braintrust-tracing skill

Braintrust Tracing for Claude Code

Comprehensive guide to tracing Claude Code sessions in Braintrust, including sub-agent correlation.

Architecture Overview

PARENT SESSION
                    +---------------------+
                    |  SessionStart       |
                    |  (creates root)     |
                    +----------+----------+
                               |
                    +----------v----------+
                    |  UserPromptSubmit   |
                    |  (creates Turn)     |
                    +----------+----------+
                               |
          +--------------------+--------------------+
          |                    |                    |
+---------v--------+  +--------v--------+  +--------v--------+
| PostToolUse      |  | PostToolUse     |  | PreToolUse      |
| (Read span)      |  | (Edit span)     |  | (Task - inject) |
+------------------+  +-----------------+  +--------+--------+
                                                    |
                                         +----------v----------+
                                         |   SUB-AGENT         |
                                         |   SessionStart      |
                                         |   (NEW root_span_id)|
                                         +----------+----------+
                            

Hook Event Flow

Trace Hierarchy

Session (task span) - root_span_id = session_id
|
+-- Turn 1 (task span)
|   |
|   +-- claude-sonnet (llm span) - model call with tool_use
|   +-- Read (tool span)
|   +-- Edit (tool span)
|   +-- claude-sonnet (llm span) - response after tools
|
+-- Turn 2 (task span)
|   |
|   +-- claude-sonnet (llm span)
|   +-- Task (tool span) -----> [Sub-agent session - SEPARATE trace]
|   +-- claude-sonnet (llm span)
|
+-- Turn 3 ...

Sub-Agent Tracing: What Works and What Doesn't

What Doesn't Work

SessionStart doesn't receive the Task prompt.

We tried injecting trace context into Task prompts via PreToolUse:

# PreToolUse hook injects:
[BRAINTRUST_TRACE_CONTEXT]
{"root_span_id": "abc", "parent_span_id": "xyz", "project_id": "123"}
[/BRAINTRUST_TRACE_CONTEXT]

But SessionStart only receives session metadata, not the modified prompt. The injected context is lost.

What DOES Work

Task spans in parent session contain everything:

• agentId - identifier for the sub-agent run
• totalTokens, totalToolUseCount - metrics
• content - full agent response/summary
• tool_input.prompt - original task prompt
• toolinput.subagenttype - agent type (e.g., "oracle")

SubagentStop hook receives the sub-agent's sessionid:**

• This equals the sub-agent's orphaned trace rootspanid
• Allows correlation between parent Task span and child trace

The Correlation Pattern

Current state: Sub-agents create orphaned traces (new rootspanid).

Correlation method:

1. Query parent session's Task spans for agent metadata
2. Match agentId or timing with orphaned traces
3. Sub-agent's sessionid = its trace's rootspan_id

Future solution (not yet implemented):

SubagentStop fires -> writes session_id to temp file
PostToolUse (Task) -> reads temp file -> adds child_session_id to Task span metadata

This would link: Task.agentId + Task.childsessionid -> orphaned trace rootspanid

State Management

Per-Session State Files

~/.claude/state/braintrust_sessions/
  {session_id}.json       # Per-session state

Each session file contains:

{
  "root_span_id": "abc-123",
  "project_id": "proj-456",
  "turn_count": 5,
  "tool_count": 23,
  "current_turn_span_id": "turn-789",
  "current_turn_start": 1703456789,
  "started": "2025-12-24T10:00:00.000Z",
  "is_subagent": false
}

Global State

~/.claude/state/braintrust_global.json   # Cached project_id
~/.claude/state/braintrust_hook.log      # Debug log

Debugging Commands

Check if Tracing is Active

# View hook logs in real-time
tail -f ~/.claude/state/braintrust_hook.log

# Check if session has state
cat ~/.claude/state/braintrust_sessions/*.json | jq -s '.'

# Verify environment
echo "TRACE_TO_BRAINTRUST=$TRACE_TO_BRAINTRUST"
echo "BRAINTRUST_API_KEY=${BRAINTRUST_API_KEY:+set}"

Query Braintrust Directly

# List recent sessions
uv run python -m runtime.harness scripts/braintrust_analyze.py --sessions 5

# Analyze last session
uv run python -m runtime.harness scripts/braintrust_analyze.py --last-session

# Replay specific session
uv run python -m runtime.harness scripts/braintrust_analyze.py --replay <session-id>

# Find sub-agent traces (orphaned roots)
uv run python -m runtime.harness scripts/braintrust_analyze.py --agent-stats

Debug Hook Execution

# Enable verbose logging
export BRAINTRUST_CC_DEBUG=true

# Test hooks manually
echo '{"session_id":"test-123","type":"resume"}' | \
  bash "$CLAUDE_PROJECT_DIR/.claude/plugins/braintrust-tracing/hooks/session_start.sh"

# Test PreToolUse (Task injection)
echo '{"session_id":"test-123","tool_name":"Task","tool_input":{"prompt":"test"}}' | \
  bash "$CLAUDE_PROJECT_DIR/.claude/plugins/braintrust-tracing/hooks/pre_tool_use.sh"

Troubleshooting Checklist

1. No traces appearing:

• Check TRACETOBRAINTRUST=true in .claude/settings.local.json
• Verify API key: echo $BRAINTRUSTAPIKEY
• Check logs: tail -20 ~/.claude/state/braintrust_hook.log

1. Sub-agents not linking:

• This is expected - sub-agents create orphaned traces
• Use --agent-stats to find agent activity
• Correlate via timing or agentId in parent Task span

1. Missing spans:

• Check currentturnspan_id in session state
• Ensure Stop hook runs (turn finalization)
• Look for "Failed to create" errors in log

1. State corruption:

• Remove session state: rm ~/.claude/state/braintrustsessions/.json
• Clear global cache: rm ~/.claude/state/braintrust_global.json

Key Files

Environment Variables

Session Learnings

What We Learned About Sub-Agent Tracing (Dec 2025)

Attempted: Inject trace context via PreToolUse into Task prompts.

Result: Failed - SessionStart only receives session metadata, not the prompt.

Discovery: Task spans already contain rich sub-agent data:

• metadata.agenttype - agent type from subagenttype
• metadata.skill_name - skill from Skill tool
• tool_input - full prompt sent to agent
• tool_output - agent response

Current correlation path:

1. Parent session Task span has agentId and timing
2. Sub-agent creates orphaned trace with rootspanid = session_id
3. SubagentStop provides the sub-agent's session_id
4. Manual correlation: match timing or use session_id link

Future work: Write childsessionid to Task span metadata from PostToolUse after SubagentStop.

What We Learned About Sub-Agent Correlation