Claude Code Skill

Autotel includes a Claude Code skill that gives AI coding agents deep knowledge of the instrumentation API. When the skill is active, Claude Code can:

Wrap functions with trace() and set semantic attributes
Convert console.log calls to getRequestLogger() wide events
Add createStructuredError() with why, fix, and link fields
Set up init() with the correct configuration
Review code for observability anti-patterns

Installation

The skill lives at .claude/skills/autotel/ in the repository root. Claude Code discovers it automatically when you work inside the repo. No extra configuration needed.

For other projects that use Autotel, copy the .claude/skills/autotel/ directory into your repo’s .claude/skills/ folder.

Skill Structure

.claude/skills/autotel/
├── SKILL.md                          # Main skill entry point
└── references/
    ├── wide-events.md                # Canonical log lines guide
    ├── structured-errors.md          # Error handling patterns
    ├── request-logger.md             # Request-scoped logging
    └── code-review.md                # Anti-pattern detection checklist

Reference Guides

Wide Events (Canonical Log Lines)

File: .claude/skills/autotel/references/wide-events.md

A wide event is a single structured record per request containing all context accumulated during processing. Instead of 10 scattered log lines, you get one queryable event with 25+ fields.

Covers:

Anatomy of a canonical log line (core fields, trace identifiers, duration)
Context grouping with dot-prefixed keys
Configuration options (keep, shouldEmit, pretty, drain, attributeRedactor)
Drain pipeline for production fan-out with batching and retry
SQL query examples for analyzing wide events

init({
  service: 'my-app',
  canonicalLogLines: {
    enabled: true,
    rootSpansOnly: true,
    keep: [
      { status: 500 },
      { durationMs: 1000 },
    ],
    pretty: true,
    drain: (event) => sendToAnalytics(event),
  },
});

Structured Errors

File: .claude/skills/autotel/references/structured-errors.md

Covers three APIs:

API	Purpose
`createStructuredError()`	Create errors with `message`, `why`, `fix`, `link`, `code`, `status`, `details`
`recordStructuredError()`	Record a structured error onto a trace context (sets status + exception + attributes)
`parseError()`	Normalize any thrown value to a consistent shape for API responses

Common patterns:

API endpoint error responses
Wrapping third-party errors (Stripe, etc.)
Domain validation errors
Span attribute mapping (error.type, error.why, error.fix, etc.)

throw createStructuredError({
  message: 'Payment declined',
  why: 'Card has insufficient funds for the requested amount',
  fix: 'Prompt the user to update their payment method',
  code: 'CARD_DECLINED',
  status: 402,
  link: 'https://docs.stripe.com/declines/codes#insufficient-funds',
});

Request Logger

File: .claude/skills/autotel/references/request-logger.md

Covers the getRequestLogger() API and framework adapters:

log.set(fields) — merge fields onto the span (nested objects flatten automatically)
log.info(message, fields?) — span event at info level
log.warn(message, fields?) — span event at warn level, promotes canonical log line
log.error(error, fields?) — record structured error on span
log.getContext() — snapshot of all accumulated fields
log.emitNow(overrides?) — emit snapshot immediately with onEmit fan-out

Framework adapters:

useLogger() with withAutotel() (Next.js, Nitro, Cloudflare Workers)
useLogger() alone when middleware already creates the span (Hono)
Custom adapters via createUseLogger()

const log = getRequestLogger(ctx);
log.set({ feature: 'checkout', tier: req.user.tier });

const cart = await loadCart(req.cartId);
log.set({ cart_items: cart.items.length, cart_total: cart.total });

Code Review Checklist

File: .claude/skills/autotel/references/code-review.md

Anti-pattern detection table with severity levels and fixes:

Anti-Pattern	Severity	Fix
`console.log` for request context	High	`getRequestLogger(ctx)` with `log.set()`
`console.error` in catch blocks	High	`log.error(err)` or `recordStructuredError()`
`throw new Error('...')` without context	Medium	`createStructuredError({ message, why, fix })`
`catch (e) { throw e }` without recording	Medium	Add `log.error(e)` before rethrowing
`JSON.stringify` for log output	High	`log.set()` — attributes flatten automatically
Manual request ID generation	Low	`ctx.correlationId` from trace context

Also includes a 4-step migration path from console.log to fully instrumented Autotel code.

How the Skill Works

When you ask Claude Code to instrument code, review for observability, or work with errors/logging, the skill activates and loads the relevant reference guides. Claude Code then applies the patterns directly — wrapping functions with trace(), replacing console.log with request loggers, and adding structured error context.

Example Prompts

“Add tracing to this handler”
“Convert the console.logs in this file to use request logger”
“Add structured errors to the catch blocks”
“Review this file for observability anti-patterns”
“Set up canonical log lines for this service”

CLI-Backed Skill (Alternative to MCP)

The skill described above teaches Claude Code how to write instrumented code. A second class of agent task — investigating telemetry (querying traces, diagnosing errors, ranking expensive LLM calls) — is served by two interchangeable mechanisms:

autotel-mcp — an MCP server the user wires into .mcp.json.
The autotel CLI itself — the same investigation surface as one-shot subcommands that emit JSON on stdout.

Both share the same backends (packages/autotel-mcp/src/backends/) and the same query helpers, so the answers are identical. The CLI route is useful when you’d rather not configure an MCP server, or when you’re driving an agent harness that doesn’t speak MCP.

Command surface

MCP tool group	CLI equivalent
`health.backend_health`	`autotel health`
`health.backend_capabilities`	`autotel capabilities`
`discovery.discover_services`	`autotel discover services`
`discovery.discover_trace_fields`	`autotel discover trace-fields`
`discovery.discover_log_fields`	`autotel discover log-fields`
`investigation.search_traces`	`autotel query traces`
`investigation.search_spans`	`autotel query spans`
`investigation.get_trace`	`autotel trace get <id>`
`investigation.summarize_trace`	`autotel trace summary <id>`
`signals.list_metrics`	`autotel query metrics`
`signals.search_logs`	`autotel query logs`
`topology.list_services`	`autotel topology services`
`topology.list_operations`	`autotel topology operations <name>`
`topology.service_map`	`autotel topology map`
`diagnosis.find_anomalies`	`autotel diagnose anomalies`
`diagnosis.find_root_cause`	`autotel diagnose root-cause <id>`
`diagnosis.find_errors`	`autotel diagnose errors`
`diagnosis.check_slos`	`autotel diagnose slos --service <name>`
`correlation.correlate`	`autotel correlate trace <id>`
`correlation.explain_slowdown`	`autotel correlate explain-slowdown --service <name>`
`llm-analytics.get_llm_usage`	`autotel llm usage`
`llm-analytics.list_llm_models`	`autotel llm models`
`llm-analytics.get_llm_model_stats`	`autotel llm model-stats --model-name <name>`
`llm-analytics.get_llm_expensive_traces`	`autotel llm expensive`
`llm-analytics.get_llm_slow_traces`	`autotel llm slow`
`llm-analytics.list_llm_tools`	`autotel llm tools`
`semantic-conventions.*`	`autotel semconv list \| get <ns> \| refresh`
`instrumentation.score_span_instrumentation`	`autotel score` (JSON span on stdin)
`instrumentation.explain_instrumentation_score`	`autotel score explain`
`collector-config.*`	`autotel collector validate \| suggest \| explain`
`collector-schema.*`	`autotel collector versions \| components \| schema \| readme \| validate-component \| refresh`

Backend selection

Every backend-touching command accepts the same flags as the MCP server’s env vars, with flags winning over env when both are set:

autotel query traces \
  --backend tempo \
  --tempo-base-url http://localhost:3200 \
  --service-name checkout \
  --error-only \
  --limit 20

Available backends: collector (built-in OTLP receiver, in-memory or libsql-persisted) · jaeger · tempo · prometheus · loki · stack (composite) · auto (probe localhost) · fixture (JSON file).

Output shape

Every investigate command emits one JSON document on stdout. Success:

{ "ok": true, "command": "query traces", "data": { /* … */ } }

Failure:

{
  "ok": false,
  "error": {
    "type": "validation" | "runtime",
    "code": "AUTOTEL_E_INVALID_INPUT" | "AUTOTEL_E_UNKNOWN",
    "message": "…",
    "retryable": false
  }
}

Exit code is 0 on success, 2 on validation, 1 on runtime. Agents check result.ok and read data (or error).

Why offer both

	MCP server	CLI subcommands
Install step	Edit `.mcp.json`, restart client	`npm i -g autotel`
Transport	stdio / HTTP per client	Shell exec
Streaming results	Yes	One-shot per command
Cross-tool state	Server holds backend connection	Each call re-opens
Works in non-MCP agents	No	Yes
Discoverability	MCP tool list	`autotel <group> --help`

For an agent driving a single investigation prompt the CLI is usually faster — no MCP handshake, no per-tool round-trip. For an interactive incident review with many follow-up queries against a heavy backend, the MCP server’s persistent backend connection wins.

Skill file

A ready-made skill that maps prompts → CLI commands ships at .claude/skills/autotel-investigate/SKILL.md. Claude Code discovers it automatically when you work inside the repo; copy the directory into another project’s .claude/skills/ to use it elsewhere.