- code_reviewer.yaml: Verifier Agent skill config for adversarial review
with structured output schema for ReviewFeedback format
- coding_harness.yaml: Example pipeline with adversarial loop
develop → test → review (Worker↔Verifier) → archive
Add ReviewIssue, ReviewFeedback, AdversarialState models and extend
PipelineStage with verifier, max_adversarial_rounds, feedback_mode,
and escalate_on_exhaust fields for Worker-Verifier adversarial loop.
DeepSeek-chat has limited/partial function calling support. Qwen3-coder-plus
(DashScope) has robust OpenAI-compatible function calling.
Also added tool usage instructions to system prompt and enhanced logging
to trace tool propagation through the pipeline.
1. Loading indicator: three-dot bouncing animation appears after
sending a message and disappears when server starts responding.
2. Tool descriptions: resolve_skill_routing now appends available
tools (name + description + parameters) to the system prompt so
the LLM knows what tools it can call.
Root cause: app.py registers tools via agent._tool_registry.register()
which adds to the ToolRegistry but NOT to agent._tools (which is only
populated by use_tool() from config). Both get_tools() and
get_system_prompt() were reading only _tools, missing all post-init
registered tools. Now both methods merge _tools with
_tool_registry.list_tools().
Previously get_system_prompt() only returned identity/instructions but
did not tell the LLM what tools are available. The LLM would therefore
refuse to call tools even when they were registered, saying it had no
tools. Now the system prompt includes a '## 可用工具' section listing
all registered tools with their descriptions and parameters.
e.isComposing is a standard KeyboardEvent property that's true during
IME composition. More reliable than compositionstart/compositionend
which can fire at unpredictable timing relative to keydown.
Added compositionstart/compositionend event listeners to track IME
composing state. Enter key now only submits when not composing,
so Chinese/Japanese/Korean input methods work correctly.
When IntentRouter matches a direct-mode agent (no tools), but the task
content suggests tool needs (shell, search, file ops, etc.), the routing
now falls through to the default agent which has full tool access.
This fixes the issue where "帮我执行个命令" would be routed to
direct_agent and fail because direct mode doesn't support tool calling.
Also restored "你好" in direct_agent keywords since it's correctly
handled now — greetings don't need tools, direct mode is fine.
Hardcoded model names like 'openai/gpt-4o-mini' or 'anthropic/claude-sonnet'
cause 'No provider available' errors when the specific provider isn't configured.
Using 'default' lets the system pick the available provider automatically.
1. InMemoryMessageBus.request(): fix param name (timeout→timeout_seconds) to match ABC
2. InMemoryMessageBus: track consumer tasks, cancel on unsubscribe
3. InMemoryMessageBus: _try_resolve_pending() in queue consumer path
4. evolve_soul(): use "default" category when patterns is empty
5. quick_classify(): use delimiter-based prompt to mitigate injection risk
6. Use asyncio.get_running_loop() instead of deprecated get_event_loop()
1. Critical: Add missing TaskResult import in plan_exec_engine.py
2. Critical: Fix ReWOOEngine param name (max_steps → max_plan_steps)
3. Major: Remove duplicate token counting in reflexion.py
4. Major: LLM audit failure now passes (trusts rule check) instead of failing
5. Major: Fix dict iteration with del using list() copy in lifecycle.py
6. Major: Fix Chinese content tokenization using regex split instead of space split
7. Minor: _is_positive_mention now checks all occurrences, not just the first
Phase B:
- U1: CostAwareRouter with 3-layer routing (rule/LLM/capability matching)
- U6: OrganizationContext with agent profiles and capability-based discovery
- U7: AlignmentGuard with constraint injection and cascade detection
Phase C:
- U8: Soul dynamic evolution with version tracking and reflection-triggered updates
- U9: Auction mechanism as optional advanced routing mode
- U10: Server integration + end-to-end integration tests
250 new tests passing across all units.
- Shell whitelist: use exact binary match instead of startswith
- Shell audit log: use deque(maxlen=10000) to cap memory
- Terminal history: use deque(maxlen) for O(1) eviction
- Path optimizer: cap _pending_paths at 50 entries per task_type
- Pitfall detector: only add tips to matching steps, not all
- Experience store: handle non-numeric _parse_time_window input
- Extract shared is_safe_url() to utils/security.py (DRY)
- Workflow condition evaluator: handle float() ValueError
- Enhanced chat CLI with adaptive mode and session management
- Added pipeline reflection and schema extensions
- Upgraded BaiduSearch and WebSearch tools with advanced capabilities
- Expanded server routes for skills and chat
- Added session store enhancements
- New chat module and pipeline reflection support
- Orchestrator accepts optional message_bus parameter; workers publish
task.progress messages via MessageBus after each subtask execution
- AgentPool accepts optional message_bus; auto-registers agents on
create and auto-unregisters on remove
- app.py initializes MessageBus from config and injects into AgentPool
- ServerConfig adds bus configuration field
- 5 new tests, all passing
- AgentMessage: message model with sender/recipient/topic/payload/correlation_id
- MessageBus Protocol: publish/subscribe/unsubscribe/request/broadcast/health_check
- InMemoryMessageBus: asyncio.Queue-based implementation for testing
- RedisMessageBus: Redis Streams (XADD/XREADGROUP) implementation with
consumer groups, message acknowledgment, and dead letter queue
- create_message_bus() factory with graceful Redis→InMemory fallback
- Request-response pattern via correlation_id + asyncio.Future
- 13 new tests, all passing
- AskHumanTool: Human-in-the-Loop tool for Chat mode, pushes questions
via WebSocket callback and waits for user reply via asyncio.Future
- Token streaming: execute_stream() now uses chat_stream() instead of
chat(), yielding token-type ReActEvents for each StreamChunk
- _build_response_from_stream() static method constructs LLMResponse
from accumulated stream data
- Export AskHumanTool from tools/__init__.py
- 12 new tests (7 AskHumanTool + 5 token streaming), all passing
Add HeadroomRetrieveTool that allows LLM to retrieve original
uncompressed data from CCR cache via Function Calling. Auto-registered
when HeadroomCompressor is active and available.
chiguyong