manage-claude-code-tracing

Setup Tracing

Configure OpenInference tracing for Claude Code sessions or Agent SDK applications to Arize AX (cloud) or Phoenix (self-hosted). Spans are sent directly to the backend from hooks -- no background process or backend-specific dependencies are needed in the user's environment.

How to Use This Skill

This skill follows a decision tree workflow. Start by asking the user where they are in the setup process:

1. Are they using the Claude Code CLI or the Agent SDK?

   • CLI -> Continue to step 2
   • Agent SDK (Python or TypeScript) -> Go to Agent SDK Setup

2. Do they already have credentials?

   • Yes -> Jump to Configure Settings
   • No -> Continue to step 3

3. Which backend do they want to use?

   • Phoenix (self-hosted) -> Go to Set Up Phoenix
   • Arize AX (cloud) -> Go to Set Up Arize AX

4. Are they troubleshooting?

   • Yes -> Jump to Troubleshoot

Important: Only follow the relevant path for the user's needs. Don't go through all sections.

Set Up Phoenix

Phoenix is self-hosted. No Python dependencies are needed for tracing -- spans are sent directly via send_span() using stdlib urllib.

Install Phoenix

Ask if they already have Phoenix running. If not, walk through:

# Option A: pip
pip install arize-phoenix && phoenix serve

# Option B: Docker
docker run -p 6006:6006 arizephoenix/phoenix:latest

Phoenix UI will be available at http://localhost:6006. Confirm it's running:

curl -sf http://localhost:6006/v1/traces >/dev/null && echo "Phoenix is running" || echo "Phoenix not reachable"

Then proceed to Configure Settings with the Phoenix endpoint.

Set Up Arize AX

Arize AX is available as a SaaS platform or as an on-prem deployment. Users need an account, a space, and an API key.

First, ask the user: "Are you using the Arize SaaS platform or an on-prem instance?"

• SaaS -> Uses the default endpoint (otlp.arize.com:443). Continue below.
• On-prem -> The user will need to provide their custom OTLP endpoint (e.g., otlp.mycompany.arize.com:443). Ask for it and note it for the Configure Settings step.

1. Create an account

If the user doesn't have an Arize account:

• SaaS: Sign up at https://app.arize.com/auth/join
• On-prem: Contact their administrator for access to the on-prem instance

2. Get Space ID and API key

Walk the user through finding their credentials:

1. Log in to their Arize instance (https://app.arize.com for SaaS, or their on-prem URL)
2. Click Settings (gear icon) in the left sidebar
3. The Space ID is shown on the Space Settings page
4. Go to the API Keys tab
5. Click Create API Key or copy an existing one

Both api_key and space_id are required for the shared config.

No Python dependencies are needed. Both Phoenix and Arize AX use HTTP/JSON — no additional Python dependencies are needed.

Then proceed to Configure Settings. If the user is on an on-prem instance, remind them to provide their custom endpoint.

Configure Settings

Important: For marketplace installs, users must run this setup skill before tracing will work. The send_span() function requires ~/.arize/harness/config.json to exist for backend credential resolution.

Configuration has two parts:

1. Backend config (~/.arize/harness/config.json) -- backend credentials and per-harness settings, read by send_span(). This skill creates it.
2. Claude settings (~/.claude/settings.json or .claude/settings.local.json) -- tracing feature flags and user-level env vars

Ask the user for:

1. Scope: Global (~/.claude/settings.json) or project-local (.claude/settings.local.json)
2. Backend choice: Phoenix or Arize AX
3. Credentials (only if no existing config):
   • Phoenix: endpoint URL (default: http://localhost:6006), optional API key
   • Arize AX: API key and Space ID
4. OTLP Endpoint (Arize AX only, optional): For hosted Arize instances using a custom endpoint. Defaults to otlp.arize.com:443.
5. Project name (optional): defaults to "claude-code", stored under harnesses.claude-code.project_name
6. User ID (optional): Set ARIZE_USER_ID to identify spans by user (useful for teams)

Write the backend config

The config file at ~/.arize/harness/config.json is the single source of truth for backend credentials and per-harness settings. Create the directory structure if needed: mkdir -p ~/.arize/harness/{bin,run,logs}

Important: read-merge-write. If ~/.arize/harness/config.json already exists, read it first, then merge in the new or updated fields (e.g., add/update the harnesses.claude-code entry) while preserving existing backend credentials. Only prompt for backend credentials if no existing config is found.

Phoenix:

{
  "harnesses": {
    "claude-code": {
      "project_name": "claude-code",
      "target": "phoenix",
      "endpoint": "<endpoint>",
      "api_key": ""
    }
  }
}

Arize AX:

{
  "harnesses": {
    "claude-code": {
      "project_name": "claude-code",
      "target": "arize",
      "endpoint": "otlp.arize.com:443",
      "api_key": "<key>",
      "space_id": "<id>"
    }
  }
}

If the user has a custom OTLP endpoint, set it in harnesses.claude-code.endpoint.

Write the Claude settings

Determine the config file:

• Global: ~/.claude/settings.json
• Project-local: .claude/settings.local.json (create directory if needed: mkdir -p .claude)

Read the file (or create {} if it doesn't exist), then merge env vars into the "env" object.

{
  "env": {
    "ARIZE_TRACE_ENABLED": "true"
  }
}

If a custom project name was provided, set it in harnesses.claude-code.project_name in the config (~/.arize/harness/config.json), not as an env var.

If a user ID was provided, also set "ARIZE_USER_ID": "<id>". This adds a user.id attribute to all traced spans.

Example workflow:

# For project-local
mkdir -p .claude
echo '{}' > .claude/settings.local.json
# Then use an editor to add env vars

Validate

1. Config exists: Run cat ~/.arize/harness/config.json to verify the config file exists and has correct backend credentials.
2. Phoenix (if applicable): Run curl -sf <endpoint>/v1/traces >/dev/null to check connectivity.

Confirm

Tell the user:

• Backend config saved to ~/.arize/harness/config.json
• Claude settings saved to the chosen file:
  • Global: ~/.claude/settings.json
  • Project-local: .claude/settings.local.json
• Restart the Claude Code session for tracing to take effect
• Spans are sent directly to the backend from hooks — no background process needed
• After restarting, traces will appear in their Phoenix UI or Arize AX dashboard under the project name
• Mention ARIZE_DRY_RUN=true to test without sending data
• Mention ARIZE_VERBOSE=true for debug output
• Errors are always written to ~/.arize/harness/logs/claude-code.log; set ARIZE_VERBOSE to "true" under env in ~/.claude/settings.json to also capture routine hook activity (session_start, span emits, state transitions)

Note: Project-local settings override global settings for the same variables.

Agent SDK Setup

For users building with the Claude Agent SDK (Python or TypeScript), the tracing plugin loads as a local plugin. This section provides code and configuration for the developer to add to their application -- the agent cannot set this up at runtime since plugin paths and settings must be configured before the SDK session starts.

Important: The user must use ClaudeSDKClient -- the standalone query() function does not support hooks, so tracing will not work with it.

How to guide the user

When a user asks about Agent SDK tracing setup, provide them with the steps below to integrate into their own code. Do NOT try to execute export commands or modify their application source -- instead, give them the snippets to copy.

1. Choose a backend

Ask the user which backend they want. If they don't have credentials yet, walk them through Set Up Phoenix or Set Up Arize AX first, then return here.

2. Get the plugin path

Ask the user: "Have you already installed this plugin via the Claude Code CLI?"

If yes (already installed via the install.sh / install.bat flow): The plugin lives inside the harness install directory at ~/.arize/harness/tracing/claude_code.

If yes (installed via the Claude marketplace): They can reference it from the CLI cache. Tell them to check ~/.claude/plugins/installed_plugins.json for the exact path.

If no: Tell them to clone the repo into their project:

git clone https://github.com/Arize-ai/coding-harness-tracing.git

The plugin path will be ./coding-harness-tracing/tracing/claude_code.

Tip: tracing.claude_code.agent_sdk.claude_options() returns a pre-configured ClaudeAgentOptions with the plugin path and setting_sources=["user"] already wired in, so users can skip the manual plumbing in step 5 below when the harness is installed via install.sh.

No Python dependencies are needed -- both Phoenix and Arize AX use HTTP/JSON.

3. Set up the backend config

Ensure ~/.arize/harness/config.json has the correct backend credentials (see Configure Settings above).

4. Create a settings file

The Agent SDK spawns a Claude Code subprocess that does not inherit the user's shell environment variables. Tracing env vars must be passed via a settings file referenced in the ClaudeAgentOptions.

Tell the user to create a settings.local.json file (or similar):

{
  "env": {
    "ARIZE_TRACE_ENABLED": "true"
  }
}

Optional env vars that can also be added to the settings file:

• ARIZE_USER_ID: User identifier added as user.id attribute to all spans (useful for teams)
• ARIZE_DRY_RUN: Set to "true" to test without sending data
• ARIZE_VERBOSE: Set to "true" for debug output

To customize the project name, set it in harnesses.claude-code.project_name in the config (~/.arize/harness/config.json) rather than as an env var.

5. Add the plugin to their code

Give the user the appropriate snippet to add to their application. They must use ClaudeSDKClient and pass both the plugin path (from step 2) and the settings file (from step 4):

Python:

from claude_agent_sdk import ClaudeAgentOptions, ClaudeSDKClient

PLUGIN_PATH = "./coding-harness-tracing/tracing/claude_code"  # or ~/.arize/harness/tracing/claude_code if installed via install.sh

options = ClaudeAgentOptions(
    plugins=[{"type": "local", "path": PLUGIN_PATH}],
    settings="./settings.local.json",
)
async with ClaudeSDKClient(options=options) as client:
    await client.query("Your prompt here")
    async for message in client.receive_response():
        print(message)

TypeScript:

import { ClaudeSDKClient } from "@anthropic-ai/claude-agent-sdk";

const PLUGIN_PATH = "./coding-harness-tracing/tracing/claude_code"; // or ~/.arize/harness/tracing/claude_code if installed via install.sh

const client = new ClaudeSDKClient({
  plugins: [{ type: "local", path: PLUGIN_PATH }],
  settings: "./settings.local.json",
});

await client.connect();
await client.query("Your prompt here");
for await (const message of client.receiveResponse()) {
  console.log(message);
}
await client.close();

6. Validate

Tell the user to add "ARIZE_DRY_RUN": "true" to their settings file to verify hooks fire without sending data, and check ~/.arize/harness/logs/claude-code.log for output.

Agent SDK Compatibility

For full Agent SDK documentation, see: https://platform.claude.com/docs/en/agent-sdk/overview

• Important: You must use ClaudeSDKClient -- the standalone query() function does not support hooks and tracing will not work.
• Hook coverage: The CLI registers a broad hook set (16 events as of this writing, including session-lifecycle, prompt/tool, compaction, and permission events). The Agent SDKs may expose a smaller subset depending on version; check the SDK's HookEvent enum / type for the events your SDK version supports.
• Python SDK in particular has historically not supported SessionStart, SessionEnd, Notification, and PermissionRequest. The plugin handles this automatically -- session state is lazily initialized on the first UserPromptSubmit, so core tracing (LLM spans, tool spans, subagent spans) works fully.
• Tracing env vars must be passed via a settings file in ClaudeAgentOptions -- the SDK subprocess does not inherit shell environment variables.
• If the user is troubleshooting an existing Agent SDK setup, you can help by checking log files (~/.arize/harness/logs/claude-code.log), verifying the settings file contains the correct env vars, verifying ~/.arize/harness/config.json has correct backend credentials, or enabling dry-run mode.

Troubleshoot

Common issues and fixes:

Problem	Fix
Traces not appearing	Check ARIZE_TRACE_ENABLED is "true" in Claude settings, and verify config exists: cat ~/.arize/harness/config.json
Config missing	Run the installer or create ~/.arize/harness/config.json manually (include harnesses section)
Phoenix unreachable	Verify Phoenix is running: curl -sf <endpoint>/v1/traces
No output in terminal	Hook stderr is discarded by Claude Code; check ~/.arize/harness/logs/claude-code.log
Want to test without sending	Set ARIZE_DRY_RUN to "true" in env config
Want verbose logging	Set ARIZE_VERBOSE to "true" in env config
Wrong project name	Set harnesses.claude-code.project_name in ~/.arize/harness/config.json (default: "claude-code")
Spans missing user attribution	Set ARIZE_USER_ID in env config to add user.id to all spans