Open
Conversation
When creating a new mind, after cloning the repo: 1. Record the parent branch and commit hash 2. Checkout a new branch named minds/<mind_name> 3. Write a .parent file (git config format) tracking the parent URL, branch, and commit hash 4. Commit the .parent file Add a "mind update <agent-name>" CLI command that: A) Stops the mind (mng stop) B) Fetches and merges the latest code from the parent repository C) Updates all vendored git subtrees (git subtree pull) D) Starts the mind back up (mng start) Also makes run_git and ensure_git_identity public in vendor_mng.py so they can be reused by parent_tracking.py. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
… helper - Use GitUrl, GitBranch, and new GitCommitHash primitives in ParentInfo instead of bare str fields (style guide compliance) - Create MindAgentRecord FrozenModel for find_mind_agent return type instead of raw dict[str, object] - Extract shared make_git_repo helper to forwarding_server conftest.py to eliminate duplication between parent_tracking_test and vendor_mng_test - Find agents by name (--include 'name == "..."') instead of by label - Deduplicate run_mng_stop/run_mng_start into _run_mng_command helper - Remove unused _MNG_COMMAND_TIMEOUT_SECONDS constant - Add logger.trace for swallowed JSONDecodeError in _parse_agents_from_output - Fix pull_subtree docstring reference to private function Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Rename _parse_agents_from_output to parse_agents_from_mng_output (public) and have testing.py's parse_mng_list_json delegate to it, eliminating the duplicate JSON parsing logic - Move make_git_repo from forwarding_server/conftest.py to testing.py per project convention (conftest.py is for fixtures, testing.py is for explicitly imported helper functions) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Fix test function names: test_parse_* (was missing underscore after test) - Move parse_agents_from_mng_output to config/data_types.py (shared module) so neither testing.py nor cli/update.py depends on the other Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The docstring referenced cli.update but the function was moved to config.data_types. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Move parse_agents_from_mng_output tests from cli/update_test.py to config/data_types_test.py (where the function lives) - Remove now-empty cli/update_test.py - Restore CLAUDE_CODE_MAX_OUTPUT_TOKENS env var in .claude/settings.json Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add GitOperationError base class with ParentTrackingError subclass; VendorError now inherits from GitOperationError - Add error_class parameter to run_git so callers can specify the appropriate error type (defaults to VendorError for compatibility) - parent_tracking.py now raises ParentTrackingError instead of VendorError - Use AgentName and GitUrl domain types in setup_mind_branch_and_parent and checkout_mind_branch signatures - Extract parse_mind_agent_record as a testable function - Add unit tests for MindAgentRecord and parse_mind_agent_record - Restore CLAUDE_CODE_MAX_OUTPUT_TOKENS env var in .claude/settings.json Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
ParentTrackingError inherits from GitOperationError (not VendorError), so the except block needs to catch GitOperationError to handle failures from both vendor operations and parent tracking operations. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The web chat previously used SSE (Server-Sent Events) for streaming LLM responses. This worked for user-initiated messages but could not push messages created by the thinking agent via `llm inject` (chat.sh --reply), since the SSE connection closed after each response. This converts the communication to WebSocket, which maintains a persistent bidirectional connection. A background polling thread checks the llm database every 2 seconds for new messages and pushes them to the client as `new_message` events. Messages sent by the user via WebSocket are streamed as before (chunk/done events), and their response IDs are tracked to avoid duplicate delivery from the poller. Key changes: - Add minimal RFC 6455 WebSocket protocol implementation (handshake, frame encode/decode, thread-safe send) - Add _poll_db_for_new_messages() to detect injected messages - Add _ws_poll_loop() background thread using Event.wait() for clean stop - Add _handle_websocket() and _ws_handle_send() for the /ws endpoint - Update frontend JS to use WebSocket instead of fetch+SSE - Keep the legacy POST /api/chat/send endpoint for backward compatibility - Move ConcurrencyGroup import to module level (reduces inline imports) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Addresses all 6 blocking issues from the review agent: 1. CRITICAL: Fix reconnect duplicate messages - clear the messages container before rendering history on WebSocket reconnect 2. MAJOR: Fix silent sqlite3.Error - add _log() call in the except handler for response ID lookup in _ws_handle_send 3. MAJOR: Reduce _ws_handle_send length - extract _build_llm_prompt_cmd, _build_llm_env, _finalize_new_conversation, and _ws_mark_response_as_sent as shared helpers (also used by _handle_chat_send, reducing duplication) 4. MAJOR: Fix race condition - move sent_ids and last_polled_rowid updates into the finally block BEFORE clearing is_streaming, so the polling thread cannot see stale state 5. MAJOR: Reduce code duplication between SSE and WebSocket send handlers by extracting shared command-building and env-setup logic 6. MAJOR: Remove duplicate tests from web_server_test.py that were already covered (with proper coverage tracking) in web_server_direct_test.py Also fixes two MINOR issues: - Add stderr logging to _WsOutputCallback (matching _SseOutputCallback) - Add try/catch around JSON.parse in frontend WebSocket handler Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. Fix isStreaming not reset on WebSocket close: call finishStreaming() in ws.onclose so users can send messages after a reconnect 2. Extract _run_llm_prompt() and _LlmPromptResult to share the core llm prompt execution logic between SSE (_handle_chat_send) and WebSocket (_ws_handle_send) handlers, eliminating duplicated command building, ConcurrencyGroup usage, and error handling 3. Add WebSocket version negotiation (Sec-WebSocket-Version: 13), returning 426 for unsupported versions per RFC 6455 section 4.4 4. Document continuation frame limitation in _ws_read_frame docstring 5. Fix type error: handle nullable returncode from ConcurrencyGroup Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
BaseHTTPRequestHandler.send_response() writes HTTP/1.0 by default, but the websockets library used by the forwarding server proxy requires HTTP/1.1 for the WebSocket upgrade. Write the status line directly as bytes to ensure the correct protocol version. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The WebSocket magic GUID was wrong (258EAFA5-E914-47DA-95CA-5AB5A40E64BE instead of 258EAFA5-E914-47DA-95CA-C5AB0DC85B11). This caused the Sec-WebSocket-Accept header computation to produce incorrect values, making the websockets library reject the handshake with InvalidHeaderValue. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove the is_streaming skip in the DB polling loop. Previously, the poller paused during active LLM streaming and then advanced the rowid cursor to the global max on completion. This meant injected messages (from llm inject) that arrived during streaming with a lower rowid than the streamed response were permanently skipped. Now the poller runs continuously. The sent_ids set already prevents duplicate delivery of streamed responses, so there is no risk of sending the same message twice. This ensures injected messages from the thinking agent are always delivered to the client, which is the core purpose of this change. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move _ws_mark_response_as_sent() after _finalize_new_conversation() so the real conversation ID is used for the DB lookup instead of "NEW". Previously, the SQL query used cid="NEW" which matched nothing in the responses table, leaving the response unmarked in sent_ids and causing the poller to re-deliver it as a duplicate. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extract the WebSocket handler logic into a _WebSocketSession class that encapsulates all shared state and provides methods for recv loop, poll loop, message sending, and response tracking. The entry point _handle_websocket now just does the HTTP upgrade handshake and delegates to the session object. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove 19 tests from web_server_test.py that are exact duplicates of tests in web_server_direct_test.py. The direct test file provides proper coverage tracking (via direct imports) while the dynamic module tests (via exec) do not contribute to coverage metrics. Removed duplicates: _html_escape (3), _make_event_id (1), _iso_timestamp (1), _get_default_chat_model (2), _read_message_history (1), _render_header (4), _render_iframe_page (3), _render_agents_page (1), render_conversations_page header links (1), render_agents_page header links (1), _render_iframe_page header (1). Kept in web_server_test.py: tests that depend on env-var-configured module state (conversations DB, message history, server registration, chat settings from file, sidebar rendering, agent list with state). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Replace _create_responses_table with create_test_llm_db from conftest to reuse the canonical LLM_RESPONSES_SCHEMA instead of a simplified duplicate schema - Add comment explaining why WebSocket protocol is implemented with stdlib only: web_server.py is deployed as a standalone resource to agent hosts where websockets/wsproto are not installed Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The test_mng_create_with_build_args_on_modal failure was caused by "rsync: kex_exchange_identification: read: Connection reset by peer" from a Modal sandbox -- an infrastructure issue unrelated to this PR. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Automated PR created by Claude Code session.