// Task paused: resume_task
} else if (isNonBlockingAsk(ask)) {
// Command running: command_output
}
```
## Debug Logging
Enable with `-d` flag. Logs go to `~/.roo/cli-debug.log`:
```bash
roo -d -P "Build something" --no-tui
```
View logs:
```bash
tail -f ~/.roo/cli-debug.log
```
Example output:
```
[MessageProcessor] State update: {
"messageCount": 5,
"lastMessage": {
"msgType": "ask:completion_result"
},
"stateTransition": "running → idle",
"currentAsk": "completion_result",
"isWaitingForInput": true
}
[MessageProcessor] EMIT waitingForInput: { "ask": "completion_result" }
[MessageProcessor] EMIT taskCompleted: { "success": true }
```
## Summary
1. **Agent communicates via `ClineMessage` stream**
2. **Last message determines state**
3. **`ask` messages (non-partial) block the agent**
4. **Ask category determines required action**
5. **`partial: true` or `api_req_started` without cost = streaming**
6. **`ExtensionClient` is the single source of truth**