@respan/cli 0.5.0 → 0.5.2

package/dist/assets/hook.py

@@ -81,6 +81,38 @@ def save_state(state: Dict[str, Any]) -> None:
         STATE_FILE.write_text(json.dumps(state, indent=2), encoding="utf-8")
 
 
+# Known config keys in respan.json that map to span fields.
+# Anything else is treated as a custom property (goes into metadata).
+KNOWN_CONFIG_KEYS = {"customer_id", "span_name", "workflow_name"}
+
+
+def load_respan_config(cwd: str) -> Dict[str, Any]:
+    """Load .claude/respan.json from the project directory.
+
+    Returns a dict with two keys:
+      - "fields": known span fields (customer_id, span_name, workflow_name)
+      - "properties": everything else (custom properties → metadata)
+    """
+    config_path = Path(cwd) / ".claude" / "respan.json"
+    if not config_path.exists():
+        return {"fields": {}, "properties": {}}
+    try:
+        raw = json.loads(config_path.read_text(encoding="utf-8"))
+        if not isinstance(raw, dict):
+            return {"fields": {}, "properties": {}}
+        fields = {}
+        properties = {}
+        for k, v in raw.items():
+            if k in KNOWN_CONFIG_KEYS:
+                fields[k] = v
+            else:
+                properties[k] = v
+        return {"fields": fields, "properties": properties}
+    except (json.JSONDecodeError, IOError) as e:
+        debug(f"Failed to load respan.json from {config_path}: {e}")
+        return {"fields": {}, "properties": {}}
+
+
 def get_content(msg: Dict[str, Any]) -> Any:
     """Extract content from a message."""
     if isinstance(msg, dict):
@@ -317,27 +349,33 @@ def create_respan_spans(
     user_msg: Dict[str, Any],
     assistant_msgs: List[Dict[str, Any]],
     tool_results: List[Dict[str, Any]],
+    config: Optional[Dict[str, Any]] = None,
 ) -> List[Dict[str, Any]]:
-    """Create Respan span logs for a single turn with all available metadata."""
+    """Create Respan span logs for a single turn with all available metadata.
+
+    Produces a proper span tree so that the Respan UI renders nested children:
+        Root (agent container)
+          ├── claude.chat  (generation – carries model, tokens, messages)
+          ├── Thinking 1   (generation, if extended thinking is present)
+          ├── Tool: Read   (tool, if tool use occurred)
+          └── Tool: Write  (tool, if tool use occurred)
+    """
     spans = []
-    
-    # Extract user text and timestamp
+
+    # ------------------------------------------------------------------
+    # 1. Extract data from the transcript messages
+    # ------------------------------------------------------------------
     user_text = get_text_content(user_msg)
     user_timestamp = user_msg.get("timestamp")
     user_time = parse_timestamp(user_timestamp) if user_timestamp else None
-    
-    # Extract assistant text from ALL messages in the turn (tool-using turns
-    # have multiple assistant messages: text before tool, then text after).
+
+    # Collect assistant text across all messages in the turn
     final_output = ""
-    first_assistant_msg = None
     if assistant_msgs:
         text_parts = [get_text_content(m) for m in assistant_msgs]
         final_output = "\n".join(p for p in text_parts if p)
-        first_assistant_msg = assistant_msgs[0]
-    
-    # Get model, usage, and timing info from assistant messages.
-    # For tool-using turns there are multiple assistant messages (multiple API
-    # calls), so we aggregate usage and take the *last* timestamp as end time.
+
+    # Aggregate model, usage, timing from (possibly multiple) API calls
     model = "claude"
     usage = None
     request_id = None
@@ -360,7 +398,6 @@ def create_respan_spans(
             last_assistant_timestamp = ts
             last_assistant_time = parse_timestamp(ts)
 
-        # Aggregate usage across all API calls in the turn
         msg_usage = msg_obj.get("usage")
         if msg_usage:
             if usage is None:
@@ -371,187 +408,191 @@ def create_respan_spans(
                             "cache_read_input_tokens"):
                     if key in msg_usage:
                         usage[key] = usage.get(key, 0) + msg_usage[key]
-                # Keep last service_tier
                 if "service_tier" in msg_usage:
                     usage["service_tier"] = msg_usage["service_tier"]
 
-    # Calculate timing
-    start_time_str = user_timestamp or first_assistant_timestamp or datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
-    timestamp_str = last_assistant_timestamp or first_assistant_timestamp or datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
+    # Timing
+    now_str = datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
+    start_time_str = user_timestamp or first_assistant_timestamp or now_str
+    timestamp_str = last_assistant_timestamp or first_assistant_timestamp or now_str
 
-    # Calculate latency from user message to final assistant response
     latency = None
     if user_time and last_assistant_time:
         latency = (last_assistant_time - user_time).total_seconds()
-    
-    # Extract messages for chat span
-    prompt_messages = []
+
+    # Messages
+    prompt_messages: List[Dict[str, Any]] = []
     if user_text:
         prompt_messages.append({"role": "user", "content": user_text})
-    
-    completion_message = None
+    completion_message: Optional[Dict[str, Any]] = None
     if final_output:
         completion_message = {"role": "assistant", "content": final_output}
-    
-    # Create trace ID for this turn
+
+    # IDs — respan.json fields, then env var overrides
+    cfg_fields = (config or {}).get("fields", {})
+    cfg_props = (config or {}).get("properties", {})
+
     trace_unique_id = f"{session_id}_turn_{turn_num}"
-    
-    # Naming: human-readable workflow + span names
-    workflow_name = "claude-code"
-    # Use first ~60 chars of user message as span name for readability
-    user_preview = (user_text[:60] + "...") if user_text and len(user_text) > 60 else (user_text or f"turn_{turn_num}")
-    root_span_name = f"Turn {turn_num}: {user_preview}"
+    workflow_name = os.environ.get("RESPAN_WORKFLOW_NAME") or cfg_fields.get("workflow_name") or "claude-code"
+    root_span_name = os.environ.get("RESPAN_SPAN_NAME") or cfg_fields.get("span_name") or "claude-code"
     thread_id = f"claudecode_{session_id}"
-    
-    # Build metadata with additional info
-    metadata = {
-        "claude_code_turn": turn_num,
-    }
+    customer_id = os.environ.get("RESPAN_CUSTOMER_ID") or cfg_fields.get("customer_id") or ""
+
+    # Metadata — custom properties from respan.json, then env overrides
+    metadata: Dict[str, Any] = {"claude_code_turn": turn_num}
+    if cfg_props:
+        metadata.update(cfg_props)
     if request_id:
         metadata["request_id"] = request_id
     if stop_reason:
         metadata["stop_reason"] = stop_reason
-
-    # Merge user-provided metadata from env var
     env_metadata = os.environ.get("RESPAN_METADATA")
     if env_metadata:
         try:
             extra = json.loads(env_metadata)
             if isinstance(extra, dict):
                 metadata.update(extra)
-            else:
-                debug("RESPAN_METADATA is not a JSON object, skipping")
-        except json.JSONDecodeError as e:
-            debug(f"Invalid JSON in RESPAN_METADATA, skipping: {e}")
-    
-    # Build usage object with cache details
-    usage_obj = None
+        except json.JSONDecodeError:
+            pass
+
+    # Usage
+    usage_fields: Dict[str, Any] = {}
     if usage:
-        usage_obj = {
-            "prompt_tokens": usage.get("input_tokens", 0),
-            "completion_tokens": usage.get("output_tokens", 0),
-        }
-        total_tokens = usage_obj["prompt_tokens"] + usage_obj["completion_tokens"]
-        if total_tokens > 0:
-            usage_obj["total_tokens"] = total_tokens
-        
-        # Add cache details
-        prompt_tokens_details = {}
+        prompt_tokens = usage.get("input_tokens", 0)
+        completion_tokens = usage.get("output_tokens", 0)
+        usage_fields["prompt_tokens"] = prompt_tokens
+        usage_fields["completion_tokens"] = completion_tokens
+        total = prompt_tokens + completion_tokens
+        if total > 0:
+            usage_fields["total_tokens"] = total
         cache_creation = usage.get("cache_creation_input_tokens", 0)
         cache_read = usage.get("cache_read_input_tokens", 0)
+        if cache_creation > 0:
+            usage_fields["cache_creation_prompt_tokens"] = cache_creation
+        prompt_tokens_details: Dict[str, int] = {}
         if cache_creation > 0:
             prompt_tokens_details["cache_creation_tokens"] = cache_creation
-            usage_obj["cache_creation_prompt_tokens"] = cache_creation
         if cache_read > 0:
             prompt_tokens_details["cached_tokens"] = cache_read
-        
         if prompt_tokens_details:
-            usage_obj["prompt_tokens_details"] = prompt_tokens_details
-        
-        # Add service tier to metadata
+            usage_fields["prompt_tokens_details"] = prompt_tokens_details
         service_tier = usage.get("service_tier")
         if service_tier:
             metadata["service_tier"] = service_tier
-    
-    # Create chat span (root)
-    chat_span_id = f"claudecode_{trace_unique_id}_chat"
-    customer_id = os.environ.get("RESPAN_CUSTOMER_ID", "claude-code")
-    chat_span = {
+
+    # ------------------------------------------------------------------
+    # 2. Root span – pure agent container (no model / token info)
+    # ------------------------------------------------------------------
+    root_span_id = f"claudecode_{trace_unique_id}_root"
+    root_span: Dict[str, Any] = {
         "trace_unique_id": trace_unique_id,
         "thread_identifier": thread_id,
         "customer_identifier": customer_id,
-        "span_unique_id": chat_span_id,
-        "span_parent_id": None,
+        "span_unique_id": root_span_id,
         "span_name": root_span_name,
         "span_workflow_name": workflow_name,
-        "log_type": "agent",
+        "model": model,
+        "provider_id": "",
+        "span_path": "",
         "input": json.dumps(prompt_messages) if prompt_messages else "",
         "output": json.dumps(completion_message) if completion_message else "",
-        "prompt_messages": prompt_messages,
-        "completion_message": completion_message,
-        "model": model,
         "timestamp": timestamp_str,
         "start_time": start_time_str,
         "metadata": metadata,
     }
-    
-    # Add usage if available
-    if usage_obj:
-        chat_span["prompt_tokens"] = usage_obj["prompt_tokens"]
-        chat_span["completion_tokens"] = usage_obj["completion_tokens"]
-        if "total_tokens" in usage_obj:
-            chat_span["total_tokens"] = usage_obj["total_tokens"]
-        if "cache_creation_prompt_tokens" in usage_obj:
-            chat_span["cache_creation_prompt_tokens"] = usage_obj["cache_creation_prompt_tokens"]
-        if "prompt_tokens_details" in usage_obj:
-            chat_span["prompt_tokens_details"] = usage_obj["prompt_tokens_details"]
-    
-    # Add latency if calculated
     if latency is not None:
-        chat_span["latency"] = latency
-    
-    spans.append(chat_span)
-    
-    # Extract thinking blocks and create spans for them
-    thinking_spans = []
-    for idx, assistant_msg in enumerate(assistant_msgs):
-        if isinstance(assistant_msg, dict) and "message" in assistant_msg:
-            content = assistant_msg["message"].get("content", [])
-            if isinstance(content, list):
-                for item in content:
-                    if isinstance(item, dict) and item.get("type") == "thinking":
-                        thinking_text = item.get("thinking", "")
-                        if thinking_text:
-                            thinking_span_id = f"claudecode_{trace_unique_id}_thinking_{len(thinking_spans) + 1}"
-                            thinking_timestamp = assistant_msg.get("timestamp", timestamp_str)
-                            thinking_spans.append({
-                                "trace_unique_id": trace_unique_id,
-                                "span_unique_id": thinking_span_id,
-                                "span_parent_id": chat_span_id,
-                                "span_name": f"Thinking {len(thinking_spans) + 1}",
-                                "span_workflow_name": workflow_name,
-                                "log_type": "generation",
-                                "input": "",
-                                "output": thinking_text,
-                                "timestamp": thinking_timestamp,
-                                "start_time": thinking_timestamp,
-                            })
-    
-    spans.extend(thinking_spans)
-    
-    # Collect all tool calls and results with metadata
-    tool_call_map = {}
+        root_span["latency"] = latency
+    spans.append(root_span)
+
+    # ------------------------------------------------------------------
+    # 3. LLM generation child span (always created → every turn has ≥1 child)
+    # ------------------------------------------------------------------
+    gen_span_id = f"claudecode_{trace_unique_id}_gen"
+    gen_start = first_assistant_timestamp or start_time_str
+    gen_end = last_assistant_timestamp or timestamp_str
+    gen_latency = None
+    gen_start_dt = parse_timestamp(gen_start) if gen_start else None
+    gen_end_dt = parse_timestamp(gen_end) if gen_end else None
+    if gen_start_dt and gen_end_dt:
+        gen_latency = (gen_end_dt - gen_start_dt).total_seconds()
+
+    gen_span: Dict[str, Any] = {
+        "trace_unique_id": trace_unique_id,
+        "span_unique_id": gen_span_id,
+        "span_parent_id": root_span_id,
+        "span_name": "claude.chat",
+        "span_workflow_name": workflow_name,
+        "span_path": "claude_chat",
+        "model": model,
+        "provider_id": "anthropic",
+        "metadata": {},
+        "input": json.dumps(prompt_messages) if prompt_messages else "",
+        "output": json.dumps(completion_message) if completion_message else "",
+        "prompt_messages": prompt_messages,
+        "completion_message": completion_message,
+        "timestamp": gen_end,
+        "start_time": gen_start,
+    }
+    if gen_latency is not None:
+        gen_span["latency"] = gen_latency
+    gen_span.update(usage_fields)
+    spans.append(gen_span)
+
+    # ------------------------------------------------------------------
+    # 4. Thinking child spans
+    # ------------------------------------------------------------------
+    thinking_num = 0
     for assistant_msg in assistant_msgs:
-        tool_calls = get_tool_calls(assistant_msg)
-        for tool_call in tool_calls:
-            tool_name = tool_call.get("name", "unknown")
-            tool_input = tool_call.get("input", {})
+        if not (isinstance(assistant_msg, dict) and "message" in assistant_msg):
+            continue
+        content = assistant_msg["message"].get("content", [])
+        if not isinstance(content, list):
+            continue
+        for item in content:
+            if isinstance(item, dict) and item.get("type") == "thinking":
+                thinking_text = item.get("thinking", "")
+                if not thinking_text:
+                    continue
+                thinking_num += 1
+                thinking_ts = assistant_msg.get("timestamp", timestamp_str)
+                spans.append({
+                    "trace_unique_id": trace_unique_id,
+                    "span_unique_id": f"claudecode_{trace_unique_id}_thinking_{thinking_num}",
+                    "span_parent_id": root_span_id,
+                    "span_name": f"Thinking {thinking_num}",
+                    "span_workflow_name": workflow_name,
+                    "span_path": "thinking",
+                    "provider_id": "",
+                    "metadata": {},
+                    "input": "",
+                    "output": thinking_text,
+                    "timestamp": thinking_ts,
+                    "start_time": thinking_ts,
+                })
+
+    # ------------------------------------------------------------------
+    # 5. Tool child spans
+    # ------------------------------------------------------------------
+    tool_call_map: Dict[str, Dict[str, Any]] = {}
+    for assistant_msg in assistant_msgs:
+        for tool_call in get_tool_calls(assistant_msg):
             tool_id = tool_call.get("id", "")
             tool_call_map[tool_id] = {
-                "name": tool_name,
-                "input": tool_input,
+                "name": tool_call.get("name", "unknown"),
+                "input": tool_call.get("input", {}),
                 "id": tool_id,
                 "timestamp": assistant_msg.get("timestamp") if isinstance(assistant_msg, dict) else None,
             }
-    
-    # Find matching tool results with metadata
+
     for tr in tool_results:
         tr_content = get_content(tr)
-        tool_result_metadata = {}
-        
-        # Extract tool result metadata
+        tool_result_metadata: Dict[str, Any] = {}
         if isinstance(tr, dict):
-            tool_use_result = tr.get("toolUseResult", {})
-            if tool_use_result:
-                if "durationMs" in tool_use_result:
-                    tool_result_metadata["duration_ms"] = tool_use_result["durationMs"]
-                if "numFiles" in tool_use_result:
-                    tool_result_metadata["num_files"] = tool_use_result["numFiles"]
-                if "filenames" in tool_use_result:
-                    tool_result_metadata["filenames"] = tool_use_result["filenames"]
-                if "truncated" in tool_use_result:
-                    tool_result_metadata["truncated"] = tool_use_result["truncated"]
-        
+            tur = tr.get("toolUseResult") or {}
+            for src, dst in [("durationMs", "duration_ms"), ("numFiles", "num_files"),
+                             ("filenames", "filenames"), ("truncated", "truncated")]:
+                if src in tur:
+                    tool_result_metadata[dst] = tur[src]
         if isinstance(tr_content, list):
             for item in tr_content:
                 if isinstance(item, dict) and item.get("type") == "tool_result":
@@ -560,44 +601,52 @@ def create_respan_spans(
                         tool_call_map[tool_use_id]["output"] = item.get("content")
                         tool_call_map[tool_use_id]["result_metadata"] = tool_result_metadata
                         tool_call_map[tool_use_id]["result_timestamp"] = tr.get("timestamp")
-    
-    # Create tool spans (children)
+
     tool_num = 0
-    for tool_id, tool_data in tool_call_map.items():
+    for tool_id, td in tool_call_map.items():
         tool_num += 1
-        tool_span_id = f"claudecode_{trace_unique_id}_tool_{tool_num}"
-        
-        # Use tool result timestamp if available, otherwise use tool call timestamp
-        tool_timestamp = tool_data.get("result_timestamp") or tool_data.get("timestamp") or timestamp_str
-        tool_start_time = tool_data.get("timestamp") or start_time_str
-        
-        # Format input and output for better readability
-        formatted_input = format_tool_input(tool_data['name'], tool_data["input"])
-        formatted_output = format_tool_output(tool_data['name'], tool_data.get("output"))
-        
-        tool_span = {
+        tool_ts = td.get("result_timestamp") or td.get("timestamp") or timestamp_str
+        tool_start = td.get("timestamp") or start_time_str
+        tool_span: Dict[str, Any] = {
             "trace_unique_id": trace_unique_id,
-            "span_unique_id": tool_span_id,
-            "span_parent_id": chat_span_id,
-            "span_name": f"Tool: {tool_data['name']}",
+            "span_unique_id": f"claudecode_{trace_unique_id}_tool_{tool_num}",
+            "span_parent_id": root_span_id,
+            "span_name": f"Tool: {td['name']}",
             "span_workflow_name": workflow_name,
-            "log_type": "tool",
-            "input": formatted_input,
-            "output": formatted_output,
-            "timestamp": tool_timestamp,
-            "start_time": tool_start_time,
+            "span_path": f"tool_{td['name'].lower()}",
+            "provider_id": "",
+            "metadata": td.get("result_metadata") or {},
+            "input": format_tool_input(td["name"], td["input"]),
+            "output": format_tool_output(td["name"], td.get("output")),
+            "timestamp": tool_ts,
+            "start_time": tool_start,
         }
-        
-        # Add tool result metadata if available
-        if tool_data.get("result_metadata"):
-            tool_span["metadata"] = tool_data["result_metadata"]
-            # Calculate latency if duration_ms is available
-            duration_ms = tool_data["result_metadata"].get("duration_ms")
+        if td.get("result_metadata"):
+            duration_ms = td["result_metadata"].get("duration_ms")
             if duration_ms:
-                tool_span["latency"] = duration_ms / 1000.0  # Convert ms to seconds
-        
+                tool_span["latency"] = duration_ms / 1000.0
         spans.append(tool_span)
-    
+
+    # Add required Respan platform fields to every span.
+    # The backend expects these on all spans (per official SDK examples).
+    respan_defaults = {
+        "warnings": "",
+        "encoding_format": "float",
+        "disable_fallback": False,
+        "respan_params": {
+            "has_webhook": False,
+            "environment": os.environ.get("RESPAN_ENVIRONMENT", "prod"),
+        },
+        "field_name": "data: ",
+        "delimiter": "\n\n",
+        "disable_log": False,
+        "request_breakdown": False,
+    }
+    for span in spans:
+        for key, value in respan_defaults.items():
+            if key not in span:
+                span[key] = value
+
     return spans
 
 
@@ -607,36 +656,49 @@ def send_spans(
     base_url: str,
     turn_num: int,
 ) -> None:
-    """Send spans to Respan with timeout and one retry on transient errors."""
+    """Send spans to Respan as a single batch (matches official SDK behaviour).
+
+    The official Respan tracing SDK sends all spans for a trace in one
+    POST request to ``/v1/traces/ingest``.  We do the same here, with
+    simple retry logic for transient server errors.
+    """
     url = f"{base_url}/v1/traces/ingest"
     headers = {"Authorization": f"Bearer {api_key}"}
 
-    for attempt in range(2):
+    span_names = [s.get("span_name", "?") for s in spans]
+    payload_json = json.dumps(spans)
+    payload_size = len(payload_json)
+    debug(f"Sending {len(spans)} spans ({payload_size} bytes) for turn {turn_num}: {span_names}")
+    if DEBUG:
+        debug_file = LOG_FILE.parent / f"respan_spans_turn_{turn_num}.json"
+        debug_file.write_text(payload_json, encoding="utf-8")
+        debug(f"Dumped spans to {debug_file}")
+
+    for attempt in range(3):
         try:
             response = requests.post(url, json=spans, headers=headers, timeout=30)
             if response.status_code < 400:
-                debug(f"Sent {len(spans)} spans for turn {turn_num}")
+                resp_text = response.text[:300] if response.text else ""
+                debug(f"Sent {len(spans)} spans for turn {turn_num} "
+                      f"(attempt {attempt + 1}): {resp_text}")
                 return
             if response.status_code < 500:
-                # 4xx — not retryable
-                log("ERROR", f"Failed to send spans for turn {turn_num}: HTTP {response.status_code}")
+                log("ERROR", f"Spans rejected for turn {turn_num}: "
+                    f"HTTP {response.status_code} - {response.text[:200]}")
                 return
-            # 5xx — retryable
-            if attempt == 0:
-                debug(f"Server error {response.status_code} for turn {turn_num}, retrying...")
-                time.sleep(1)
-                continue
-            log("ERROR", f"Failed to send spans for turn {turn_num} after retry: HTTP {response.status_code}")
-        except (requests.exceptions.Timeout, requests.exceptions.ConnectionError) as e:
-            if attempt == 0:
-                debug(f"Transient error for turn {turn_num}: {e}, retrying...")
-                time.sleep(1)
-                continue
-            log("ERROR", f"Failed to send spans for turn {turn_num} after retry: {e}")
+            # 5xx — retry after short delay
+            debug(f"Server error for turn {turn_num} "
+                  f"(attempt {attempt + 1}), retrying...")
+            time.sleep(1.0)
+        except (requests.exceptions.Timeout, requests.exceptions.ConnectionError):
+            time.sleep(1.0)
         except Exception as e:
             log("ERROR", f"Failed to send spans for turn {turn_num}: {e}")
             return
 
+    log("ERROR", f"Failed to send {len(spans)} spans for turn {turn_num} "
+        f"after 3 attempts")
+
 
 def process_transcript(
     session_id: str,
@@ -644,6 +706,7 @@ def process_transcript(
     state: Dict[str, Any],
     api_key: str,
     base_url: str,
+    config: Optional[Dict[str, Any]] = None,
 ) -> int:
     """Process a transcript file and create traces for new turns."""
     # Get previous state for this session
@@ -693,7 +756,8 @@ def process_transcript(
         turns_processed += 1
         turn_num = turn_count + turns_processed
         spans = create_respan_spans(
-            session_id, turn_num, current_user, current_assistants, current_tool_results
+            session_id, turn_num, current_user, current_assistants, current_tool_results,
+            config=config,
         )
         send_spans(spans, api_key, base_url, turn_num)
         last_committed_line = total_lines  # safe default, refined below
@@ -755,8 +819,17 @@ def process_transcript(
         current_assistants.append(merged)
 
     if current_user and current_assistants:
-        _commit_turn()
-        last_committed_line = total_lines
+        # Check if the turn has actual text output.  The Stop hook can fire
+        # before the final assistant text block is flushed to disk, leaving
+        # only thinking/tool_use blocks.  If no text content is found, treat
+        # the turn as incomplete so the retry logic re-reads it.
+        has_text = any(get_text_content(m) for m in current_assistants)
+        if has_text:
+            _commit_turn()
+            last_committed_line = total_lines
+        else:
+            last_committed_line = current_user_line
+            debug(f"Turn has assistant msgs but no text output yet (likely not flushed), will retry")
     else:
         # Incomplete turn — rewind so the next run re-reads from the
         # unmatched user message (or from where we left off if no user).
@@ -869,13 +942,35 @@ def main():
         debug("Tracing disabled (TRACE_TO_RESPAN != true)")
         sys.exit(0)
 
-    # Check for required environment variables
+    # Resolve API key: env var > ~/.respan/credentials.json
     api_key = os.getenv("RESPAN_API_KEY")
-    # Default: api.respan.ai | Enterprise: endpoint.respan.ai (set RESPAN_BASE_URL)
     base_url = os.getenv("RESPAN_BASE_URL", "https://api.respan.ai/api")
 
     if not api_key:
-        log("ERROR", "Respan API key not set (RESPAN_API_KEY)")
+        creds_file = Path.home() / ".respan" / "credentials.json"
+        if creds_file.exists():
+            try:
+                creds = json.loads(creds_file.read_text(encoding="utf-8"))
+                # Find the active profile's credential
+                config_file = Path.home() / ".respan" / "config.json"
+                profile = "default"
+                if config_file.exists():
+                    cfg = json.loads(config_file.read_text(encoding="utf-8"))
+                    profile = cfg.get("activeProfile", "default")
+                cred = creds.get(profile, {})
+                api_key = cred.get("apiKey") or cred.get("accessToken")
+                if not base_url or base_url == "https://api.respan.ai/api":
+                    base_url = cred.get("baseUrl", base_url)
+                # Ensure base_url ends with /api (credentials store the host only)
+                if base_url and not base_url.rstrip("/").endswith("/api"):
+                    base_url = base_url.rstrip("/") + "/api"
+                if api_key:
+                    debug(f"Using API key from credentials.json (profile: {profile})")
+            except (json.JSONDecodeError, IOError) as e:
+                debug(f"Failed to read credentials.json: {e}")
+
+    if not api_key:
+        log("ERROR", "No API key found. Run: respan auth login")
         sys.exit(0)
 
     # Try stdin payload first, fall back to filesystem scan
@@ -894,11 +989,49 @@ def main():
 
     debug(f"Processing session: {session_id}")
 
-    # Process the transcript under file lock
+    # Load respan.json config from the project directory.
+    # Extract the project CWD from the first user message in the transcript.
+    config: Dict[str, Any] = {"fields": {}, "properties": {}}
     try:
-        with state_lock():
-            state = load_state()
-            turns = process_transcript(session_id, transcript_file, state, api_key, base_url)
+        first_line = transcript_file.read_text(encoding="utf-8").split("\n")[0]
+        if first_line:
+            first_msg = json.loads(first_line)
+            cwd = first_msg.get("cwd")
+            if not cwd:
+                # Try second line (first is often file-history-snapshot)
+                lines = transcript_file.read_text(encoding="utf-8").split("\n")
+                for line in lines[:5]:
+                    if line.strip():
+                        msg = json.loads(line)
+                        cwd = msg.get("cwd")
+                        if cwd:
+                            break
+            if cwd:
+                config = load_respan_config(cwd)
+                debug(f"Loaded respan.json config from {cwd}: {config}")
+    except Exception as e:
+        debug(f"Failed to extract CWD or load config: {e}")
+
+    # Process the transcript under file lock.
+    # Retry up to 3 times with a short delay — the Stop hook can fire
+    # before Claude Code finishes flushing the assistant response to
+    # the transcript file, causing an incomplete turn on the first read.
+    max_attempts = 3
+    turns = 0
+    try:
+        for attempt in range(max_attempts):
+            with state_lock():
+                state = load_state()
+                turns = process_transcript(session_id, transcript_file, state, api_key, base_url, config=config)
+
+            if turns > 0:
+                break
+
+            if attempt < max_attempts - 1:
+                delay = 0.5 * (attempt + 1)
+                debug(f"No turns processed (attempt {attempt + 1}/{max_attempts}), "
+                      f"retrying in {delay}s...")
+                time.sleep(delay)
 
         # Log execution time
         duration = (datetime.now() - script_start).total_seconds()