Add all CherryClaw-related UI strings for soul, scheduler, heartbeat,
channels (placeholder), agent type names, and bypass warning.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…spatch

- Create AgentServiceRegistry mapping AgentType → AgentServiceInterface
- Refactor SessionMessageService to use registry instead of hardcoded ClaudeCodeService
- Register ClaudeCodeService for 'claude-code' in services module init

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Update BaseService.listMcpTools, BaseService.listSlashCommands, and
SessionService.listSlashCommands to handle cherry-claw alongside claude-code.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…beatReader

- SoulReader: reads soul.md from workspace with mtime-based caching
- HeartbeatReader: reads heartbeat.md on demand with path traversal protection
- CherryClawService: delegates to claude-code with soul-enhanced system prompt
- Register cherry-claw in AgentServiceRegistry

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…operation

- Add cron-parser dependency for cron expression parsing
- Implement SchedulerService with cron/interval/one-time scheduling
- Per-agent tick guard prevents overlapping invocations
- Pause after 3 consecutive errors
- Restore schedulers on app boot via restoreSchedulers()
- Deliver heartbeat content to most recent session on each tick

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

- Restore schedulers on app startup (after API server init)
- Stop all schedulers on app quit
- Stop scheduler when agent is deleted
- Restart scheduler when cherry-claw agent is updated/patched

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

- Add type dropdown (Claude Code / CherryClaw) in creation modal
- Auto-apply CherryClaw defaults when cherry-claw type is selected
- Show bypassPermissions warning for CherryClaw agents
- Type selection only shown during creation (not editing)

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

… channels)

- Add SoulSettings, SchedulerSettings, HeartbeatSettings, ChannelsSettings components
- Conditionally show CherryClaw-specific tabs in AgentSettingsPopup
- Add cherry-claw to SettingsPopupTab type union
- Add cherry-claw label to getAgentTypeLabel

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

- CherryClaw agents show 🦞 default emoji instead of ⭐️
- Agent name falls back to type label which returns 'CherryClaw'

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

- AgentServiceRegistry: 4 tests (register, getService, hasService, overwrite)
- SchedulerService: 9 tests (start/stop, tick guard, status, lifecycle)
- SoulReader: 4 tests (read, missing file, mtime cache, re-read)
- HeartbeatReader: 5 tests (read, missing, custom file, path traversal, defaults)

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

- Fix cron-parser import (CronExpressionParser.parse instead of parseExpression)
- Fix logger error arg types in soul.ts and heartbeat.ts
- Fix oxlint consistent-type-imports in test files
- Apply biome formatting

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Auto-create a default CherryClaw agent on first list if none exists,
using the first available Anthropic model. Cherry-claw agents are
always sorted to the top of the agent list response.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Remove soul, scheduler, heartbeat, prompt, and permission mode tabs
for CherryClaw agents. Rename "Tools & MCP" to "MCP" with a new
lean McpSettings component that only shows external MCP server config.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Replace per-agent setTimeout scheduler with a poll-loop architecture
inspired by nanoclaw. Tasks are now first-class DB entities that users
can create, edit, pause, and delete through the agent settings UI.

- Add scheduled_tasks + task_run_logs Drizzle tables with FK cascades
- Add TaskService with CRUD, getDueTasks, drift-resistant computeNextRun
- Rewrite SchedulerService as 60s poll loop querying DB for due tasks
- Add REST API routes: /agents/:agentId/tasks (CRUD + logs)
- Add AgentApiClient task methods and SWR hooks (useTasks, etc.)
- Add Tasks tab in CherryClaw agent settings (replaces Channels)
- Add TaskFormModal, TaskListItem, TaskLogsModal UI components
- Update i18n with properly nested task UI strings (11 locales)
- Rewrite SchedulerService tests for poll-loop API

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Add an internal MCP server auto-injected into CherryClaw agent sessions,
exposing a `cron` tool (add/list/remove) that lets the agent autonomously
create and manage its own scheduled tasks without human intervention.

- New ClawServer in src/main/mcpServers/claw.ts with cron tool
- Streamable HTTP route at /v1/claw/:agentId/claw-mcp
- CherryClawService injects _internalMcpServers before delegation
- ClaudeCodeService merges internal MCPs into SDK options
- 12 unit tests for tool actions, duration parsing, and validation

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…ools

Two issues preventing the claw MCP cron tool from working:

1. GET requests (SSE streaming) to the claw MCP route caused a 500
   because req.body is undefined for GET/DELETE — only parse JSON-RPC
   body for POST requests, let transport handle GET/DELETE natively.

2. The Claude Code SDK has builtin CronCreate/CronDelete/CronList tools
   that shadow our MCP cron tool. Add disallowedTools to suppress them
   so the agent uses our persistent DB-backed cron tool instead.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…claw'

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Add ChannelAdapter (abstract base), ChannelMessageHandler (message routing,
session tracking, stream response collection), ChannelManager (lifecycle,
adapter factory registry), and barrel exports. No existing files modified.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Implements TelegramAdapter with auth middleware (allowed_chat_ids),
command handlers (/new, /compact, /help), message forwarding, typing
indicators, and message chunking at 4096 chars. Self-registers via
adapter factory pattern.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…ndlers

Start channel adapters on app ready (after scheduler), stop on quit.
Sync channels on agent update (PUT/PATCH) and delete.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Add TelegramChannelConfigSchema with bot_token and allowed_chat_ids.
Update CherryClawChannelSchema with enum type, enabled field, and typed
config union. Enable the channel enabled guard in ChannelManager.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Add ChannelsSettings tab with list view, add/edit modal (bot token,
allowed chat IDs), enable/disable toggle, and delete with confirmation.
Includes ChannelListItem and ChannelFormModal components following
existing TasksSettings patterns.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…, and TelegramAdapter

21 tests covering: stream response accumulation, message chunking, slash
commands (/new, /compact, /help), session tracking, adapter lifecycle,
agent sync, disabled channel filtering, auth middleware, and event emission.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…nfig

Replace CRUD list + modal pattern with a catalog of available channel
types. Each type is a card with enable switch and inline config fields
(blur-to-save). Delete ChannelFormModal and ChannelListItem components.
Add notify receiver checkbox. Future channels shown as "coming soon".

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…ndicators

Use Telegram's sendMessageDraft API to progressively stream agent
responses instead of sending a single message at completion. Typing
indicators run every 4s throughout the request to signal the agent is
working during tool-use gaps.

Also fixes multi-turn accumulation: text-delta values are cumulative
within a block, so each turn's text is committed on text-end and
concatenated across turns to preserve all response content.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…eduler

Channel (Telegram) and scheduler callers of createSessionMessage() never
persisted messages because persistence was entirely UI-driven via IPC.

- Add `persist` option to SessionMessageService.createSessionMessage()
- When true, build minimal AgentPersistedMessage payloads and call
  persistExchange() on stream complete
- Use `agent-session:` prefixed topicId so UI routes correctly
- Drain stream in SchedulerService so completion promise resolves
- Update tests to match new call signatures

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

TextStreamAccumulator used `+=` for text-delta events, but the transform
layer emits cumulative deltas (each contains full text so far). This
caused persisted assistant messages to contain all intermediate states
concatenated together. Use `=` (replace) to match ChannelMessageHandler.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…operation

Disable TodoWrite, AskUserQuestion, EnterPlanMode, ExitPlanMode,
EnterWorktree, and NotebookEdit via _disallowedTools for CherryClaw
agents. These tools require interactive user input or are irrelevant
for autonomous agent operation.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…rcement

Add sandbox_enabled toggle to CherryClaw agent settings that restricts
filesystem access to the workspace path via two layers:

1. PreToolUse hook in ClaudeCodeService inspects every tool call's target
   paths and denies access outside allowed directories. Works regardless
   of permissionMode (including bypassPermissions) since hooks always
   fire before permission checks.

2. SDK OS-level sandbox (sandbox.enabled, allowUnsandboxedCommands: false)
   as a fallback for Bash command isolation.

Tool path extraction covers Read, Edit, Write, MultiEdit, NotebookEdit
(file_path/notebook_path), Glob, Grep (path field), and Bash (regex
extraction of absolute paths from command string).

Note: This is a basic restriction for well-behaved agents, not a security
boundary. Bash regex has known bypasses (relative paths, variable
expansion, subshells). Needs hardening in follow-up work.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Add POST /:taskId/run API endpoint with SchedulerService.runTaskNow()
for manually triggering scheduled tasks. Wire through API client, hook,
and "Run" button in task settings UI. Update handoff.md with Phase 11
(notify tool) and Phase 12 (manual task run) progress.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

- Capture SDK session_id from init message in ClaudeCodeService and
  store on AgentStream.sdkSessionId for headless callers to persist
- SessionMessageService reads sdkSessionId on stream complete and
  passes to persistHeadlessExchange for resume on next scheduler run
- CherryClawService auto-appends mcp__claw__cron and mcp__claw__notify
  to allowed_tools when agent has explicit tool whitelist
- Add debug logging for MCP injection and session ID resolution

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…ifecycle

Add `skills` tool to claw MCP server with search/install/remove/list
actions. Reuses PluginService for install/uninstall/list and queries
the marketplace API (claude-plugins.dev) for search via net.fetch.

Fix MCP session lifecycle: the SDK Server class only supports one
transport at a time (connect() throws if already connected). Changed
from per-agent server caching to per-session Server+Transport pairs
to prevent "No such tool available" errors on reconnect.

- Add SKILLS_TOOL definition with search/install/remove/list actions
- Add SkillSearchResult type and buildSkillIdentifier() helper
- Delegate to PluginService.install/uninstall/listInstalled
- Add mcp__claw__skills to auto-allow list in CherryClawService
- Refactor claw-mcp.ts: per-session ClawServer+Transport pairs
- Add 10 new tests for skills tool (27 total claw tests)

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Replace Claude Code preset system prompt with a full custom prompt
assembled by PromptBuilder from workspace files. Add memory MCP tool
for persistent knowledge management across sessions.

- PromptBuilder reads system.md (optional override), soul.md, user.md,
  and memory/FACT.md with mtime-based caching, assembles anna-style
  XML-tagged memories section with exclusive scope documentation
- memory MCP tool with update (atomic FACT.md write), append (JOURNAL.jsonl
  timestamped entries with tags), and search (case-insensitive, tag filter)
- _systemPrompt field on EnhancedSessionFields for full system prompt
  replacement in ClaudeCodeService (falls back to preset+append for
  regular claude-code agents)
- system.md only overrides the preset/basic part; memories section is
  always appended regardless

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Add resolveFile/resolveFileCI helpers (inspired by anna's resolveFile)
that scan directories with case-insensitive matching. Applied to all
workspace file lookups: system.md, soul.md, user.md, FACT.md, and
JOURNAL.jsonl. Exact match is preferred; falls back to directory scan.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Include explicit file layout diagram in the memories section so the
agent knows exactly where each file lives and how to update it.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Replace the HTTP-based claw MCP connection with an in-memory McpServer
instance, eliminating the network round-trip for tool calls.

- Refactor ClawServer to use McpServer (high-level API) instead of
  low-level Server, exposing mcpServer property
- Add InternalMcpInMemServerConfig type alongside existing HTTP config
- Map inmem configs to SDK's { type: 'sdk', instance } format in
  ClaudeCodeService merge logic
- CherryClawService now instantiates ClawServer directly instead of
  connecting over HTTP via apiConfigService

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…memories section

Replace lines.push() pattern with template literals and consolidate
per-file subsections into a single Memories section with a table layout
for clearer file purpose and update method documentation.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Remove all sandbox-related code including path enforcement in
PreToolUse hook, sandbox settings UI toggle, configuration type
fields, and default config values.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…in system prompt

Add Chinese reference documentation for CherryClaw covering overall design,
scheduler, channels, and MCP claw tools. Add a CherryClaw Tools section to
the system prompt so the agent knows to use mcp__claw__* tools, and update
the Memories section to reference mcp__claw__memory by exact name.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

…d task

Heartbeat is now a special scheduled task (name='heartbeat') managed by
the existing TaskService + SchedulerService poll loop, replacing the old
file-prepend approach.

- HeartbeatReader simplified: always reads heartbeat.md, trims content,
  returns undefined for empty/missing files
- SchedulerService.runTask() detects heartbeat tasks and reads
  heartbeat.md from workspace; wraps content with context prompt
- SchedulerService.ensureHeartbeatTask() creates/updates heartbeat task
  on agent create and config update
- TaskService.listTasks() excludes heartbeat tasks by default
- CherryClawConfiguration: heartbeat_file replaced with
  heartbeat_interval (minutes, default 30)
- UI: HeartbeatSection (toggle + interval) added to Tasks settings page
- Agent handlers sync heartbeat task on create/update/patch
- i18n: simplified heartbeat keys (enabled, enabledHelper, interval,
  intervalHelper) across all 11 locale files

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

Replace placeholder cherry-claw.png with proper lobster+cherries icon.
Add square 768x768 icon to overview doc header.

Signed-off-by: Vaayne <liu.vaayne@gmail.com>

### What this PR does

Before this PR:
CherryClaw only supported Telegram as a channel adapter for receiving
and sending messages.

After this PR:
CherryClaw now supports QQ Bot as an additional channel adapter,
allowing users to connect their autonomous agents to QQ messaging
platform.

Fixes # N/A

### Why we need it and why it was done in this way

The following tradeoffs were made:
- **No streaming support**: QQ Bot API has no native draft/streaming API
like Telegram's `sendMessageDraft`, so streaming responses are not
supported. Full responses are sent as final messages only.
- **No typing indicator**: QQ Bot API doesn't support typing indicators
for most message types, so `sendTypingIndicator` is a no-op.
- **WebSocket over long polling**: Uses WebSocket gateway (like the
reference implementation) for better reliability and lower latency.

The following alternatives were considered:
- Using the `@sliverp/qqbot` package directly: Decided against this as
it's designed as an OpenClaw plugin with heavy dependencies. Instead,
implemented a standalone adapter following the same patterns.

Links to places where the discussion took place:
- Reference implementation: https://github.com/sliverp/qqbot

### Breaking changes

None. This is an additive feature.

### Special notes for your reviewer

- The adapter follows the same pattern as TelegramAdapter
- Supports 4 message types: c2c (private), group, guild (channel), and
dm
- Includes sandbox mode toggle for testing with QQ's sandbox API
- All main process tests pass (38 files, 528 tests)

### Checklist

- [x] PR: The PR description is expressive enough and will help future
contributors
- [x] Code: Write code that humans can understand and Keep it simple
- [x] Refactor: You have left the code cleaner than you found it (Boy
Scout Rule)
- [x] Upgrade: Impact of this change on upgrade flows was considered and
addressed if required
- [ ] Documentation: A user-guide update was considered and is present
(link) or not required
- [x] Self-review: I have reviewed my own code before requesting review
from others

### Release note

```release-note
Add QQ channel adapter for CherryClaw agents - connect your autonomous agent to QQ messaging platform via the official QQ Bot API
```

---------

Signed-off-by: Vaayne <liu.vaayne@gmail.com>
Co-authored-by: kangfenmao <kangfenmao@qq.com>