agent-telemetry（Agent 遥测）

Agent Telemetry

Make application runtime behavior queryable by coding agents through structured logging and telemetry endpoints.

Core Problem

Coding agents debugging issues often can't answer "what actually happened at runtime?" because:

• Logs don't exist, or are unstructured console.log noise
• Logs exist but there's no documented way for agents to query them
• Agent docs (CLAUDE.md, AGENTS.md) don't mention how to access telemetry

Workflow

Phase 1: Audit Current State

Determine what telemetry already exists.

1. Check for logging infrastructure:

# Find logging configuration and usage
grep -r "winston\|pino\|bunyan\|log4j\|slog\|Logger\|logging\.config" --include="*.{ts,js,py,rb,go,rs}" -l .

# Find log output configuration
grep -r "LOG_LEVEL\|LOG_FORMAT\|LOG_FILE\|OTEL_\|SENTRY_DSN" .env* config/ -l 2>/dev/null

2. Check for existing telemetry endpoints:

# Health/debug/metrics endpoints
grep -r "health\|metrics\|debug\|status\|readiness\|liveness" --include="*.{ts,js,py,rb,go}" -l src/ app/ 2>/dev/null

3. Check agent docs for log access instructions:

# Do agent docs mention logs?
grep -ri "log\|telemetry\|debug\|observ" CLAUDE.md AGENTS.md .claude/*.md .cursor/*.md 2>/dev/null

4. Classify the result:

Phase 2: Add Structured Logging

If no structured logging exists, add it. See references/logging-setup.md for framework-specific patterns.

Principles:

• Use structured JSON logs, not string interpolation
• Include correlation IDs for request tracing
• Log at boundaries: incoming requests, outgoing calls, errors, state transitions
• Use consistent field names: timestamp, level, message, requestId, userId, duration, error

Where to add logging (priority order):

1. Request/response middleware (every request gets logged)
2. Error handlers (unhandled errors get captured with context)
3. External service calls (DB queries, API calls, queue operations)
4. Business logic decision points (state transitions, authorization decisions)

Minimum viable logging — add a request logger middleware that captures:

{timestamp, level, requestId, method, path, statusCode, duration, userId?}

This single addition makes most debugging possible.

Phase 3: Expose Logs to Agents

Agents need a way to query logs without SSH access or cloud console dashboards. Provide at least one of:

Option A: Log file (simplest) Write structured logs to a known file path agents can read directly.

# Agent reads recent errors
tail -100 logs/app.json | jq 'select(.level == "error")'

# Agent reads logs for a specific request
grep "requestId.*abc123" logs/app.json | jq .

Option B: Dev log endpoint (recommended for web apps) Add a development-only endpoint that returns recent log entries with filtering.

GET /__dev/logs?level=error&last=50
GET /__dev/logs?path=/api/users&last=20
GET /__dev/logs?requestId=abc-123

This endpoint must:

• Only be available in development (NODE_ENV=development or equivalent)
• Return JSON array of log entries
• Support filtering by level, path, timerange, requestId
• Limit response size (default 100 entries)

See references/dev-endpoint.md for implementation patterns by framework.

Option C: CLI query tool Wrap log access in a script agents can execute:

# Query recent errors
./scripts/query-logs.sh --level error --last 50

# Query by request path
./scripts/query-logs.sh --path /api/users --since "5 minutes ago"

Choose based on project context:

Phase 4: Document in Agent Docs

This is critical. Without documentation, agents won't know telemetry exists.

Update CLAUDE.md (or equivalent agent doc) with a Debugging section:

## Debugging

### Querying Application Logs

Structured JSON logs are available at [location].

**Quick commands:**

View recent errors

[command to view errors]

View logs for a specific endpoint

[command to filter by path]

View logs for a specific request

[command to filter by request ID]

View logs from the last N minutes

[command to filter by time]

**Log format:**

{ "timestamp": "ISO-8601", "level": "info|warn|error", "message": "Human-readable description", "requestId": "correlation-id", "method": "GET", "path": "/api/resource", "statusCode": 200, "duration": 45 }

**Common debugging workflows:**

- User reports error → query by time range and error level
- Flaky test → query by endpoint path during test run
- Performance issue → query by path, sort by duration

Key rules for the documentation:

• Include copy-pasteable commands (agents execute, not read)
• Show the log schema so agents know what fields to filter on
• List 3-4 common debugging workflows with exact commands
• Mention where log config lives for agents that need to adjust log levels

Phase 5: Validate

Test the full loop:

1. Trigger a request — hit an endpoint or run an operation
2. Query the logs — use the documented method to find the log entry
3. Verify agent usability — can an agent find the relevant log in <3 commands?
4. Check error capture — trigger an error and verify it appears with full context

If any step fails, iterate on the logging or documentation.

Anti-Patterns