Session Observability & Monitoring

Track Claude Code usage, estimate costs, and identify patterns across your development sessions.

For systems with explicit loops or workflow graphs, instrument the decision layer as well as the model and tools. Loop & Graph Engineering defines the minimum evidence for routes, state transitions, checkpoints, verdicts, and human overrides.

Table of Contents

1. Why Monitor Sessions
2. Session Search & Resume
3. Setting Up Session Logging
4. Analyzing Session Data
5. Cost Tracking
6. Activity Monitoring
7. External Monitoring Tools
8. Proxying Claude Code
9. Patterns & Best Practices
10. Limitations

---

Why Monitor Sessions

Claude Code usage can accumulate quickly, especially in active development. Monitoring helps you:

• Understand costs: Estimate API spend before invoices arrive
• Identify patterns: See which tools you use most, which files get edited repeatedly
• Optimize workflow: Find inefficiencies (e.g., repeatedly reading the same large file)
• Track projects: Compare usage across different codebases
• Team visibility: Aggregate usage for team budgeting (when combining logs)

---

Observability by Harness Layer

Capture evidence at the layer that made the decision. A runtime harness needs traces of model calls, tool calls, permissions, context, and recovery. A repository harness needs the setup command, changed artifact, and verification result. An orchestrator needs dispatch, handoff, retry, queue, and human-escalation events. A dashboard that reports only tokens cannot establish whether the loop or its delivery gates behaved correctly.

The Agent Harness Map separates loop-owning runtimes from the broader directory of frameworks and control-plane-adjacent projects. Agent Harness Engineering explains those layers, while Agent Evaluation defines the measures to score them. Apply Security Hardening before exporting traces or granting a monitoring tool access. The glossary keeps the terms consistent.

Use stable event names and version every harness field that can change behavior. The OpenTelemetry generative AI attribute registry defines attributes for agent, model, provider, operation, request, response, token, and tool data. Treat prompt text, tool arguments, tool results, file paths, and user identifiers as sensitive payloads: redact or hash them before export, and document which fields were dropped.

An optimizer needs a second trace level. Link each candidate harness version to its parent, mutation, evaluation tasks, budget, scores, and promotion decision. Without that lineage, a higher final score cannot establish which change caused it or whether the candidate consumed more search and execution budget.

Graph-Level Observability

A graph-based workflow needs more than model and tool spans. The trace must reconstruct why work moved, waited, repeated, or stopped.

Field	Question it answers
graph.version, policy.version	Which executable topology and control policy governed the run?
node.id, edge.id, route.reason	Which branch ran, and why was it selected?
state.before, state.after	What changed at the transition boundary?
join.expected, join.received, join.wait_ms	Was parallel work complete, missing, or blocked?
retry.count, interrupt.reason, resume.from	Did recovery repeat work or resume from a checkpoint?
reviewer.model, reviewer.provider, evidence.refs	How independent and evidence-backed was the verdict?
human.checkpoint, human.wait_ms, verdict.overturned	Where did judgment return to a person, and with what effect?

The official LangGraph Graph API exposes state, nodes, edges, conditional routing, parallel super-steps, and runtime metadata. Its persistence model adds checkpoints, replay, human-in-the-loop interruption, and fault recovery. Those are traceable events, not merely implementation details.

Liza illustrates the same need in a coding control plane. Its pinned architectural issues ledger identifies manual checkpoint latency, cross-pair review, provider-diversity gaps, and unmeasured reviewer accuracy as open concerns. Tokens and task status alone cannot expose those failure modes. Record routing, quorum composition, review evidence, wait time, and verdict outcomes separately.

---

Session Search & Resume

After weeks of using Claude Code, finding past conversations becomes challenging. This section covers native options and community tools.

Native Commands

Command	Use Case
claude -c / claude --continue	Resume most recent session
claude -r <id> / claude --resume <id>	Resume specific session by ID
claude --resume	Interactive session picker

Sessions are stored locally at ~/.claude/projects/<project>/ as JSONL files.

Community Tools Comparison

Tool	Install	List Speed	Search Speed	Dependencies	Resume Command
session-search.sh (this repo)	Copy script	10ms	400ms	None (bash)	✅ Displayed
claude-conversation-extractor	pip install	230ms	1.7s	Python	❌
claude-code-transcripts	uvx	N/A	N/A	Python	❌
ran CLI	npm -g	N/A	Fast	Node.js	❌ (commands only)

Recommended: session-search.sh

Zero-dependency bash script optimized for speed with ready-to-use resume commands.

Install:

cp examples/scripts/session-search.sh ~/.claude/scripts/cs
chmod +x ~/.claude/scripts/cs
echo "alias cs='~/.claude/scripts/cs'" >> ~/.zshrc
source ~/.zshrc

Usage:

cs                          # List 10 most recent sessions (~15ms)
cs "authentication"         # Single keyword search (~400ms)
cs "Prisma migration"       # Multi-word AND search (both must match)
cs -n 20                    # Show 20 results
cs -p myproject "bug"       # Filter by project name
cs --since 7d               # Sessions from last 7 days
cs --since today            # Today's sessions only
cs --json "api" | jq .      # JSON output for scripting
cs --rebuild                # Force index rebuild

Output:

2026-01-15 08:32 │ my-project             │ Implement OAuth flow for...
  claude --resume 84287c0d-8778-4a8d-abf1-eb2807e327a8

2026-01-14 21:13 │ other-project          │ Fix database migration...
  claude --resume 1340c42e-eac5-4181-8407-cc76e1a76219

Copy-paste the claude --resume command to continue any session.

How It Works

1. Index mode (no filters): Uses cached TSV index. Auto-refreshes when sessions change. ~15ms lookup.
2. Search mode (with keyword/filters): Full-text search with 3s timeout. Multi-word queries use AND logic.
3. Filters: --project (substring match), --since (supports today, yesterday, 7d, YYYY-MM-DD)
4. Output: Human-readable by default, --json for scripting. Excludes agent/subagent sessions.

Alternative: Python Tools

If you prefer richer features (HTML export, multiple formats):

# Install
pip install claude-conversation-extractor

# Interactive UI
claude-start

# Direct search
claude-search "keyword"

# Export to markdown
claude-extract --format markdown

See session-search.sh for the complete script.

---

Session Resume Limitations & Cross-Folder Migration

TL;DR: Native --resume is limited to the current working directory by design. For cross-folder migration, use manual filesystem operations (recommended) or community automation tools (untested).

Why Resume is Directory-Scoped

Claude Code stores sessions at ~/.claude/projects/<encoded-path>/ where <encoded-path> is derived from your project’s absolute path. For example:

• Project at /home/user/myapp → Sessions in ~/.claude/projects/-home-user-myapp-/
• Project moved to /home/user/projects/myapp → Claude looks for ~/.claude/projects/-home-user-projects-myapp-/ (different directory)

Design rationale: Sessions store absolute file paths, project-specific context (MCP server configs, .claudeignore rules, environment variables). Cross-folder resume would require path rewriting and context validation, which isn’t implemented yet.

Related: GitHub issue #1516 tracks community requests for native cross-folder support.

Manual Migration (Recommended)

When moving a project folder:

# Before moving project
cd ~/.claude/projects/
ls -la  # Note the current encoded path

# Move your project
mv /old/location/myapp /new/location/myapp

# Rename session directory to match new path
cd ~/.claude/projects/
mv -- -old-location-myapp- -new-location-myapp-

# Verify
cd /new/location/myapp
claude --continue  # Should resume successfully

When forking sessions to a new project:

# Copy session files (preserves original)
cd ~/.claude/projects/
cp -n ./-source-project-/*.jsonl ./-target-project-/

# Copy subagents directory if exists
if [ -d ./-source-project-/subagents ]; then
  cp -r ./-source-project-/subagents ./-target-project-/
fi

# Resume in target project
cd /path/to/target/project
claude --continue

⚠️ Migration Risks & Caveats

Before migrating sessions, verify compatibility:

Risk	Impact	Mitigation
Hardcoded secrets	Credentials exposed in new context	Audit .jsonl files before migration, redact if needed
Absolute paths	File references break if paths differ	Verify paths exist in target, or accept broken references
MCP server configs	Source MCP servers missing in target	Install matching MCP servers before resuming
.claudeignore rules	Different ignore patterns	Review differences, merge if needed
Environment variables	process.env context mismatch	Check .env files compatibility

When NOT to migrate sessions:

• Conflicting dependencies (e.g., different Node.js versions, package managers)
• Database state differences (migrations applied in source, not in target)
• Authentication context (API tokens, OAuth sessions specific to source project)
• Security boundaries (migrating from private to public repo)

Community Automation Tool

claude-migrate-session by Jim Weller (inspired by Alexis Laporte) automates the manual process above:

• Repository: jimweller/dotfiles
• Features: Global search with filtering, preserves .jsonl + subagents, uses ripgrep for performance
• Status: Personal dotfiles (0 stars/forks as of Feb 2026), limited adoption
• Command: /claude-migrate-session <source> <target>

⚠️ Caveat: This tool has minimal community testing. The manual approach is safer and gives you explicit control over what gets migrated. Test thoroughly before using in production workflows.

Use cases for migration:

• Forking prototype work into production codebase
• Moving debugging session to isolated test repository
• Continuing architecture discussion in a new project

Alternative: Entire CLI Session Portability

Native limitation: Claude Code’s --resume is tied to absolute file paths, breaking on folder moves.

Entire CLI solution: Checkpoints are path-agnostic, enabling true session portability across project locations.

How it works:

# In source project
cd /old/location/myapp
entire capture --agent="claude-code"
[... work in Claude Code ...]
entire checkpoint --name="migration-complete"

# Move project to new location
mv /old/location/myapp /new/location/myapp

# Resume in target (works because Entire stores relative paths)
cd /new/location/myapp
entire resume --checkpoint="migration-complete"
claude --continue  # Resumes with full context

Why Entire checkpoints are portable:

Aspect	Native --resume	Entire CLI
Path storage	Absolute paths in JSONL	Relative paths in checkpoints
Cross-folder	Breaks (different project encoding)	Works (path-agnostic)
Context preservation	Prompt history only	Prompts + reasoning + file states
Agent handoffs	No	Yes (between Claude/Gemini)

When to use Entire over manual migration:

• ✅ Frequent project moves/forks
• ✅ Multi-agent workflows (Claude → Gemini handoffs)
• ✅ Session replay for debugging (rewind to exact state)
• ✅ Governance (approval gates on resume)

Trade-off: Adds tool dependency + storage overhead (~5-10% project size).

Full docs: AI Traceability Guide

---

Multi-Agent Orchestration Monitoring

For monitoring multiple concurrent Claude Code instances via external orchestrators (Gas Town, multiclaude), see:

• agent-chat (https://github.com/justinabrahms/agent-chat): Real-time Slack-like UI for agent communications
• Architecture guide: guide/ecosystem/ai-ecosystem.md Section 8.1 - Multi-Agent Orchestration Systems

Architecture pattern (for custom implementations):

1. Hook logs Task agent spawns: .claude/hooks/multi-agent-logger.sh
2. Store in SQLite: ~/.claude/logs/agents.db (parent_id, child_id, timestamp, task)
3. Stream via SSE: Simple Go/Node HTTP server
4. Dashboard: React/HTML consuming SSE stream

Native Claude Code monitoring (this guide):

• Session search: session-search.sh (see Session Search & Resume)
• Activity logs: session-logger.sh hook (see Setting Up Session Logging)
• Stats analysis: session-stats.sh (see Analyzing Session Data)

When to use external orchestrator monitoring:

• Running Gas Town or multiclaude with 5+ concurrent agents
• Need real-time visibility into agent coordination
• Debugging orchestration failures (agent conflicts, merge issues)

When native monitoring suffices:

• Single Claude Code session or --delegate with <3 subagents
• Post-hoc analysis (logs, stats) is enough
• Budget/complexity constraints

---

Setting Up Session Logging

1. Install the Logger Hook

Copy the session logger to your hooks directory:

# Create hooks directory if needed
mkdir -p ~/.claude/hooks

# Copy the logger (from this repo's examples)
cp examples/hooks/bash/session-logger.sh ~/.claude/hooks/
chmod +x ~/.claude/hooks/session-logger.sh

2. Register in Settings

Add to ~/.claude/settings.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "type": "command",
        "command": "~/.claude/hooks/session-logger.sh"
      }
    ]
  }
}

3. Verify Installation

Run a few Claude Code commands, then check logs:

ls ~/.claude/logs/
# Should see: activity-2026-01-14.jsonl

# View recent entries
tail -5 ~/.claude/logs/activity-$(date +%Y-%m-%d).jsonl | jq .

Configuration Options

Environment Variable	Default	Description
CLAUDE_LOG_DIR	~/.claude/logs	Where to store logs
CLAUDE_LOG_TOKENS	true	Enable token estimation
CLAUDE_SESSION_ID	auto-generated	Custom session identifier

---

Analyzing Session Data

Using session-stats.sh

# Copy the script
cp examples/scripts/session-stats.sh ~/.local/bin/
chmod +x ~/.local/bin/session-stats.sh

# Today's summary
session-stats.sh

# Last 7 days
session-stats.sh --range week

# Specific date
session-stats.sh --date 2026-01-14

# Filter by project
session-stats.sh --project my-app

# Machine-readable output
session-stats.sh --json

Sample Output

═══════════════════════════════════════════════════════════
        Claude Code Session Statistics - today
═══════════════════════════════════════════════════════════

Summary
  Total operations:  127
  Sessions:          3

Token Usage
  Input tokens:      45,230
  Output tokens:     12,450
  Total tokens:      57,680

Estimated Cost (Sonnet rates)
  Input:   $0.1357
  Output:  $0.1868
  Total:   $0.3225

Tools Used
  Edit: 45
  Read: 38
  Bash: 24
  Grep: 12
  Write: 8

Projects
  my-app: 89
  other-project: 38

Reading for Quality, Not Just Quantity

Token counts tell you how much you used Claude Code. JSONL logs can also tell you how well your configuration is working, if you know what to look for.

Beyond cost metrics, three patterns reliably signal that a skill, rule, or CLAUDE.md section needs updating:

Repeated reads of the same file

If Claude reads the same file 3+ times in one session, the content it needs probably isn’t where it expects to find it. Consider moving the relevant context into a skill or CLAUDE.md section.

# Files read more than 3x in recent sessions
jq -r 'select(.tool == "Read") | .file' ~/.claude/logs/activity-*.jsonl \
  | sort | uniq -c | sort -rn | awk '$1 > 3'

Tool failures on the same command

A Bash command that fails repeatedly across sessions usually means a skill has an outdated path, renamed binary, or command that no longer works with your current stack.

# Failing commands
jq -r 'select(.tool == "Bash" and (.exit_code // 0) != 0) | .command' \
  ~/.claude/logs/activity-*.jsonl | sort | uniq -c | sort -rn | head -10

High edit frequency on the same file

Files edited heavily across sessions often indicate missing context: the file’s purpose isn’t clear to the agent, or conventions around it aren’t documented.

# Most-edited files (proxy for context gaps)
jq -r 'select(.tool == "Edit") | .file' ~/.claude/logs/activity-*.jsonl \
  | sort | uniq -c | sort -rn | head -10

For each pattern you surface, ask: is there a skill, rule, or CLAUDE.md section that should cover this? See §9.23 Configuration Lifecycle & The Update Loop for the full workflow.

---

Log Format

Each log entry is a JSON object:

{
  "timestamp": "2026-01-14T15:30:00Z",
  "session_id": "1705234567-12345",
  "tool": "Edit",
  "file": "src/components/Button.tsx",
  "project": "my-app",
  "tokens": {
    "input": 350,
    "output": 120,
    "total": 470
  }
}

---

Cost Tracking

Token Estimation Method

The logger estimates tokens using a simple heuristic: ~4 characters per token. This is approximate and tends to slightly overestimate.

Cost Rates

Default rates are for Claude Sonnet. Adjust via environment variables:

# Sonnet rates (default)
export CLAUDE_RATE_INPUT=0.003   # $3/1M tokens
export CLAUDE_RATE_OUTPUT=0.015  # $15/1M tokens

# Opus rates (if using Opus)
export CLAUDE_RATE_INPUT=0.015   # $15/1M tokens
export CLAUDE_RATE_OUTPUT=0.075  # $75/1M tokens

# Haiku rates
export CLAUDE_RATE_INPUT=0.00025 # $0.25/1M tokens
export CLAUDE_RATE_OUTPUT=0.00125 # $1.25/1M tokens

Budget Alerts (Manual Pattern)

Add to your shell profile for daily budget warnings:

# ~/.zshrc or ~/.bashrc
claude_budget_check() {
  local cost=$(session-stats.sh --json 2>/dev/null | jq -r '.summary.estimated_cost.total // 0')
  local threshold=5.00  # $5 daily budget

  if (( $(echo "$cost > $threshold" | bc -l) )); then
    echo "⚠️  Claude Code daily spend: \$$cost (threshold: \$$threshold)"
  fi
}

# Run on shell start
claude_budget_check

---

Activity Monitoring

Cost tracking tells you how much you spend. Activity monitoring tells you what Claude Code actually did: which files it read, which commands it ran, which URLs it fetched. This is the audit layer.

Session JSONL: The Ground Truth

Every tool call Claude Code makes is recorded in the session JSONL files at ~/.claude/projects/<project>/. Each entry with type: "assistant" contains a content array where type: "tool_use" blocks document every action.

# Find your session files
ls ~/.claude/projects/-$(pwd | tr '/' '-')-/

# Inspect tool calls in a session
cat ~/.claude/projects/-your-project-/SESSION_ID.jsonl | \
  jq 'select(.type == "assistant") | .message.content[]? | select(.type == "tool_use") | {tool: .name, input: .input}'

What Tool Calls Reveal

Tool	What It Exposes
Read	Files accessed (path, line range)
Write / Edit	Files modified (path, content delta)
Bash	Commands executed (full command string)
WebFetch	URLs fetched (may include data sent in POST)
Task	Subagent spawns (prompt passed to sub-model)
Glob / Grep	Search patterns and scope

Practical Audit Queries

# All files read in a session
SESSION=~/.claude/projects/-your-project-/SESSION_ID.jsonl
jq 'select(.type == "assistant") | .message.content[]? | select(.type == "tool_use" and .name == "Read") | .input.file_path' "$SESSION"

# All bash commands executed
jq 'select(.type == "assistant") | .message.content[]? | select(.type == "tool_use" and .name == "Bash") | .input.command' "$SESSION"

# All URLs fetched
jq 'select(.type == "assistant") | .message.content[]? | select(.type == "tool_use" and .name == "WebFetch") | .input.url' "$SESSION"

# Count tool usage by type
jq -r 'select(.type == "assistant") | .message.content[]? | select(.type == "tool_use") | .name' "$SESSION" | sort | uniq -c | sort -rn

Sensitive Patterns to Watch

These tool call patterns are worth flagging in automated audits:

Pattern	Risk	Detection
Read on .env, *.pem, id_rsa	Credential access	`jq ‘…
Bash with rm -rf, git push --force	Destructive action	`jq ‘…
WebFetch on external URLs	Data exfiltration risk	`jq ‘…
Write on files outside project root	Scope creep	Check paths against working directory

Security context: Claude Code operates read-write on your filesystem with your user permissions. The JSONL audit trail is your record of what happened. For teams, consider syncing these logs to immutable storage.

---

External Monitoring Tools

Beyond the hook-based approach above, the community has built purpose-specific tools. This is a factual snapshot as of early 2026.

Instrumenting the coding assistant itself, not just the application it produces, is what turns AI usage into something auditable. Tracing every decision an agent makes, including tools like Claude Code and Codex, with OpenTelemetry closes the gap between “the AI probably did the right thing” and having the exact chain of tool calls that produced a given output. A system that cannot be observed at that level cannot be structurally trusted with real autonomy.

Annie Freeman, Devoxx, 2026, talk “Un-observable AI is un-trustworthy AI”

Tool	Type	What It Does	Install
ccusage	CLI / TUI	Cost tracking from JSONL. The de-facto reference for pricing data. ~10K GitHub stars.	npm i -g ccusage
claude-code-otel	OpenTelemetry exporter	Emits spans to any OTEL collector. Integrates with Prometheus + Grafana dashboards. Enterprise-focused.	npm i -g claude-code-otel
Akto	SaaS / self-hosted	API security guardrails + audit trail. Intercepts at the API level, flags policy violations.	akto.io
MLflow Tracing	CLI + SDK	Exact token counts, tool spans, LLM-as-judge evaluation. CLI mode: zero Python required. Best for ML/MLOps teams.	pip install mlflow → see section below
ccboard	TUI + Web	Unified dashboard for sessions, costs, stats. Activity/audit tab in development.	cargo install ccboard
claude-crusts	CLI	One-command context pollution scanner: flags stale files, oversized memories, and redundant rule loading. Spot what’s inflating your context before running sessions.	github.com/Abinesh-L/claude-crusts

Decision Guide

Want cost numbers fast?          → ccusage (CLI, 0 config)
Need enterprise audit trail?     → claude-code-otel + Grafana or Akto
Already using MLflow for ML?     → MLflow tracing integration (see below)
Need agent regression detection? → MLflow tracing + LLM-as-judge
Want a persistent TUI/Web UI?    → ccboard
Context pollution audit?         → claude-crusts (1-command scan)

ccusage

npm i -g ccusage
ccusage          # Today's usage
ccusage --days 7 # Last 7 days

Reads directly from ~/.claude/projects/**/*.jsonl. No API keys, no data sent externally. Source: github.com/ryoppippi/ccusage.

claude-code-otel

Exports Claude Code activity as OpenTelemetry spans:

npm i -g claude-code-otel
claude-code-otel --collector http://localhost:4318

Spans include tool name, duration, token counts. Plug into any OTEL-compatible backend (Jaeger, Tempo, Datadog). Source: github.com/badger-99/claude-code-otel.

ccboard

cargo install ccboard
ccboard              # Launch TUI
ccboard --web        # Launch Web UI (localhost:3000)

Source: github.com/FlorianBruniaux/ccboard. An Activity tab covering file access, bash commands, and network calls is planned (see docs/resource-evaluations/ccboard-activity-module-plan.md).

MLflow Tracing

When to use: Teams already in the MLflow/MLOps ecosystem, or anyone needing exact token counts + LLM-based quality evaluation. Not the right fit for solo devs wanting quick cost numbers (use ccusage instead).

What makes it different from the other tools: MLflow intercepts at the API level, not post-hoc from JSONL. It captures exact token counts (vs the ~15-25% variance of hook-based estimation) and enables LLM-as-judge regression detection, not just “what happened” but “was it good?”.

Setup: CLI mode (no Python required)

Works with interactive claude sessions. Hooks into .claude/settings.json:

pip install "mlflow[genai]>=3.4"

# Enable tracing in current project directory
mlflow autolog claude

# With custom backend (recommended for persistence)
mlflow autolog claude -u sqlite:///mlflow.db

# With named experiment
mlflow autolog claude -n "my-project"

# Check status / disable
mlflow autolog claude --status
mlflow autolog claude --disable

Launch the UI to inspect traces:

mlflow server  # → http://localhost:5000

What gets captured automatically: user prompts, assistant responses, tool calls (name + inputs + outputs), token counts (exact), latency per call, session metadata.

Setup: SDK mode (Python agents)

import mlflow
mlflow.anthropic.autolog()         # one line, before anything else
mlflow.set_experiment("my-agent")

# Use ClaudeSDKClient normally — all interactions are traced
# ⚠️ Only ClaudeSDKClient is supported. Direct API calls are not traced.
from anthropic import claude_agent_sdk
async with ClaudeSDKClient(options=AGENT_OPTIONS) as client:
    await client.query(query)

Requires: mlflow>=3.5 + claude-agent-sdk>=0.1.0.

MCP server: bidirectional integration

Claude Code can query its own traces directly. Add to .claude/settings.json:

{
  "mcpServers": {
    "mlflow-mcp": {
      "command": "uv",
      "args": ["run", "--with", "mlflow[mcp]>=3.5.1", "mlflow", "mcp", "run"],
      "env": { "MLFLOW_TRACKING_URI": "<your-tracking-uri>" }
    }
  }
}

Once configured, you can ask Claude Code: “Find all sessions where the backend-architect agent used more than 20 tool calls”. It queries MLflow directly without copy-pasting IDs.

LLM-as-judge: agent regression detection

The key capability absent from all other tools in this section. After modifying an agent’s instructions, measure whether quality improved or degraded:

from mlflow.genai.scorers import scorer, ConversationCompleteness, RelevanceToQuery
from mlflow.entities.model_registry import Feedback

@scorer
def tool_efficiency(trace) -> int:
    """Count tool calls — lower is better for well-scoped tasks."""
    return len(trace.search_spans(span_type="TOOL"))

@scorer
def permission_blocks(trace) -> int:
    """Detect how often the agent was blocked by permission gates."""
    return sum(
        1 for span in trace.search_spans(span_type="TOOL")
        if span.outputs and "requires approval" in str(span.outputs).lower()
    )

# Run evaluation against recorded traces
traces = mlflow.search_traces(experiment_ids=["<id>"], max_results=50)
results = mlflow.genai.evaluate(
    data=traces,
    scorers=[
        tool_efficiency,
        permission_blocks,
        ConversationCompleteness(),
        RelevanceToQuery(),
    ]
)

Built-in scorers: ConversationCompleteness, RelevanceToQuery, UserFrustration, SafetyScorer.

Custom scorers: full access to the trace object (all spans, inputs, outputs, token counts).

Limitations

Limitation	Detail
CLI mode audience	Best for interactive sessions; SDK mode required for programmatic agents
SDK restriction	ClaudeSDKClient only; direct API calls bypass tracing
PII risk	Traces capture full conversation content. Redact before storing if working with sensitive data
Production backend	SQLite = dev only. Use PostgreSQL/MySQL for production
OpenTelemetry	MLflow 3.6+ exports to any OTEL-compatible backend (Datadog, Grafana, etc.)

---

Proxying Claude Code

A common question: “Can I run Proxyman/Charles to see what Claude Code sends to Anthropic?”

Short answer: Not directly. Here’s why, and what works instead.

Why System Proxies Don’t Work

Claude Code is a Node.js process. By default, Node.js ignores system-level proxy settings (HTTP_PROXY, HTTPS_PROXY). It uses its own TLS stack and doesn’t read macOS/Windows proxy configurations.

Additionally, even if traffic flows through your proxy, the TLS certificate mismatch causes Claude Code to fail (CERT_UNTRUSTED).

Option 1: Trust a MITM Certificate (Proxyman / Charles)

Force Node.js to trust your proxy’s CA certificate:

# Export Proxyman's CA cert (File → Export → Root Certificate)
# Then point Node.js at it:
export NODE_EXTRA_CA_CERTS="/path/to/proxyman-ca.pem"

# Start Claude Code — traffic will now route through Proxyman
claude

Same approach works for Charles: Help → SSL Proxying → Export Charles Root Certificate.

Caveats:

• Some Claude Code versions use certificate pinning for api.anthropic.com; this may still fail
• This approach requires a running Proxyman/Charles instance listening on the configured port

Option 2: Redirect API Traffic with ANTHROPIC_API_URL

Point Claude Code at a local interceptor instead of api.anthropic.com:

export ANTHROPIC_API_URL="http://localhost:8080"
claude

Run any HTTP proxy/logger on port 8080 that forwards to https://api.anthropic.com. This bypasses TLS entirely for the Claude Code → proxy hop.

Use cases: Logging request payloads, injecting headers, rate-limiting locally, replaying requests.

Option 3: mitmproxy (Recommended)

mitmproxy is the cleanest open-source solution. It provides a scriptable HTTPS proxy with a web UI and terminal interface.

# Install
brew install mitmproxy  # macOS
# or: pip install mitmproxy

# Start transparent proxy on port 8080
mitmproxy --listen-port 8080

# In a new terminal, point Claude Code at it
export NODE_EXTRA_CA_CERTS="$(python3 -c 'import mitmproxy.certs; print(mitmproxy.certs.Cert.default_ca_path())')"
export HTTPS_PROXY="http://localhost:8080"
claude

The mitmproxy web UI (mitmweb) at http://localhost:8081 shows full request/response bodies, including the JSON payloads Claude Code sends to Anthropic.

What you’ll see: System prompt, user messages, tool definitions, tool results, model parameters.

Option 4: Minimal Python Logging Proxy

For a zero-dependency approach:

# proxy.py — simple HTTPS logging proxy
from http.server import HTTPServer, BaseHTTPRequestHandler
import urllib.request, json, sys

TARGET = "https://api.anthropic.com"

class LoggingProxy(BaseHTTPRequestHandler):
    def do_POST(self):
        length = int(self.headers["Content-Length"])
        body = self.rfile.read(length)
        print(json.dumps(json.loads(body), indent=2))  # Log request
        # Forward to Anthropic...

HTTPServer(("localhost", 8080), LoggingProxy).serve_forever()

python3 proxy.py &
export ANTHROPIC_API_URL="http://localhost:8080"
claude

Privacy note: Proxied traffic includes everything in the conversation context: file contents Claude has read, your code, any secrets it encountered. Handle proxy logs accordingly.

---

Patterns & Best Practices

1. Weekly Review

Set a calendar reminder to review weekly stats:

session-stats.sh --range week

Look for:

• Unusually high token usage days
• Repeated operations on same files (inefficiency signal)
• Project distribution (where time is spent)

2. Per-Project Tracking

Use CLAUDE_SESSION_ID to tag sessions by project:

export CLAUDE_SESSION_ID="project-myapp-$(date +%s)"
claude

3. Team Aggregation

For team-wide tracking, sync logs to shared storage:

# Example: sync to S3 daily
aws s3 sync ~/.claude/logs/ s3://company-claude-logs/$(whoami)/

Then aggregate with:

# Download all team logs
aws s3 sync s3://company-claude-logs/ /tmp/team-logs/

# Combine and analyze
cat /tmp/team-logs/*/activity-$(date +%Y-%m-%d).jsonl | \
  jq -s 'group_by(.project) | map({project: .[0].project, total_tokens: [.[].tokens.total] | add})'

4. Log Rotation

Logs accumulate over time. Add cleanup to cron:

# Clean logs older than 30 days
find ~/.claude/logs -name "*.jsonl" -mtime +30 -delete

5. Resilience Testing for AI Dependencies

Deliberately simulate AI unavailability or degraded output to confirm the team can keep working without it. Provider outages, extreme latency, and low-quality responses are failure modes worth rehearsing on purpose, the same way chaos engineering rehearses server or network failures. A workflow that silently assumes the AI is always available and always correct only gets that assumption tested for the first time during an actual incident.

Konstantin Pavlov, Devoxx, 2025

---

Limitations

What This Monitoring CANNOT Do

Limitation	Reason
Exact token counts	Claude Code CLI doesn’t expose API token metrics
TTFT (Time to First Token)	Hook runs after tool completes, not during streaming
Real-time streaming metrics	No hook event during response generation
Actual API costs	Token estimates are heuristic, not from billing
Model selection	Log doesn’t capture which model was used per request
Context window usage	No visibility into current context percentage

Accuracy Notes

• Token estimates: ~15-25% variance from actual billing
• Cost estimates: Use as directional guidance, not accounting
• Session boundaries: Sessions are approximated by ID, not exact API sessions

What You CAN Trust

• Tool usage counts: Exact count of each tool invocation
• File access patterns: Which files were touched
• Relative comparisons: Day-to-day/project-to-project trends
• Operation timing: When tools were used (timestamp)

---

---

Manager Audit Checklist

For engineering managers and team leads who need to verify Claude Code is being used appropriately within their team, these are the practical audit queries.

Weekly Spot Check (5 minutes)

# Did anything unusual happen this week?

# 1. Files accessed outside project scope
find ~/.claude/projects/ -name "*.jsonl" -newer "$(date -d '7 days ago' +%Y-%m-%d 2>/dev/null || date -v-7d +%Y-%m-%d)" 2>/dev/null | \
  xargs jq -r 'select(.type == "assistant") |
    .message.content[]? |
    select(.type == "tool_use" and .name == "Read") |
    .input.file_path' 2>/dev/null | \
  grep -v "^$(pwd)" | sort -u

# 2. Destructive commands run
find ~/.claude/projects/ -name "*.jsonl" -newer "$(date -d '7 days ago' +%Y-%m-%d 2>/dev/null || date -v-7d +%Y-%m-%d)" 2>/dev/null | \
  xargs jq -r 'select(.type == "assistant") |
    .message.content[]? |
    select(.type == "tool_use" and .name == "Bash") |
    .input.command' 2>/dev/null | \
  grep -iE "(drop|delete|truncate|rm -rf|git push --force)"

Compliance Reporting

For regulated environments, generate a summary of AI activity for auditors:

#!/bin/bash
# Monthly compliance report for Claude Code activity

START_DATE=${1:-$(date -d '30 days ago' +%Y-%m-%d 2>/dev/null || date -v-30d +%Y-%m-%d)}
END_DATE=${2:-$(date +%Y-%m-%d)}
REPORT_FILE="ai-activity-report-${START_DATE}-${END_DATE}.json"

echo "Generating compliance report: $START_DATE to $END_DATE"

# Count sessions, tool calls, and file accesses
jq -s '{
  report_period: {start: "'"$START_DATE"'", end: "'"$END_DATE"'"},
  tool_usage: (group_by(.tool) | map({tool: .[0].tool, count: length})),
  unique_files_accessed: ([.[].file] | unique | length),
  sessions: ([.[].session_id] | unique | length)
}' ~/.claude/logs/activity-*.jsonl 2>/dev/null > "$REPORT_FILE" || \
  echo "No activity logs found. Set up session-logger.sh hook to enable."

echo "Report saved: $REPORT_FILE"

For a full governance setup with automatic audit trail logging, see Enterprise AI Governance §6.2.

---

10. Team-level Log Aggregation

The sections above cover individual developer monitoring. For a team of 10+ developers, you need logs flowing into a central store to track total spend, flag unusual patterns, and answer compliance questions.

Option A: Route Configured API Traffic Through LiteLLM

The API Gateway guide describes server-side telemetry for clients configured with the gateway URL and a team-scoped virtual key. After that client configuration, the team does not need per-machine cron jobs to collect routed request metadata. The gateway still observes only traffic sent through it. Subscription-authenticated Claude usage and clients that retain direct provider credentials remain outside this dataset.

For routed API traffic, this avoids synchronizing local usage logs. The session JSONL approach below remains necessary when you need file-level detail, such as which files were read and written, rather than model and token telemetry alone.

Option B: Ship Session JSONL to a Central Store

Claude Code writes session logs to ~/.claude/projects/**/*.jsonl. A lightweight cron job or PostToolUse hook can ship new entries to Loki:

#!/bin/bash
# Ship new session log entries to Loki
# Add to crontab: */5 * * * * ~/.claude/hooks/ship-logs.sh

LOKI_ENDPOINT="${CENTRAL_LOG_ENDPOINT:-http://loki:3100/loki/api/v1/push}"
TEAM_ID="${CLAUDE_TEAM_ID:-unknown}"
MARKER="/tmp/.claude-last-ship-${USER}"

find ~/.claude/projects -name "*.jsonl" -newer "$MARKER" 2>/dev/null | while read -r f; do
  while IFS= read -r line; do
    curl -s -X POST "$LOKI_ENDPOINT" \
      -H "Content-Type: application/json" \
      -d "{\"streams\":[{\"stream\":{\"app\":\"claude_code\",\"team\":\"$TEAM_ID\",\"user\":\"$(whoami)\"},\"values\":[[\"$(date +%s%N)\",$(echo "$line" | jq -Rs .)]}]}" \
      > /dev/null
  done < "$f"
done
touch "$MARKER"

Set CLAUDE_TEAM_ID and CENTRAL_LOG_ENDPOINT in each developer’s shell profile. Coordinate with your platform team to provision the Loki endpoint.

OTel Collector Configuration

If you are using the LiteLLM Gateway with OTel export, a minimal collector config to forward spans to Grafana Tempo:

receivers:
  otlp:
    protocols:
      grpc:
        endpoint: 0.0.0.0:4317

processors:
  batch:
    timeout: 5s

exporters:
  otlp:
    endpoint: tempo:4317
    tls:
      insecure: true

service:
  pipelines:
    traces:
      receivers: [otlp]
      processors: [batch]
      exporters: [otlp]

This feeds Claude Code usage spans into the same Tempo backend as your application traces, so you can correlate an AI session with a deployment event on a single timeline.

Key Grafana Queries

With LiteLLM Prometheus metrics and Loki for session logs, three panels cover most team needs:

# Daily spend by team (USD)
sum by (team_id) (increase(litellm_spend_total[1d]))

# Request volume by model
sum by (model) (rate(litellm_requests_total[1h]))

# Budget headroom (fraction remaining per team)
1 - (litellm_spend_total / litellm_budget_limit)

---

Related Resources

• Session Search Script - Fast session search & resume
• Session Logger Hook
• Stats Analysis Script
• Enterprise AI Governance - Org-level governance, audit trails, compliance
• Third-Party Tools - Community GUIs, TUIs, and dashboards (ccusage, ccburn, claude-code-viewer)
• Data Privacy Guide - What data leaves your machine
• Cost Optimization - Tips to reduce spend