Dd Instrument Llmo

Datadog LLM Observability Instrumentation

This skill assesses the current project (backend runtime, LLM/agent framework, existing instrumentation) and routes you to the right reference file under references/ for the actual setup steps. It is fully self-contained — it does not call any Datadog MCP server. Credential provisioning uses the local CreateApiKey tool, and all code changes are made by you, directly, using your own file-editing tools.

Stay within LLM Observability scope. Do not add RUM, APM application instrumentation, or unrelated Datadog products. RUM and a Datadog Agent are relevant only because they change what session/trace linking is achievable (see the “Beyond SDK init” section in the reference files).

Do NOT invent tool names. Use only CreateApiKey as described in references/common-credentials.md; every other step is done with your normal file-editing/search tools.

Do NOT write anything to memory during or after this skill. Project paths, frameworks, credentials, and ML app names are project-specific and must not be stored in persistent memory.

Ground rules while instrumenting

• Verify before you assert. If you’re not sure about file content or codebase structure, read the files — do not guess.
• Check for existing instrumentation first. Before touching the backend, check whether ddtrace/dd-trace is already initialized with LLM Observability enabled. If it is, do not add a second, competing init/enable call. This only means skip re-init — it does not mean skip the work. SDK presence is not the same as a well-formed trace or a session ID that actually flows; run the session-ID plumbing audit in Phase 1d and close any gap it finds, even when the SDK is already present.
• Only use packages/features you’ve explicitly been told to add. Don’t decide on your own to enable an additional Datadog product or SDK feature beyond what this skill specifies.
• Persist added dependencies into the deploy’s manifest. Any Datadog package you add (ddtrace, dd-trace) must be written into the dependency manifest the build/deploy installs from — the one identified in Phase 1e — not just installed into the local environment. A clean deploy install reads only the manifest and will crash with a missing-module error (e.g. ModuleNotFoundError: No module named 'ddtrace') if the package isn’t declared there.
• Don’t make stylistic changes to code you’re not otherwise touching.
• No package aliases when importing Datadog packages.
• Copy Datadog SDK package names, import paths, and init keyword arguments verbatim from the applicable reference section — do not paraphrase them from memory. The Python LLM Observability import is exactly from ddtrace.llmobs import LLMObs — the module is ddtrace.llmobs, not ddtrace.llm_observability (that module does not exist and produces a deploy-time ModuleNotFoundError). The Node package is dd-trace, initialized with the form that matches the project: CommonJS require('dd-trace').init(...), an ESM import-based form for "type": "module"/.mjs projects, or Next.js’s dd-trace/initialize.mjs in instrumentation.ts — all shown in references/llmobs-nodejs.md; never force require(...) into an ESM project. Use the exact LLMObs.enable(...) / .init({...}) keyword arguments shown; do not add, drop, or rename kwargs based on general Datadog knowledge.
• Use a real edit tool for existing files (Edit/Write, not sed -i, awk, or scripted find/replace via Bash). If an edit-by-text-match fails, re-read the file first rather than retrying the identical edit.
• Checklist discipline. Before starting the instrumentation steps, post a short checklist of the steps you’re about to take. Check items off as you go, and review the checklist at the end of the run.
• After any code change, check package.json for a lint:fix, fix, or format script (or eslint --fix / prettier --write config) and run it automatically — no need to ask permission.
• Verify the app still builds/runs before declaring success (see references/common-verify-report.md).
• LLM Observability session IDs propagate from the root span only. Never pass a session ID to a child span/decorator — see the “Adding spans” section in references/llmobs-python.md / references/llmobs-nodejs.md.
• Don’t promise what the environment doesn’t support. If no RUM SDK is detected, don’t claim RUM session linking; if no dd-agent is detected, don’t claim navigable APM trace linking. State the gap and the upgrade path instead (see the “Beyond SDK init” section in the reference files).

---

Phase 1: Analysis

Inspect the relevant application directory before asking questions or editing files.

1a. Detecting the backend LLM/agent runtime

• Python signals: requirements.txt, pyproject.toml, Pipfile, or *.py files → runtime python
• Node.js signals: a backend entry point (server.js, index.js, an Express/Next.js/Fastify app) → runtime nodejs
• If both exist (e.g. a Next.js app with a Python worker), instrument each backend runtime independently.
• If neither is present, there is nothing to instrument — stop and tell the user.

1b. Detecting the LLM framework/SDK (optional, informational)

Check dependency files for signals of: openai, @anthropic-ai/sdk / anthropic, langchain, langgraph, ai (Vercel AI SDK), boto3 + Bedrock usage, google-generativeai / google-genai, crewai, litellm, pydantic-ai, an MCP SDK, google-adk. This doesn’t change the init code (ddtrace/dd-trace auto-instruments these SDKs once the tracer is initialized) — it’s only used for confirming the setup with the user and for special-cased frameworks noted in the reference files (Next.js, Vercel AI SDK).

1c. Detecting the backend application framework

• Python: fastapi, flask, or django dependency
• Node.js: express dependency, or next (Next.js API routes / server actions)

1d. Detecting existing instrumentation, and auditing session-ID plumbing

• Grep for ddtrace init (LLMObs.enable(, ddtrace-run) in Python, or dd-trace init (require('dd-trace').init(, dd-trace/initialize) in Node.js.
• Also grep for a frontend RUM SDK (datadogRum.init(, @datadog/browser-rum) — not to set it up, but because its presence determines whether the opt-in RUM↔LLMObs pivot is available (that pivot reuses the RUM session ID as the LLMObs session ID).
• If LLMObs init is already present, do not add a second enable/init call — but do not stop there. SDK presence only means “don’t re-init”; it says nothing about whether a session ID actually flows. Run this audit whenever its prerequisite surfaces are present:
  • Session-ID plumbing — applicable whenever Phase 1a found an LLM/agent backend. Check whether a stable, conversation/operation-scoped session ID flows from the appropriate source: for web apps, the frontend sends a per-conversation ID and the backend request model/route reads it; for CLI/background jobs, the backend reuses an existing job/request/task ID or mints one UUID per invocation. A root agent/workflow span must then set it as session_id/sessionId (see “Beyond SDK init” and “Session ID intake by environment” in references/llmobs-python.md / references/llmobs-nodejs.md). LLMObs.enable()/dd-trace().init() alone does not establish this. If a RUM SDK is also present, the RUM↔LLMObs pivot is available as an opt-in — reusing the RUM session ID as the session_id — but that is a deliberate trade-off (the LLMObs session then spans the whole browser session), not the default; don’t flag its absence as a gap.
  • If the check finds a gap, tell the user what’s missing and fix it (same reference files, same ground rules) even though the SDK itself doesn’t need re-initializing. If it passes, say so explicitly. Note the RUM pivot as not applicable rather than a gap when there’s no RUM SDK.

1e. Detecting the dependency manifest to persist into

For each backend runtime found in 1a, identify the dependency manifest the build/deploy actually installs from. This is where any Datadog package you add (ddtrace/dd-trace) must be persisted so a clean deploy install includes it. It is not necessarily “whichever manifest file happens to exist”: a repo can have several (e.g. an empty requirements.txt alongside a pyproject.toml, or multiple package.json files where only one is the deployed workspace), and editing a non-authoritative one is a silent no-op at deploy time.

Resolve it in this order:

1. Deploy/build install command first (authoritative). Read the project’s build/deploy configuration and use the install source it names: render.yaml (buildCommand), Procfile, Dockerfile (RUN … install …), Makefile, CI workflows, package.json scripts. Examples: pip install -r <file> → that requirements file; poetry install / uv sync / pdm install → pyproject.toml (+ its lockfile); pipenv install → Pipfile; npm ci / yarn install / pnpm i → package.json.
2. Else, the highest-priority manifest present. Python: a lockfile-backed manager (uv.lock or poetry.lock → pyproject.toml) > pyproject.toml [project.dependencies]/[tool.poetry.dependencies] > requirements*.txt > Pipfile > setup.py/setup.cfg. Node.js: package.json (always).

Record the manifest path and the manager that owns it. Persisting commands (poetry add, uv add, pdm add, pipenv install, npm/yarn/pnpm/bun add) write the manifest. Non-persisting commands (pip install, uv pip install, conda install) touch only the current environment — with those you must also hand-edit the manifest. If requirements.txt is generated from a requirements.in (pip-tools), edit the .in and recompile. After hand-editing a manifest that has a lockfile, regenerate the lock so a frozen deploy install picks up the new package.

---

Phase 2: Instrumentation — routing table

Provision DD_API_KEY via references/common-credentials.md, then follow the reference for each backend runtime found in Phase 1a:

If Phase 1d found existing LLMObs instrumentation on a runtime, skip only that runtime’s init/enable call and credential provisioning — still act on any gap the Phase 1d session-ID plumbing audit found, using the same reference files.

---

Phase 3: Verification and Reporting

See references/common-verify-report.md for the verify step and the JSON report shape.