fix: handle rate_limit_event crash in chat sessions

.claude/commands/review-pr.md

@@ -55,10 +55,10 @@ Pull request(s): $ARGUMENTS
    - Reviewing large, unfocused PRs is impractical and error-prone; the review cannot provide adequate assurance for such changes
 
 6. **Vision Alignment Check**
-   - Read the project's README.md and CLAUDE.md to understand the application's core purpose
-   - Assess whether this PR aligns with the application's intended functionality
-   - If the changes deviate significantly from the core vision or add functionality that doesn't serve the application's purpose, note this in the review
-   - This is not a blocker, but should be flagged for the reviewer's consideration
+   - **VISION.md protection**: First, check whether the PR diff modifies `VISION.md` in any way (edits, deletions, renames). If it does, **stop the review immediately** — verdict is **DON'T MERGE**. VISION.md is immutable and no PR is permitted to alter it. Explain this to the user and skip all remaining steps.
+   - Read the project's `VISION.md`, `README.md`, and `CLAUDE.md` to understand the application's core purpose and mandatory architectural constraints
+   - Assess whether this PR aligns with the vision defined in `VISION.md`
+   - **Vision deviation is a merge blocker.** If the PR introduces functionality, integrations, or architectural changes that conflict with `VISION.md`, the verdict must be **DON'T MERGE**. This is not negotiable — the vision document takes precedence over any PR rationale.
 
 7. **Safety Assessment**
    - Provide a review on whether the PR is safe to merge as-is

.claude/launch.json

@@ -0,0 +1,18 @@
+{
+  "version": "0.0.1",
+  "configurations": [
+    {
+      "name": "backend",
+      "runtimeExecutable": "python",
+      "runtimeArgs": ["-m", "uvicorn", "server.main:app", "--host", "127.0.0.1", "--port", "8888", "--reload"],
+      "port": 8888
+    },
+    {
+      "name": "frontend",
+      "runtimeExecutable": "cmd",
+      "runtimeArgs": ["/c", "cd ui && npx vite"],
+      "port": 5173
+    }
+  ],
+  "autoVerify": true
+}

VISION.md

@@ -0,0 +1,22 @@
+# VISION
+
+This document defines the mandatory project vision for AutoForge. All contributions must align with these principles. PRs that deviate from this vision will be rejected. This file itself is immutable via PR — any PR that modifies VISION.md will be rejected outright.
+
+## Claude Agent SDK Exclusivity
+
+AutoForge is a wrapper around the **Claude Agent SDK**. This is a foundational architectural decision, not a preference.
+
+**What this means:**
+
+- AutoForge only supports providers, models, and integrations that work through the Claude Agent SDK.
+- We will not integrate with, accommodate, or add support for other AI SDKs, CLIs, or coding agent platforms (e.g., Codex, OpenCode, Aider, Continue, Cursor agents, or similar tools).
+
+**Why:**
+
+Each platform has its own approach to MCP tools, skills, context management, and feature integration. Attempting to support multiple agent frameworks creates an unsustainable maintenance burden and dilutes the quality of the core experience. By committing to the Claude Agent SDK exclusively, we can build deep, reliable integration rather than shallow compatibility across many targets.
+
+**In practice:**
+
+- PRs adding support for non-Claude agent frameworks will be rejected.
+- PRs introducing abstractions designed to make AutoForge "agent-agnostic" will be rejected.
+- Alternative API providers (e.g., Vertex AI, AWS Bedrock) are acceptable only when accessed through the Claude Agent SDK's own configuration.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "autoforge-ai",
-  "version": "0.1.13",
+  "version": "0.1.14",
   "description": "Autonomous coding agent with web UI - build complete apps with AI",
   "license": "AGPL-3.0",
   "bin": {

requirements-prod.txt

@@ -1,6 +1,6 @@
 # Production runtime dependencies only
 # For development, use requirements.txt (includes ruff, mypy, pytest)
-claude-agent-sdk>=0.1.0,<0.2.0
+claude-agent-sdk>=0.1.39,<0.2.0
 python-dotenv>=1.0.0
 sqlalchemy>=2.0.0
 fastapi>=0.115.0

requirements.txt

@@ -1,4 +1,4 @@
-claude-agent-sdk>=0.1.0,<0.2.0
+claude-agent-sdk>=0.1.39,<0.2.0
 python-dotenv>=1.0.0
 sqlalchemy>=2.0.0
 fastapi>=0.115.0

server/services/assistant_chat_session.py

@@ -7,6 +7,7 @@
 but cannot modify any files.
 """
 
+import asyncio
 import json
 import logging
 import os
@@ -25,7 +26,12 @@
     create_conversation,
     get_messages,
 )
-from .chat_constants import ROOT_DIR
+from .chat_constants import (
+    MAX_CHAT_RATE_LIMIT_RETRIES,
+    ROOT_DIR,
+    calculate_rate_limit_backoff,
+    check_rate_limit_error,
+)
 
 # Load environment variables from .env file if present
 load_dotenv()
@@ -393,39 +399,66 @@ async def _query_claude(self, message: str) -> AsyncGenerator[dict, None]:
 
         full_response = ""
 
-        # Stream the response
-        async for msg in self.client.receive_response():
-            msg_type = type(msg).__name__
-
-            if msg_type == "AssistantMessage" and hasattr(msg, "content"):
-                for block in msg.content:
-                    block_type = type(block).__name__
-
-                    if block_type == "TextBlock" and hasattr(block, "text"):
-                        text = block.text
-                        if text:
-                            full_response += text
-                            yield {"type": "text", "content": text}
-
-                    elif block_type == "ToolUseBlock" and hasattr(block, "name"):
-                        tool_name = block.name
-                        tool_input = getattr(block, "input", {})
+        # Stream the response (with rate-limit retry)
+        for _attempt in range(MAX_CHAT_RATE_LIMIT_RETRIES + 1):
+            try:
+                async for msg in self.client.receive_response():
+                    msg_type = type(msg).__name__
+
+                    if msg_type == "AssistantMessage" and hasattr(msg, "content"):
+                        for block in msg.content:
+                            block_type = type(block).__name__
+
+                            if block_type == "TextBlock" and hasattr(block, "text"):
+                                text = block.text
+                                if text:
+                                    full_response += text
+                                    yield {"type": "text", "content": text}
+
+                            elif block_type == "ToolUseBlock" and hasattr(block, "name"):
+                                tool_name = block.name
+                                tool_input = getattr(block, "input", {})
+
+                                # Intercept ask_user tool calls -> yield as question message
+                                if tool_name == "mcp__features__ask_user":
+                                    questions = tool_input.get("questions", [])
+                                    if questions:
+                                        yield {
+                                            "type": "question",
+                                            "questions": questions,
+                                        }
+                                        continue
 
-                        # Intercept ask_user tool calls -> yield as question message
-                        if tool_name == "mcp__features__ask_user":
-                            questions = tool_input.get("questions", [])
-                            if questions:
                                 yield {
-                                    "type": "question",
-                                    "questions": questions,
+                                    "type": "tool_call",
+                                    "tool": tool_name,
+                                    "input": tool_input,
                                 }
-                                continue
-
-                        yield {
-                            "type": "tool_call",
-                            "tool": tool_name,
-                            "input": tool_input,
-                        }
+                # Completed successfully — break out of retry loop
+                break
+            except Exception as exc:
+                is_rate_limit, retry_secs = check_rate_limit_error(exc)
+                if is_rate_limit and _attempt < MAX_CHAT_RATE_LIMIT_RETRIES:
+                    delay = retry_secs if retry_secs else calculate_rate_limit_backoff(_attempt)
+                    logger.warning(f"Rate limited (attempt {_attempt + 1}/{MAX_CHAT_RATE_LIMIT_RETRIES}), retrying in {delay}s")
+                    yield {
+                        "type": "rate_limited",
+                        "retry_in": delay,
+                        "attempt": _attempt + 1,
+                        "max_attempts": MAX_CHAT_RATE_LIMIT_RETRIES,
+                    }
+                    await asyncio.sleep(delay)
+                    await self.client.query(message)
+                    continue
+                if is_rate_limit:
+                    logger.error("Rate limit retries exhausted for assistant chat")
+                    yield {"type": "error", "content": "Rate limited. Please try again later."}
+                    return
+                # Non-rate-limit MessageParseError: log and break (don't crash)
+                if type(exc).__name__ == "MessageParseError":
+                    logger.warning(f"Ignoring unrecognized message from Claude CLI: {exc}")
+                    break
+                raise
 
         # Store the complete response in the database
         if full_response and self.conversation_id:

server/services/chat_constants.py

@@ -9,6 +9,7 @@
 imports (``from .chat_constants import API_ENV_VARS``) continue to work.
 """
 
+import logging
 import sys
 from pathlib import Path
 from typing import AsyncGenerator
@@ -32,6 +33,45 @@
 # imports continue to work unchanged.
 # -------------------------------------------------------------------
 from env_constants import API_ENV_VARS  # noqa: E402, F401
+from rate_limit_utils import calculate_rate_limit_backoff, is_rate_limit_error, parse_retry_after  # noqa: E402, F401
+
+logger = logging.getLogger(__name__)
+
+# -------------------------------------------------------------------
+# Rate-limit handling for chat sessions
+# -------------------------------------------------------------------
+MAX_CHAT_RATE_LIMIT_RETRIES = 3
+
+
+def check_rate_limit_error(exc: Exception) -> tuple[bool, int | None]:
+    """Inspect an exception and determine if it represents a rate-limit.
+
+    Returns ``(is_rate_limit, retry_seconds)``.  ``retry_seconds`` is the
+    parsed Retry-After value when available, otherwise ``None`` (caller
+    should use exponential backoff).
+
+    Handles:
+    - ``MessageParseError`` whose raw *data* dict has
+      ``type == "rate_limit_event"`` (Claude CLI sends this).
+    - Any exception whose string representation matches known rate-limit
+      patterns (via ``rate_limit_utils.is_rate_limit_error``).
+    """
+    exc_str = str(exc)
+
+    # Check for MessageParseError with a rate_limit_event payload
+    cls_name = type(exc).__name__
+    if cls_name == "MessageParseError":
+        raw_data = getattr(exc, "data", None)
+        if isinstance(raw_data, dict) and raw_data.get("type") == "rate_limit_event":
+            retry = parse_retry_after(str(raw_data)) if raw_data else None
+            return True, retry
+
+    # Fallback: match error text against known rate-limit patterns
+    if is_rate_limit_error(exc_str):
+        retry = parse_retry_after(exc_str)
+        return True, retry
+
+    return False, None
 
 
 async def make_multimodal_message(content_blocks: list[dict]) -> AsyncGenerator[dict, None]:

server/services/expand_chat_session.py

@@ -22,7 +22,13 @@
 from dotenv import load_dotenv
 
 from ..schemas import ImageAttachment
-from .chat_constants import ROOT_DIR, make_multimodal_message
+from .chat_constants import (
+    MAX_CHAT_RATE_LIMIT_RETRIES,
+    ROOT_DIR,
+    calculate_rate_limit_backoff,
+    check_rate_limit_error,
+    make_multimodal_message,
+)
 
 # Load environment variables from .env file if present
 load_dotenv()
@@ -298,24 +304,67 @@ async def _query_claude(
         else:
             await self.client.query(message)
 
-        # Stream the response
-        async for msg in self.client.receive_response():
-            msg_type = type(msg).__name__
-
-            if msg_type == "AssistantMessage" and hasattr(msg, "content"):
-                for block in msg.content:
-                    block_type = type(block).__name__
-
-                    if block_type == "TextBlock" and hasattr(block, "text"):
-                        text = block.text
-                        if text:
-                            yield {"type": "text", "content": text}
-
-                            self.messages.append({
-                                "role": "assistant",
-                                "content": text,
-                                "timestamp": datetime.now().isoformat()
+        # Stream the response (with rate-limit retry)
+        for _attempt in range(MAX_CHAT_RATE_LIMIT_RETRIES + 1):
+            try:
+                async for msg in self.client.receive_response():
+                    msg_type = type(msg).__name__
+
+                    if msg_type == "AssistantMessage" and hasattr(msg, "content"):
+                        for block in msg.content:
+                            block_type = type(block).__name__
+
+                            if block_type == "TextBlock" and hasattr(block, "text"):
+                                text = block.text
+                                if text:
+                                    yield {"type": "text", "content": text}
+
+                                    self.messages.append({
+                                        "role": "assistant",
+                                        "content": text,
+                                        "timestamp": datetime.now().isoformat()
+                                    })
+                # Completed successfully — break out of retry loop
+                break
+            except Exception as exc:
+                is_rate_limit, retry_secs = check_rate_limit_error(exc)
+                if is_rate_limit and _attempt < MAX_CHAT_RATE_LIMIT_RETRIES:
+                    delay = retry_secs if retry_secs else calculate_rate_limit_backoff(_attempt)
+                    logger.warning(f"Rate limited (attempt {_attempt + 1}/{MAX_CHAT_RATE_LIMIT_RETRIES}), retrying in {delay}s")
+                    yield {
+                        "type": "rate_limited",
+                        "retry_in": delay,
+                        "attempt": _attempt + 1,
+                        "max_attempts": MAX_CHAT_RATE_LIMIT_RETRIES,
+                    }
+                    await asyncio.sleep(delay)
+                    # Re-send the query before retrying receive_response
+                    if attachments and len(attachments) > 0:
+                        content_blocks_retry: list[dict[str, Any]] = []
+                        if message:
+                            content_blocks_retry.append({"type": "text", "text": message})
+                        for att in attachments:
+                            content_blocks_retry.append({
+                                "type": "image",
+                                "source": {
+                                    "type": "base64",
+                                    "media_type": att.mimeType,
+                                    "data": att.base64Data,
+                                }
                             })
+                        await self.client.query(make_multimodal_message(content_blocks_retry))
+                    else:
+                        await self.client.query(message)
+                    continue
+                if is_rate_limit:
+                    logger.error("Rate limit retries exhausted for expand chat")
+                    yield {"type": "error", "content": "Rate limited. Please try again later."}
+                    return
+                # Non-rate-limit MessageParseError: log and break (don't crash)
+                if type(exc).__name__ == "MessageParseError":
+                    logger.warning(f"Ignoring unrecognized message from Claude CLI: {exc}")
+                    break
+                raise
 
     def get_features_created(self) -> int:
         """Get the total number of features created in this session."""