The Claude Code CLI v2.1.45+ emits a `rate_limit_event` message type that
the Python SDK v0.1.19 cannot parse, raising MessageParseError. Two bugs
resulted:
1. **False-positive rate limit**: check_rate_limit_error() matched
"rate_limit" in the exception string "Unknown message type:
rate_limit_event" via both an explicit type check and a regex fallback,
triggering 15-19s backoff + query re-send on every session.
2. **One-message-behind**: The MessageParseError killed the
receive_response() async generator, but the CLI subprocess was still
alive with buffered response data. Catching and returning meant the
response was never consumed. The next send_message() would read the
previous response first, creating a one-behind offset.
Changes:
- chat_constants.py: check_rate_limit_error() now returns (False, None)
for any MessageParseError, blocking both false-positive paths. Added
safe_receive_response() helper that retries receive_response() on
MessageParseError — the SDK's decoupled producer/consumer architecture
(anyio memory channel) allows the new generator to continue reading
remaining messages without data loss. Removed calculate_rate_limit_backoff
re-export and MAX_CHAT_RATE_LIMIT_RETRIES constant.
- spec_chat_session.py, assistant_chat_session.py, expand_chat_session.py:
Replaced retry-with-backoff loops with safe_receive_response() wrapper.
Removed asyncio.sleep backoff, query re-send, and rate_limited yield.
Cleaned up unused imports (asyncio, calculate_rate_limit_backoff,
MAX_CHAT_RATE_LIMIT_RETRIES).
- agent.py: Added inner retry loop around receive_response() with same
MessageParseError skip-and-restart pattern. Removed early-return that
truncated responses.
- types.ts: Removed SpecChatRateLimitedMessage,
AssistantChatRateLimitedMessage, and their union entries.
- useSpecChat.ts, useAssistantChat.ts, useExpandChat.ts: Removed dead
'rate_limited' case handlers.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Major changes across 21 files (755 additions, 196 deletions):
Browser Automation Migration:
- Add versioned project migration system (prompts.py) with content-based
detection and section-level regex replacement for coding/testing prompts
- Migrate STEP 5 (browser verification) and BROWSER AUTOMATION sections
in coding prompt template to use playwright-cli commands
- Migrate STEP 2 and AVAILABLE TOOLS sections in testing prompt template
- Migration auto-runs at agent startup (autonomous_agent_demo.py), copies
playwright-cli skill, scaffolds .playwright/cli.config.json, updates
.gitignore, and stamps .migration_version file
- Add playwright-cli command validation to security allowlist (security.py)
with tests for allowed subcommands and blocked eval/run-code
Headless Browser Setting Fix:
- Add _apply_playwright_headless() to process_manager.py that reads/updates
.playwright/cli.config.json before agent subprocess launch
- Remove dead PLAYWRIGHT_HEADLESS env var that was never consumed
- Settings UI toggle now correctly controls visible browser window
Playwright CLI Auto-Install:
- Add ensurePlaywrightCli() to lib/cli.js for npm global entry point
- Add playwright-cli detection + npm install to start.bat, start.sh,
start_ui.bat, start_ui.sh for all startup paths
Other Improvements:
- Add project folder path tooltip to ProjectSelector.tsx dropdown items
- Remove legacy Playwright MCP server configuration from client.py
- Update CLAUDE.md with playwright-cli skill documentation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Agents can now request structured human input when they encounter
genuine blockers (API keys, design choices, external configs). The
request is displayed in the UI with a dynamic form, and the human's
response is stored and made available when the agent resumes.
Changes span 21 files + 1 new component:
- Database: 3 new columns (needs_human_input, human_input_request,
human_input_response) with migration
- MCP: new feature_request_human_input tool + guards on existing tools
- API: new resolve-human-input endpoint, 4th feature bucket
- Orchestrator: skip needs_human_input features in scheduling
- Progress: 4-tuple return from count_passing_tests
- WebSocket: needs_human_input count in progress messages
- UI: conditional 4th Kanban column, HumanInputForm component,
amber status indicators, dependency graph support
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Enable the orchestrator to assign 1-3 features per coding agent subprocess,
selected via dependency chain extension + same-category fill. This reduces
cold-start overhead and leverages shared context across related features.
Orchestrator (parallel_orchestrator.py):
- Add batch tracking: _batch_features and _feature_to_primary data structures
- Add build_feature_batches() with dependency chain + category fill algorithm
- Add start_feature_batch() and _spawn_coding_agent_batch() methods
- Update _on_agent_complete() for batch cleanup across all features
- Update stop_feature() with _feature_to_primary lookup
- Update get_ready_features() to exclude all batch feature IDs
- Update main loop to build batches then spawn per available slot
CLI and agent layer:
- Add --feature-ids (comma-separated) and --batch-size CLI args
- Add feature_ids parameter to run_autonomous_agent() with batch prompt selection
- Add get_batch_feature_prompt() with sequential workflow instructions
WebSocket layer (server/websocket.py):
- Add BATCH_CODING_AGENT_START_PATTERN and BATCH_FEATURES_COMPLETE_PATTERN
- Add _handle_batch_agent_start() and _handle_batch_agent_complete() methods
- Add featureIds field to all agent_update messages
- Track current_feature_id updates as agent moves through batch
Frontend (React UI):
- Add featureIds to ActiveAgent and WSAgentUpdateMessage types
- Update KanbanColumn and DependencyGraph agent-feature maps for batch
- Update AgentCard to show "Batch: #X, #Y, #Z" with active feature highlight
- Add "Features per Agent" segmented control (1-3) in SettingsModal
Settings integration (full stack):
- Add batch_size to schemas, settings router, agent router, process manager
- Default batch_size=3, user-configurable 1-3 via settings UI
- batch_size=1 is functionally identical to pre-batching behavior
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Narrow `\boverloaded\b` regex to require server/api/system context,
preventing false positives when Claude discusses method/operator
overloading in OOP code (C++, Java, C#, etc.)
- Restore 24-hour cap for absolute reset-time delays instead of 1-hour
clamp, avoiding unnecessary retry loops when rate limits reset hours
in the future
- Add test for Retry-After: 0 returning 0 (regression lock for the
`is not None` fix)
- Add false positive tests for "overloaded" in programming context
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Use the shared clamp_retry_delay() function (1-hour cap) for parsed
reset-time delays instead of a separate 24-hour cap. This aligns with
the PR's consistent 1-hour maximum delay objective.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
1. agent.py: Reset opposite retry counter when entering rate_limit or error
status to prevent mixed events from inflating delays
2. rate_limit_utils.py: Fix parse_retry_after() regex to reject minute/hour
units - patterns now require explicit "seconds"/"s" unit or end of string
3. test_rate_limit_utils.py: Add tests for "retry after 5 minutes" and other
minute/hour variants to ensure they return None
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Create rate_limit_utils.py with shared constants and functions
- Update agent.py to import from shared module
- Update test_agent.py to import from shared module (removes duplication)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Fix comment: "exponential" -> "linear" for error backoff (30 * retries)
- Fix rate limit counter reset: only reset when no rate limit signal detected
- Apply exponential backoff to rate limit in response text (not just exceptions)
- Use explicit `is not None` check for retry_seconds to handle Retry-After: 0
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When Claude API hits rate limits via HTTP 429 exceptions (rather than
response text), the agent now properly detects and handles them:
- Add RATE_LIMIT_PATTERNS constant for comprehensive detection
- Add parse_retry_after() to extract wait times from error messages
- Add is_rate_limit_error() helper for pattern matching
- Return new "rate_limit" status from exception handler
- Implement exponential backoff: 60s → 120s → 240s... (max 1 hour)
- Improve generic error backoff: 30s → 60s → 90s... (max 5 minutes)
- Expand text-based detection patterns in response handling
- Add unit tests for new functions
Fixes#41
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace agent-initiated feature selection with orchestrator pre-selection
for both coding and testing agents. This ensures Mission Control displays
correct feature numbers for testing agents (previously showed "Feature #0").
Key changes:
MCP Server (mcp_server/feature_mcp.py):
- Add feature_get_by_id tool for agents to fetch assigned feature details
- Remove obsolete tools: feature_get_next, feature_claim_next,
feature_claim_for_testing, feature_get_for_regression
- Remove helper functions and unused imports (text, OperationalError, func)
Orchestrator (parallel_orchestrator.py):
- Change running_testing_agents from list to dict[int, Popen]
- Add claim_feature_for_testing() with random selection
- Add release_testing_claim() method
- Pass --testing-feature-id to spawned testing agents
- Use unified [Feature #X] output format for both agent types
Agent Entry Points:
- autonomous_agent_demo.py: Add --testing-feature-id CLI argument
- agent.py: Pass testing_feature_id to get_testing_prompt()
Prompt Templates:
- coding_prompt.template.md: Update to use feature_get_by_id
- testing_prompt.template.md: Update workflow for pre-assigned features
- prompts.py: Update pre-claimed headers for both agent types
WebSocket (server/websocket.py):
- Simplify tracking with unified [Feature #X] pattern
- Remove testing-specific parsing code
Assistant (server/services/assistant_chat_session.py):
- Update help text with current available tools
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Changes:
- Revert incorrect import from claude_code_sdk to claude_agent_sdk in agent.py
(PR #50 introduced an undocumented change to a deprecated package)
- Clear activeAgents and recentActivity in useWebSocket when agent stops
to prevent stale UI state
The claude_code_sdk package is deprecated (last updated Sep 2025) while
claude_agent_sdk is the active, maintained package. The import change in
PR #50 was undocumented and would have caused ImportError since only
claude-agent-sdk is specified in requirements.txt.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This fix addresses two root causes that caused testing agents to
accumulate (10-12 agents) instead of maintaining a 1:1 ratio with
coding agents:
1. Testing agents now exit after one session (agent.py)
- Added `or agent_type == "testing"` to the exit condition
- Previously, testing agents never hit the exit condition since
they're spawned with feature_id=None
2. Testing agents now spawn when coding agents START, not complete
- Moved spawn logic from _on_agent_complete() to start_feature()
- Removed the old spawn logic from _on_agent_complete()
- This ensures proper 1:1 ratio and prevents accumulation
Expected behavior after fix:
- First coding agent: no testing agent (no passing features yet)
- Subsequent coding agents: one testing agent spawns per start
- Each testing agent tests ONE feature then terminates immediately
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The PR #77 introduced a bug where `is_first_run` was used in the
completion detection check, but this variable is only defined when
`agent_type is None`. When the orchestrator runs agents with explicit
`--agent-type` or `--feature-id`, the variable is undefined causing
a NameError crash.
Changed to use `is_initializer` which is always defined and has the
correct semantic meaning for this check.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Introduce a new testing agent architecture that runs regression tests
independently from coding agents, improving quality assurance in
parallel mode.
Key changes:
Testing Agent System:
- Add testing_prompt.template.md for dedicated testing agent role
- Add feature_mark_failing MCP tool for regression detection
- Add --agent-type flag to select initializer/coding/testing mode
- Remove regression testing from coding prompt (now handled by testing agents)
Parallel Orchestrator Enhancements:
- Add testing agent spawning with configurable ratio (--testing-agent-ratio)
- Add comprehensive debug logging system (DebugLog class)
- Improve database session management to prevent stale reads
- Add engine.dispose() calls to refresh connections after subprocess commits
- Fix f-string linting issues (remove unnecessary f-prefixes)
UI Improvements:
- Add testing agent mascot (Chip) to AgentAvatar
- Enhance AgentCard to display testing agent status
- Add testing agent ratio slider in SettingsModal
- Update WebSocket handling for testing agent updates
- Improve ActivityFeed to show testing agent activity
API & Server Updates:
- Add testing_agent_ratio to settings schema and endpoints
- Update process manager to support testing agent type
- Enhance WebSocket messages for agent_update events
Template Changes:
- Delete coding_prompt_yolo.template.md (consolidated into main prompt)
- Update initializer_prompt.template.md with improved structure
- Streamline coding_prompt.template.md workflow
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The agent was running in an infinite loop when all kanban features were
completed. This happened because:
1. The main loop in agent.py had no completion detection
2. The coding prompt instructs Claude to run regression tests BEFORE
checking for new features
3. feature_get_next() returns "All features passing!" but nothing acted on it
This fix adds three completion checks:
1. Pre-loop check: Exits immediately if project is already 100% complete
when the agent starts (avoids running unnecessary sessions)
2. Post-session check: After each session, checks if all features are now
passing and exits gracefully with a success message
3. Single-feature mode: Exits after one session since the parallel
orchestrator manages spawning new agents for other features
Tested with a project that had 240/240 features passing - agent now exits
immediately with "ALL FEATURES ALREADY COMPLETE!" message.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Changes:
- Add per-agent log viewer with copy-to-clipboard functionality
- New AgentLogEntry type for structured log entries
- Logs stored per-agent in WebSocket state (up to 500 entries)
- Log modal rendered via React Portal to avoid overflow issues
- Click log icon on agent card to view full activity history
- Fix agents getting stuck in "failed" state
- Wrap client context manager in try/except (agent.py)
- Remove failed agents from UI on error state (useWebSocket.ts)
- Handle permanently failed features in get_all_complete()
- Add friendlier agent state labels
- "Hit an issue" → "Trying plan B..."
- "Retrying..." → "Being persistent..."
- Softer colors (yellow/orange instead of red)
- Add scheduling scores for smarter feature ordering
- compute_scheduling_scores() in dependency_resolver.py
- Features that unblock others get higher priority
- Update CLAUDE.md with parallel mode documentation
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Major feature implementation for parallel agent execution with dependency-aware
scheduling and an engaging multi-agent UI experience.
Backend Changes:
- Add parallel_orchestrator.py for concurrent feature processing
- Add api/dependency_resolver.py with cycle detection (Kahn's algorithm + DFS)
- Add atomic feature_claim_next() with retry limit and exponential backoff
- Fix circular dependency check arguments in 4 locations
- Add AgentTracker class for parsing agent output and emitting updates
- Add browser isolation with --isolated flag for Playwright MCP
- Extend WebSocket protocol with agent_update messages and log attribution
- Add WSAgentUpdateMessage schema with agent states and mascot names
- Fix WSProgressMessage to include in_progress field
New UI Components:
- AgentMissionControl: Dashboard showing active agents with collapsible activity
- AgentCard: Individual agent status with avatar and thought bubble
- AgentAvatar: SVG mascots (Spark, Fizz, Octo, Hoot, Buzz) with animations
- ActivityFeed: Recent activity stream with stable keys (no flickering)
- CelebrationOverlay: Confetti animation with click/Escape dismiss
- DependencyGraph: Interactive node graph visualization with dagre layout
- DependencyBadge: Visual indicator for feature dependencies
- ViewToggle: Switch between Kanban and Graph views
- KeyboardShortcutsHelp: Help overlay accessible via ? key
UI/UX Improvements:
- Celebration queue system to handle rapid success messages
- Accessibility attributes on AgentAvatar (role, aria-label, aria-live)
- Collapsible Recent Activity section with persisted preference
- Agent count display in header
- Keyboard shortcut G to toggle Kanban/Graph view
- Real-time thought bubbles and state animations
Bug Fixes:
- Fix circular dependency validation (swapped source/target arguments)
- Add MAX_CLAIM_RETRIES=10 to prevent stack overflow under contention
- Fix THOUGHT_PATTERNS to match actual [Tool: name] format
- Fix ActivityFeed key prop to prevent re-renders on new items
- Add featureId/agentIndex to log messages for proper attribution
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Previously, the autonomous agent would continue running indefinitely
even after all features passed verification. The agent would enter a
verification loop, repeatedly checking 'All features are passing!'
without ever exiting.
This fix detects the completion message from feature_get_next() and
gracefully exits the main loop with a victory banner, preventing
unnecessary API calls and resource consumption.
Fixes infinite loop when project reaches 100% completion.
Applied Fixes
More flexible string matching: Changed from response.lower().strip().startswith("limit reached") to "limit reached" in response.lower() to handle cases where the message has prefix text or variations in whitespace.
Improved regex pattern: Updated to r"(?i)\bresets(?:\s+at)?\s+(\d+)(?::(\d+))?\s*(am|pm)\s*\(([^)]+)\)" which now handles:
Optional "at" after "resets" (e.g., "resets at 3pm" or "resets 3pm")
Flexible whitespace around components
Word boundaries to prevent partial matches
Timezone sanitization: Added .strip() to tz_name = match.group(4).strip() to remove any leading/trailing whitespace that could cause ZoneInfo() to fail.
Safety clamp: Added delay_seconds = min(delta.total_seconds(), 24 * 60 * 60) to ensure the delay never exceeds 24 hours, preventing the agent from being stuck waiting for extremely long periods.
- Add CI workflow with Python (ruff lint, security tests) and UI (ESLint, TypeScript, build) jobs
- Add ruff, mypy, pytest to requirements.txt
- Add pyproject.toml with ruff configuration
- Fix import sorting across Python files (ruff --fix)
- Fix test_security.py expectations to match actual security policy
- Remove invalid 'eof' command from ALLOWED_COMMANDS
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add a new YOLO (You Only Live Once) mode that skips all browser testing
and regression tests for faster feature iteration during prototyping.
Changes made:
**Core YOLO Mode Implementation:**
- Add --yolo CLI flag to autonomous_agent_demo.py
- Update agent.py to accept yolo_mode parameter and select appropriate prompt
- Modify client.py to conditionally include Playwright MCP server (excluded in YOLO mode)
- Add coding_prompt_yolo.template.md with static analysis only verification
- Add get_coding_prompt_yolo() to prompts.py
**Server/API Updates:**
- Add AgentStartRequest schema with yolo_mode field
- Update AgentStatus to include yolo_mode
- Modify process_manager.py to pass --yolo flag to subprocess
- Update agent router to accept yolo_mode in start request
**UI Updates:**
- Add YOLO toggle button (lightning bolt icon) in AgentControl
- Show YOLO mode indicator when agent is running in YOLO mode
- Add useAgentStatus hook to track current mode
- Update startAgent API to accept yoloMode parameter
- Add YOLO toggle in SpecCreationChat completion flow
**Spec Creation Improvements:**
- Fix create-spec.md to properly replace [FEATURE_COUNT] placeholder
- Add REQUIRED FEATURE COUNT section to initializer_prompt.template.md
- Fix spec_chat_session.py to create security settings file for Claude SDK
- Delete app_spec.txt before spec creation to allow fresh creation
**Documentation:**
- Add YOLO mode section to CLAUDE.md with usage examples
- Add checkpoint.md slash command for creating detailed commits
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
