Imported from tracewayapp/traceway (skills/traceway-setup/SKILL.md). Install upstream with npx skills add tracewayapp/traceway --skill traceway-setup. Copyright stays with the author.

Set Up Traceway

Analyze the application first, agree on its Traceway project structure with the user, help them create that structure in the dashboard, and only then integrate and verify it.

Operating Contract

• Inspect before asking for credentials or changing code.
• Explain what was detected, what is already tracked, and how each production component should report to Traceway.
• Present a proposed project map and wait for the user to confirm it before implementation.
• There are two classes of credentials, treated differently:
  • Setup tokens (prefix tws_) are org-scoped, expire in 6 hours, and can only propose a setup plan; nothing is created until the user approves the plan on the Traceway website. They are designed to transit chat, and asking the user to paste one is expected.
  • Project ingest tokens and upload tokens are long-lived. Never print their values, never ask the user to paste them into chat, and never commit them. When Step 3's apply script (or the traceway CLI) writes an approved plan's ingest tokens into env files, read only the variable names and statuses from its output, never the values.
• Never print token values found in files or the environment. Report only that a value exists and the variable name that carries it.
• Never commit tokens or filled connection strings.

Invocation Forms

Every backend integrates with OpenTelemetry. There is one backend path, not one per language or web framework: Go, Node, Python, PHP, Java, .NET, Ruby, and everything else export OTLP/HTTP to <instance>/api/otel/*. The Traceway project is created with framework OpenTelemetry, which is the only backend option in the dashboard's framework picker. The native Traceway Go SDK is a deliberate exception used only when the user explicitly asks for it.

Fast Path

When all of the following hold, skip Steps 2 and 3 and go straight to the integration steps:

• the repository has exactly one deployable component, and
• the user supplied (or the environment already carries) an instance URL and a project ingest token (not a tws_ setup token, which exists to create projects), and
• the token's project already matches that component.

Still run Step 1, because it decides how to instrument. Collapse "Propose and Confirm the Project Map" to a single confirmation line ("This is a Go API; it reports to your existing <name> project over OpenTelemetry") and skip the project-creation step entirely. The full map-and-creation ceremony exists for repositories with more than one deployable component, or where no project exists yet. Do not make a user who handed you a token sit through a project-map review for a single service.

Step 1: Analyze the Architecture

Before changing anything, build a picture of what is deployed and what is already instrumented:

1. Frameworks and languages: detect them from package.json, go.mod, composer.json, requirements.txt/pyproject.toml, pubspec.yaml, build.gradle(.kts), Package.swift, *.xcodeproj/*.xcworkspace, Podfile, and source extensions. For iOS/Apple targets, note whether the sources are Swift or Objective-C. That choice picks the path in "Frontend and Mobile" (Swift gets the native SDK; Objective-C-only has none and falls back to OTel).

2. Deployable components and entry points: inventory APIs, SSR servers, browser apps, workers, schedulers, CLIs, and mobile applications. Treat code organization as evidence, not proof that something is deployed. Note the deployment target while you are here: Kubernetes manifests, a Helm chart, Kustomize overlays, Dockerfile plus a compose file, or cloud-init / Ansible / Terraform. It decides the host-metrics path in Step 8.

3. Existing observability: find OpenTelemetry configuration, Traceway SDKs, Sentry, Datadog, New Relic, Honeycomb, logging exporters, tracing middleware, source map or symbol upload steps, and observability environment-variable names. Explain whether Traceway will replace, coexist with, or extend each integration. Never display discovered credential values.

4. Production role of JS meta-frameworks (Next.js, SvelteKit, Remix, Nuxt): never assume full-stack.

   • Frontend-only signals: output: 'export' in next.config.* (static export, so there is no server in production); no API routes (app/api/**/route.*, pages/api/*) or only trivial ones; no 'use server' server actions; rewrites/proxy config or a NEXT_PUBLIC_API_URL-style env var pointing the browser at an external API; a separate backend service in this repo, another repo, or another language; static hosting in the deploy config (S3/CloudFront, GitHub Pages, nginx serving out/).
   • Server signals: API routes with real logic, 'use server' server actions, database clients imported in server code, SSR reading its own data layer, next start or standalone output in the deploy config.
   • Mixed or unclear: ask how the application is deployed and whether the framework server does production work worth tracking.

   Frontend-only means integrate ONLY the browser side with the frontend SDK; do NOT add OTel to, or otherwise instrument, the framework's server side. A separate backend is its own component with its own Traceway project.

5. Background work: find cron jobs, queue consumers, schedulers, CLI commands, and long-running workers. Record whether libraries already emit CONSUMER spans.

6. AI/LLM usage: check dependencies for openai, @anthropic-ai/sdk, anthropic, langchain / @langchain/*, ai (Vercel AI SDK), litellm, google-generativeai, cohere, openrouter, and agent frameworks (langgraph, crewai, autogen / ag2, pydantic-ai, @openai/agents, @mastra/*, llamaindex / llama-index, semantic-kernel). Then decide whether this is a conversational product (chatbot, assistant, agent) rather than one-shot LLM calls: look for chat or streaming-chat routes (/chat, /completions, SSE or websocket handlers feeding an LLM), message-history persistence (a messages/conversations table or thread ids), tool/function-calling definitions passed to the model, or any agent framework above. If conversational or tool-calling, read ai-agent.md in this skill directory before proposing the AI integration.

7. Browser build and release flow: identify the bundler, output directory, source-map settings, deploy pipeline, and existing artifact uploads.

8. Mobile build and release flow: determine whether Flutter is obfuscated, Android uses R8/minification, iOS produces dSYMs, and whether multiple platform directories are one cross-platform product or independently released apps.

9. Deployment and host metrics: inspect Dockerfiles, Compose, Kubernetes, Helm, Terraform, Ansible, cloud-init, PaaS configs, and CI/CD. Determine whether the backend runs on a host where the Traceway OTel Agent is applicable.

Present the findings before asking questions:

Use actual paths and findings. Include unresolved deployment assumptions explicitly.

Step 2: Propose and Confirm the Project Map

Explain these default boundaries:

The dashboard's framework picker offers exactly these nine options: OpenTelemetry for every backend, React / Svelte / Vue.js / jQuery for browsers, and Flutter / React Native / Android / iOS for mobile. There is no Gin, Django, Laravel, Next.js, or Remix entry, and none is needed. A Next.js or Remix browser project selects React, and its server side selects OpenTelemetry like any other backend. The framework-specific setup for a backend lives on the project's Connection page, which asks for the language and web framework after the project exists.

Do not create one project per backend process by default. Keeping the API, workers, AI calls, and server metrics together preserves their operational context and distributed traces. Split backend projects only for a real product, ownership, access-control, compliance, or data-isolation boundary.

Ask the user to confirm only what the repository cannot establish safely:

1. Which detected components are deployed and should be tracked?
2. Traceway Cloud or self-hosted, and which organization should own the projects?
3. Are the proposed names and project boundaries correct?
4. Does the backend run on a VM/host, Kubernetes, or serverless/PaaS, and should host metrics be collected?
5. Do mobile directories represent one cross-platform product or separate released apps?

Each confirmed row must also settle what Step 3's plan file needs:

• The framework value, one of the nine: opentelemetry, react, svelte, vuejs, jquery, flutter, react-native, android, ios.
• The env wiring: which untracked env file holds the credential (envFile) and under which name (envVar), following the credential-naming rules above (public prefixes like VITE_/PUBLIC_/NEXT_PUBLIC_ for browser values). envFormat is token when the code composes the connection string itself or the value feeds an OTLP Authorization header, and connectionString when the SDK init reads the full <token>@<instance>/api/report string straight from the variable.
• The deployment placement, from Step 1's deployment analysis: when the credential cannot be hardcoded into the repository because of how the component is deployed (Vercel/Fly/K8s/CI-injected env, secret managers), add a deployment block with the platform name and the exact command or UI path, using the literal placeholder <token> for the value. The website substitutes the real value after approval; never put real tokens in the plan. A mobile project whose credential lives in build config usually gets a deployment block and no envFile.

Do not modify code until the user confirms this map.

Step 3: Create the Projects

The default path: submit the confirmed map as a setup plan, the user approves it visually on the Traceway website, and a short curl script writes the resulting ingest tokens into env files. Nothing needs to be installed beyond curl and jq. Existing correctly mapped Traceway projects may be reused (the plan matches projects by name, so listing an existing name reuses it instead of duplicating).

3a. Get a setup token (skip when the invocation carried one)

Branch on account state; ask only if the repository and conversation do not establish it:

• Has a Traceway account: "Open <instance>/setup (Traceway Cloud: https://cloud.tracewayapp.com/setup). Log in if prompted; the page shows a setup token starting with tws_. Paste it here and keep the page open." The token expires in 6 hours; the page has a Generate New Token button.
• No account yet: "Open <instance>/register (Traceway Cloud: https://cloud.tracewayapp.com/register). Step 1 creates your account and organization. On step 2, keep 'AI' selected; it shows a setup token starting with tws_. Paste it here and leave that page open: my proposal will appear there for your approval, and the projects show up live once you approve."

Do not have the user create projects in the web UI on this path, and do not proceed without the token.

3b. Write the plan file

Write traceway-setup-plan.json from the confirmed map, one entry per project:

{
  "projects": [
    {"name": "Product Backend", "framework": "opentelemetry",
     "envFile": "backend/.env", "envVar": "TRACEWAY_BACKEND_TOKEN"},
    {"name": "Product Web", "framework": "svelte",
     "envFile": "frontend/.env", "envVar": "PUBLIC_TRACEWAY_WEB_CONNECTION_STRING",
     "envFormat": "connectionString",
     "deployment": {"platform": "Vercel",
       "instructions": "vercel env add PUBLIC_TRACEWAY_WEB_CONNECTION_STRING production\n# paste: <token>"}}
  ]
}

Rules: framework is one of the nine values from Step 2; every envFile must be untracked (create it and confirm it is gitignored first); envVar follows the credential-naming rules; deployment.instructions uses the literal <token> placeholder, never a real value.

3c. Submit the plan and wait for approval (curl)

Use curl and jq directly for this step. Do NOT install the traceway CLI for it; nothing gets installed on this path. (Only if a CLI new enough to have setup apply is already present is printf '%s' "$TRACEWAY_SETUP_TOKEN" | traceway setup apply --url "$TRACEWAY_URL" --plan traceway-setup-plan.json --token-stdin an equivalent one-step alternative.) If jq is unavailable and cannot be installed, use the manual fallback (3e).

Export the credentials first; setup tokens may transit chat, so this is fine:

export TRACEWAY_URL="<instance>"          # no trailing slash
export TRACEWAY_SETUP_TOKEN="tws_..."

Validate the token and show the organization (this response carries project names only, safe to display):

curl -s -w '\nHTTP %{http_code}\n' -H "Authorization: Bearer $TRACEWAY_SETUP_TOKEN" \
  "$TRACEWAY_URL/api/setup/session"

Submit the plan as a draft:

curl -s -w '\nHTTP %{http_code}\n' -X PUT \
  -H "Authorization: Bearer $TRACEWAY_SETUP_TOKEN" -H 'Content-Type: application/json' \
  --data-binary @traceway-setup-plan.json "$TRACEWAY_URL/api/setup/plan"

200 {"status":"pending"} means the draft is up; tell the user: "Review the proposal on the Traceway page you have open and press Approve Setup." A 422 is a validation message: fix the plan and resubmit. A 401 means the token is invalid or expired: ask the user for a fresh one from <instance>/setup.

Then wait for the decision with the script below. The script exists because the decision response carries the ingest tokens: values must flow API to env file without ever reaching stdout, your context, or the chat. Write it to traceway-setup-wait.sh exactly as given and run sh traceway-setup-wait.sh. Never fetch GET /api/setup/plan directly, and never read the env files it writes.

#!/bin/sh
# Waits for the setup decision, then writes each ingest token into the env
# file named by the plan. Prints statuses and variable names only.
set -eu
PLAN=${1:-traceway-setup-plan.json}
AUTH="Authorization: Bearer ${TRACEWAY_SETUP_TOKEN:?}"
URL=${TRACEWAY_URL:?}
deadline=$(($(date +%s) + 1800))
while :; do
  resp=$(curl -s -H "$AUTH" "$URL/api/setup/plan")
  status=$(printf '%s' "$resp" | jq -r '.status // "error"')
  case "$status" in
    approved) break ;;
    rejected)
      echo "rejected: $(printf '%s' "$resp" | jq -r '.reason // "(no reason given)"')"
      exit 2 ;;
    pending|none) ;;
    *) echo "unexpected response, the setup token may have expired"; exit 1 ;;
  esac
  if [ "$(date +%s)" -ge "$deadline" ]; then
    echo "timed out waiting for approval; rerun when the user is ready"
    exit 3
  fi
  sleep 3
done
i=0
n=$(jq '.projects | length' "$PLAN")
while [ "$i" -lt "$n" ]; do
  name=$(jq -r ".projects[$i].name" "$PLAN")
  envfile=$(jq -r ".projects[$i].envFile // empty" "$PLAN")
  envvar=$(jq -r ".projects[$i].envVar // empty" "$PLAN")
  format=$(jq -r ".projects[$i].envFormat // \"token\"" "$PLAN")
  pstatus=$(printf '%s' "$resp" | jq -r --arg n "$name" \
    '[.projects[] | select(.name == $n)][0].status // "missing"')
  if [ "$pstatus" = missing ]; then
    echo "$name: missing from the approved plan"
    i=$((i + 1)); continue
  fi
  if [ -n "$envfile" ] && [ -n "$envvar" ]; then
    value=$(printf '%s' "$resp" | jq -r --arg n "$name" --arg f "$format" \
      '[.projects[] | select(.name == $n)][0]
       | if $f == "connectionString"
         then .token + "@" + (.backendUrl | sub("/+$"; "")) + "/api/report"
         else .token end')
    dir=$(dirname "$envfile"); mkdir -p "$dir"
    tmp=$(mktemp "$dir/.tmp-env-XXXXXX")
    { [ -f "$envfile" ] && grep -v "^${envvar}=" "$envfile" > "$tmp"; } || true
    printf '%s=%s\n' "$envvar" "$value" >> "$tmp"
    mv "$tmp" "$envfile"
    echo "$name: $pstatus, wrote $envvar to $envfile"
  else
    echo "$name: $pstatus"
  fi
  i=$((i + 1))
done
echo "done: deployment credentials, if any, are shown on the Traceway page the user has open"

Env files come out mode 0600 via mktemp. Failure handling:

• Rejected (exit 2): the script prints the user's rejection reason. Revise the map with the user, update the plan file, submit again (a new PUT replaces the pending draft), and rerun the wait script. Already-approved projects are matched by name, never duplicated.
• Expired or invalid token (401 / exit 1): ask the user for a fresh token from <instance>/setup, re-export it, resubmit, rerun.
• Timeout (exit 3): resubmit and rerun when the user is ready; it is safe.

Delete the plan file and traceway-setup-wait.sh after a successful apply; they hold no secrets but they are setup litter.

3d. Confirm and handle deployment credentials

On approval the Traceway page celebrates and lands the user on the new project's dashboard, which doubles as the confirmation that the projects exist. For every project with a deployment block, the page first shows a Next Steps panel with the exact command and the real credential value; walk the user through completing it there before they continue, because the setup page is not reachable again once the account has projects. Never relay those values through chat; if a value is needed later, each project's Connection page in the dashboard carries its token.

3e. Manual fallback

When the user prefers clicking, curl or jq is unavailable, or apply keeps failing, follow dashboard-project-setup.md: the user creates each project in the dashboard UI and puts the tokens straight into their env files or secret store. On this path the old rule applies in full: ingest tokens never transit chat.

Upload tokens stay manual on every path: the uploaders read fixed variable names. traceway-sourcemaps reads TRACEWAY_SOURCEMAP_TOKEN, and dart run traceway:upload_symbols and the iOS dSYM script read TRACEWAY_UPLOAD_TOKEN (all three also read TRACEWAY_URL). Either name the CI secret exactly that, or keep a component-specific secret and pass it explicitly:

traceway-sourcemaps --url "$TRACEWAY_URL" --token "$TRACEWAY_WEB_UPLOAD_TOKEN" --directory ./dist

Inventing a name like TRACEWAY_WEB_UPLOAD_TOKEN and then calling the uploader with no --token fails at release time, not at setup time, so make the choice explicit in the CI step. Upload tokens are generated per project from Connection -> Source Maps or Symbol Upload in the dashboard.

Integration Paths

Pick the path by project type, and pick the project type from the "Analyze the Architecture" findings, never from the framework name alone (a Next.js repo can be a full-stack app or just the frontend of a separate backend). Per path, this is not negotiable per framework; it is how Traceway is designed to receive data:

The rules every backend integration must satisfy, and the table of what each span becomes, are in Step 4 where they are applied.

Step 4: Backend OTel Setup

The same shape in every language:

1. Install the language's OpenTelemetry SDK plus the auto-instrumentation for the web framework and database clients.
2. Point the OTLP/HTTP exporter at Traceway with the project token as a Bearer header.
3. Set service.name (becomes the Server Name in Traceway) and service.version (enables release comparison) on the resource.
4. Verify endpoint grouping before anything else.

Three rules that are not negotiable

1. Endpoints MUST arrive parametrized. http.route must be the route pattern (/api/users/:id), never the concrete URL. Traceway uses the value as-is, but only when it starts with /: a route name like app_user_show is discarded exactly like a missing value. With no usable route it falls back to url.path, and the Endpoints page explodes into one row per unique URL.
2. Background work MUST use SpanKind.CONSUMER. A root span with the default INTERNAL kind and no HTTP attributes is silently dropped (exceptions recorded on it still reach Issues, but the task run itself is lost).
3. The exporter MUST use an OpenTelemetry project's token. If OTLP arrives with a browser project's token (framework React, Svelte, Vue.js, jQuery, or React Native), Traceway keeps only the Issues and discards every endpoint, task, child span, and AI trace. The symptom is "errors appear but the Endpoints page is empty", and the cause is almost always two projects whose tokens got swapped.

How Traceway classifies spans

The rules are evaluated in this order and the first match wins.

Because the order is fixed, a CONSUMER span carrying gen_ai.* is a Task, and a SERVER request span carrying gen_ai.* is an Endpoint. Put the model call on its own child span (Step 6).

Two more rules that do not depend on span kind:

• An exception event, or exception.* attributes, on any span produces an Issue, whatever the span itself became.
• An endpoint that returns 404 with no real route matched (http.route missing, or a catch-all like /, /*) is renamed to the literal endpoint UNMATCHED, so bot scans and typo'd URLs collapse into one row. A concrete matched route returning 404 keeps its own name.

For the exact rules, endpoint naming, metric conversion, and the remaining quirks, read data-model.md in this skill directory. It is the authoritative reference.

Exporter configuration

Where the SDK supports the standard env vars, prefer them; they work identically across languages:

OTEL_SERVICE_NAME=my-service
OTEL_RESOURCE_ATTRIBUTES=service.version=1.2.3   # there is no OTEL_SERVICE_VERSION variable
OTEL_EXPORTER_OTLP_ENDPOINT=https://<instance>/api/otel
OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf        # required, several SDKs default to gRPC
OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer <project-token>"
OTEL_TRACES_EXPORTER=otlp
OTEL_METRICS_EXPORTER=otlp
OTEL_LOGS_EXPORTER=otlp

Two of those lines are load-bearing in a way that fails silently:

• OTEL_EXPORTER_OTLP_PROTOCOL is not optional. Left unset, the Python SDK and the Java agent resolve otlp to gRPC, and Traceway has no gRPC listener. The app starts, serves traffic, exits 0, prints no warning, and every export is lost. (PHP is the one exception to the value: it needs http/json.)
• There is no OTEL_SERVICE_VERSION. service.version only reaches Traceway through OTEL_RESOURCE_ATTRIBUTES. Without it, every endpoint row comes back with an empty App Version and release comparison does nothing.

SDKs append /v1/traces, /v1/metrics, /v1/logs to the endpoint automatically, so set the base URL only. A full signal path in OTEL_EXPORTER_OTLP_ENDPOINT produces /v1/traces/v1/traces, and the signal-specific variables (OTEL_EXPORTER_OTLP_TRACES_ENDPOINT) are used verbatim with nothing appended. When configuring in code instead, the full URLs are https://<instance>/api/otel/v1/traces (and /v1/metrics, /v1/logs) with header Authorization: Bearer <project-token>.

When Step 3's apply script wrote the env files, the plan's variables already hold real values locally, so wire the exporter to exactly those names (e.g. Authorization=Bearer ${TRACEWAY_BACKEND_TOKEN}) and run live verification immediately. The script writes only each project's one plan variable: companion variables the integration reads, like TRACEWAY_URL for the instance URL, are yours to add to the same env file now. On the manual fallback the user populates the variables first.

Constraints: OTLP/HTTP only (protobuf or JSON). OTLP/gRPC is not supported, there is no listener on port 4317. Content-Encoding: gzip is fine. The body is capped at 10 MB after decompression, and an oversized batch is rejected outright with 413 and {"error":"request body exceeds the 10MB limit"}. Nothing is ingested partially. A wrong or missing token answers 401, and any other path answers 404. All three are invisible from the application: most SDKs log exporter failures at debug level only, so when nothing arrives, turn the SDK's own diagnostic logging on first.

Node.js example

npm install @opentelemetry/sdk-node @opentelemetry/auto-instrumentations-node \
  @opentelemetry/instrumentation @opentelemetry/api @opentelemetry/sdk-metrics \
  @opentelemetry/exporter-trace-otlp-http @opentelemetry/exporter-metrics-otlp-http

Declare @opentelemetry/api yourself even though it also arrives transitively. Relying on hoisting breaks under pnpm's strict node_modules and under Yarn PnP.

Create instrumentation.mjs at the project root and load it before the app: node --import ./instrumentation.mjs server.js. Use the .mjs extension. A .js file holding import syntax fails to load in a CommonJS project, which is what most projects still are.

import { register } from "node:module";
register("@opentelemetry/instrumentation/hook.mjs", import.meta.url);

import { NodeSDK } from "@opentelemetry/sdk-node";
import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-http";
import { OTLPMetricExporter } from "@opentelemetry/exporter-metrics-otlp-http";
import { PeriodicExportingMetricReader } from "@opentelemetry/sdk-metrics";
import { getNodeAutoInstrumentations } from "@opentelemetry/auto-instrumentations-node";

const url = process.env.TRACEWAY_URL;
const headers = { Authorization: `Bearer ${process.env.TRACEWAY_BACKEND_TOKEN}` };

const sdk = new NodeSDK({
  serviceName: process.env.OTEL_SERVICE_NAME ?? "my-service",
  traceExporter: new OTLPTraceExporter({ url: `${url}/api/otel/v1/traces`, headers }),
  metricReaders: [
    new PeriodicExportingMetricReader({
      exporter: new OTLPMetricExporter({ url: `${url}/api/otel/v1/metrics`, headers }),
      exportIntervalMillis: 30_000,
    }),
  ],
  instrumentations: [getNodeAutoInstrumentations()],
});

sdk.start();

Three parts of that snippet are load-bearing:

• The two register lines install the ESM loader hook. Without them an ESM app ("type": "module", import express from "express") still gets HTTP spans but no http.route, so every URL becomes its own endpoint row. A CommonJS app (require("express")) is patched without the hook, and the lines are harmless there, so keep them either way.
• serviceName becomes the Server Name in Traceway. Leave it out and every span reports unknown_service:node. For release comparison also set OTEL_RESOURCE_ATTRIBUTES=service.version=1.2.3.
• metricReaders is a list. The older singular metricReader option is deprecated.

Auto-instrumentation covers Express routes (sets http.route), status codes, errors, and database clients (pg, mysql2, mongodb, ioredis). SQLite and custom business logic need manual tracer.startActiveSpan() child spans.

Next.js server exception

Next.js must not use the generic instrumentation.mjs preload above, node --import, or NODE_OPTIONS. Next.js creates its own incoming request root span with the matched http.route; preloading the generic Node HTTP instrumentation creates a competing root named from the literal URL, so /api/users/1 and /api/users/2 become separate endpoints.

For a Next.js server, create a minimal instrumentation.ts hook and put the SDK in a separate instrumentation.node.ts. Next.js compiles the hook for both Node and Edge; importing OTel packages directly in the hook, even with dynamic import() calls after a runtime guard, can make the development Edge compilation resolve Node built-ins and fail. The wrapper must conditionally import only the Node module:

export async function register() {
  if (process.env.NEXT_RUNTIME === "nodejs") {
    await import("./instrumentation.node");
  }
}

Keep every Node-only import and the SDK startup in the sibling module:

import { getNodeAutoInstrumentations } from "@opentelemetry/auto-instrumentations-node";
import { OTLPLogExporter } from "@opentelemetry/exporter-logs-otlp-http";
import { OTLPMetricExporter } from "@opentelemetry/exporter-metrics-otlp-http";
import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-http";
import { resourceFromAttributes } from "@opentelemetry/resources";
import { BatchLogRecordProcessor } from "@opentelemetry/sdk-logs";
import { PeriodicExportingMetricReader } from "@opentelemetry/sdk-metrics";
import { NodeSDK } from "@opentelemetry/sdk-node";

(Error as unknown as { prepareStackTrace?: unknown }).prepareStackTrace = undefined;

const instance = process.env.TRACEWAY_URL!.replace(/\/+$/, "");
const base = `${instance}/api/otel`;
const headers = { Authorization: `Bearer ${process.env.TRACEWAY_BACKEND_TOKEN}` };

new NodeSDK({
  resource: resourceFromAttributes({
    "service.name": process.env.OTEL_SERVICE_NAME ?? "nextjs-server",
    "service.version": process.env.APP_VERSION ?? "development",
  }),
  traceExporter: new OTLPTraceExporter({ url: `${base}/v1/traces`, headers }),
  metricReaders: [
    new PeriodicExportingMetricReader({
      exporter: new OTLPMetricExporter({ url: `${base}/v1/metrics`, headers }),
      exportIntervalMillis: 30_000,
    }),
  ],
  logRecordProcessors: [
    new BatchLogRecordProcessor({
      exporter: new OTLPLogExporter({ url: `${base}/v1/logs`, headers }),
    }),
  ],
  instrumentations: [getNodeAutoInstrumentations()],
}).start();

Install the log exporter and SDK packages in addition to the generic Node dependencies: @opentelemetry/exporter-logs-otlp-http, @opentelemetry/sdk-logs, @opentelemetry/api-logs, and @opentelemetry/resources. Next.js versions before 15 also need experimental.instrumentationHook: true. Full setup and verification: https://docs.tracewayapp.com/client/otel/nextjs

Resetting Error.prepareStackTrace in instrumentation.node.ts lets Node apply source maps instead of leaving server Issues pointed at minified Next chunks. Production must also start with NODE_OPTIONS=--enable-source-maps, and the build must emit .next/server/**/*.js.map; verify both rather than assuming readable server stacks. If the webpack build omits them, add experimental: { serverSourceMaps: true } to next.config. Also list @opentelemetry/auto-instrumentations-node in serverExternalPackages so Next does not bundle the instrumentation registry or warn about its optional transports.

Per-language notes

• Go: use the framework's OTel middleware. go.opentelemetry.io/contrib/instrumentation/github.com/gin-gonic/gin/otelgin, github.com/riandyrn/otelchi (a community package, and pass otelchi.WithChiRoutes(r) or it reports raw URLs), github.com/gofiber/contrib/otelfiber/v2. All three set http.route from the matched route pattern. Exporter: otlptracehttp.WithEndpointURL(...) + WithHeaders. Two Go-only traps:

  • Set the propagator. Go's global propagator is a no-op by default, so traceparent is never sent or read and cross-service traces never link. One line, right after otel.SetTracerProvider(tp):

    otel.SetTextMapPropagator(propagation.NewCompositeTextMapPropagator(
        propagation.TraceContext{}, propagation.Baggage{},
    ))

  • stdlib net/http does not report its route. otelhttp reads the route from r.Pattern, which ServeMux fills in only after it has matched, so the usual top-level otelhttp.NewHandler(mux, "server") records no http.route at all and every URL becomes its own endpoint row. Either wrap each registered handler, which puts the span start inside the mux:

    mux.Handle("GET /api/users/{id}", otelhttp.NewHandler(usersHandler, "users"))

    or keep the single top-level wrap and stamp the route yourself:

    // imports: strings, go.opentelemetry.io/otel/trace,
    //          semconv "go.opentelemetry.io/otel/semconv/v1.26.0"
    func withRoute(mux *http.ServeMux) http.Handler {
        return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
            if _, pattern := mux.Handler(r); pattern != "" {
                if i := strings.IndexByte(pattern, '/'); i >= 0 {
                    route := pattern[i:] // drop the "GET " prefix, http.route must start with "/"
                    trace.SpanFromContext(r.Context()).SetAttributes(semconv.HTTPRoute(route))
                }
            }
            mux.ServeHTTP(w, r)
        })
    }
    
    handler := otelhttp.NewHandler(withRoute(mux), "server")

• Python: pip install opentelemetry-distro opentelemetry-exporter-otlp, then opentelemetry-bootstrap -a install, then run the app under opentelemetry-instrument with the env vars above (opentelemetry-instrument uvicorn app:app, opentelemetry-instrument gunicorn wsgi:app, opentelemetry-instrument python worker.py). Starting the server without that prefix sends nothing. FastAPI and Flask instrumentation sets http.route correctly with zero application code, in the framework's own syntax (GET /users/{user_id}, GET /orders/<order_id>), and both frameworks already answer 500 on an unhandled exception, so the endpoint status and the Issue line up without extra work. Three Python-only gotchas:

  • Logs need two more variables. OTEL_LOGS_EXPORTER=otlp attaches the OTel handler to the root logger, whose level defaults to WARNING, so every logger.info(...) is filtered out before the bridge sees it and only WARN and above reach Traceway. Attaching the handler also replaces the root handler list, so the app's own records stop appearing on stdout. Add both:

    OTEL_PYTHON_LOG_CORRELATION=true   # restores console output, stamps otelTraceID/otelSpanID
    OTEL_PYTHON_LOG_LEVEL=info         # only read alongside the line above

    OTEL_PYTHON_LOG_LEVEL on its own does nothing. Do not set OTEL_PYTHON_LOGGING_AUTO_INSTRUMENTATION_ENABLED=true; on current versions the bridge is already on and that variable switches it to the SDK's deprecated handler.

  • Metrics take a minute. The SDK default for OTEL_METRIC_EXPORT_INTERVAL is 60000 ms, so a correct pipeline looks empty while you are checking it. Set OTEL_METRIC_EXPORT_INTERVAL=10000 during verification.

  • Django needs two extra things: export DJANGO_SETTINGS_MODULE or opentelemetry-instrument fails to start Django at all, and add the route middleware from the Django guide, because opentelemetry-instrumentation-django reports http.route as Django's own pattern (api/users/<int:user_id>/), which has no leading / and is therefore discarded.

  Guides: https://docs.tracewayapp.com/client/otel/python and https://docs.tracewayapp.com/client/otel/django

• PHP: Laravel via composer require keepsuit/laravel-opentelemetry open-telemetry/exporter-otlp php-http/guzzle7-adapter; Symfony via composer require traceway/opentelemetry-symfony open-telemetry/exporter-otlp php-http/guzzle7-adapter (the stock Symfony auto-instrumentation sets http.route to the route NAME, which Traceway discards; see data-model.md). PHP does not take OTEL_PHP_AUTOLOAD_ENABLED as a plain env var next to the block above. Laravel needs nothing extra, the keepsuit service provider starts the SDK. Symfony starts the SDK through open_telemetry.sdk.autoload_enabled: true in config/packages/open_telemetry.yaml, or through a real process environment variable set in php-fpm, the Dockerfile, or Apache. Never put OTEL_PHP_AUTOLOAD_ENABLED in .env: Dotenv reads it after Composer autoload, the bundle then skips starting the SDK, and every signal silently becomes a no-op. PHP also prefers OTEL_EXPORTER_OTLP_PROTOCOL=http/json, though http/protobuf works too and is only slower without ext-protobuf.

  • Symfony workers: never instrument messenger:consume. The bundle already excludes it. traces.console.excluded_commands defaults to ['messenger:consume', 'messenger:consume-messages'], and that key is a prototyped array node, so setting it in config/packages/open_telemetry.yaml REPLACES the default instead of adding to it. If you override it for any reason you MUST re-list both entries. A consumer runs in an endless loop, so its command span never ends, never exports, and swallows every child span for the life of the worker. Excluding the command does not silence the work inside it: the Messenger middleware still emits one CONSUMER span per message, and those are what become Tasks in Traceway. Check what actually resolved with bin/console debug:config open_telemetry traces.console.

• Java / .NET / anything else: the standard OTel agent or SDK with the env vars above works as-is.

• Full per-framework docs: https://docs.tracewayapp.com/client/otel

Verify endpoint grouping (do this first)

Hit a parametrized route a few times with different IDs and check the Traceway Endpoints page: you must see ONE row (GET /api/users/:id), not one row per ID. If you see raw IDs, the instrumentation is not setting http.route; fix that before continuing. Last resort, set it manually in a middleware that knows the matched route pattern (the value must start with /, or it is discarded):

import { trace } from "@opentelemetry/api";

trace.getActiveSpan()?.setAttribute("http.route", matchedRoutePattern);

Errors

Thrown errors must be recorded as exception events to appear as Issues. Auto-instrumentation handles uncaught errors; for caught-and-handled ones:

import { trace, SpanStatusCode } from "@opentelemetry/api";

const span = trace.getActiveSpan();
span?.recordException(error);
span?.setStatus({ code: SpanStatusCode.ERROR, message: error.message });

(Go: span.RecordError(err, trace.WithStackTrace(true)); the stack trace option is what produces the exception.stacktrace attribute.)

Return HTTP 500 when an exception happens. A request that throws or panics must respond with 500 and set the span status to ERROR. Never let it fall through to a 200/2xx. The common trap is a recover/panic middleware that reports the exception but lets the response writer keep its default 200: Traceway then records the endpoint transaction as a success, so the Endpoints page looks healthy while Issues fills up, and the exception becomes an unlinked island instead of correlating to a failed request (and, with distributed tracing, to the frontend call that triggered it). Always pair "an exception was recorded" with "respond 500 and set span status ERROR". Genuine client errors that are NOT bugs (validation 422, auth 401, not-found 404) keep their real 4xx status and should not be recorded as exceptions in the first place.

Go has a sharper version of the same trap. otelgin records the status only after c.Next() returns, which never happens during a panic unwind, so an unrecovered panic leaves the endpoint row at status 0 even though Gin's own recovery already returned 500 to the client, and the Issue arrives with a title but no stack trace. Register a recovery middleware after otelgin.Middleware so it runs inside the server span:

// imports: fmt, net/http, go.opentelemetry.io/otel/codes, go.opentelemetry.io/otel/trace
func otelRecovery() gin.HandlerFunc {
    return func(c *gin.Context) {
        defer func() {
            if r := recover(); r != nil {
                err, ok := r.(error)
                if !ok {
                    err = fmt.Errorf("%v", r)
                }
                span := trace.SpanFromContext(c.Request.Context())
                span.RecordError(err, trace.WithStackTrace(true))
                span.SetStatus(codes.Error, err.Error())
                c.AbortWithStatus(http.StatusInternalServerError)
            }
        }()
        c.Next()
    }
}

router.Use(otelgin.Middleware(serviceName))
router.Use(otelRecovery())

Logs

OTEL_LOGS_EXPORTER=otlp wires the exporter only. The application's own log calls still need a bridge, or the Logs page stays empty while traces and metrics look perfectly healthy.

• Node: add @opentelemetry/winston-transport, or the Pino or Bunyan instrumentation, so each record picks up trace_id and span_id.
• Python, PHP, Java, .NET: opentelemetry-instrument and the language agents bridge the standard logger automatically.
• Go: build an sdklog.LoggerProvider with otlploghttp.WithEndpointURL(...), register it with global.SetLoggerProvider(lp), and defer lp.Shutdown(ctx) or the last batch is lost on exit.

Emit from the request context (c.Request.Context() in Gin, r.Context() in net/http, the active context in Node). A line emitted from a background context carries no trace id and cannot be opened from the endpoint it belongs to. Traceway reads severity_text (or derives it from severity_number), the body, service.name, and the trace and span ids.

Go uses OpenTelemetry

There is no separate Traceway Go SDK. It was retired, and Go instruments with opentelemetry-go exactly like every other backend: go get go.opentelemetry.io/otel/sdk go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracehttp, then the same env vars as the block above. Host CPU and memory come from the Traceway OTel Agent (see "Deployment and Server Metrics"), not from the app.

Step 5: Background Tasks (Boundaries and Labeling)

If "Analyze the Architecture" found background work, instrument it as Tasks. The rules:

• Boundary: one Task = one execution of a unit of background work. A whole cron job run is one task. One queue message or job is one task. One CLI command invocation is one task. Per-item work inside a run (each email in a batch, each row in an import) is a child span of the task span, never a separate CONSUMER span.
• Do not double-wrap. If a library's auto-instrumentation already emits CONSUMER spans (Kafka, RabbitMQ, Symfony Messenger consumers), wrapping them again creates duplicate Task entries.
• Labeling: the span name IS the task name and the grouping key. Use a stable identifier like cleanup-expired-sessions or process-email-queue. Never embed job IDs, timestamps, or user IDs in the name; each unique name becomes a separate task group. Dynamic context (job ID, batch size) belongs in span attributes, where it shows on the task detail page without affecting grouping.
• The kind must be CONSUMER. A root span with the default SpanKind.INTERNAL is dropped silently; this is the most common reason "my cron job doesn't show up". CLI commands may alternatively be root INTERNAL spans with a console.command attribute (Laravel/Symfony console instrumentation does this).

import { trace, SpanKind, SpanStatusCode } from "@opentelemetry/api";

const tracer = trace.getTracer("my-app");

async function runScheduledJob() {
  await tracer.startActiveSpan(
    "cleanup-expired-sessions",
    { kind: SpanKind.CONSUMER },
    async (span) => {
      try {
        await doWork();
        span.setStatus({ code: SpanStatusCode.OK });
      } catch (error) {
        const err = error instanceof Error ? error : new Error(String(error));
        span.recordException(err);
        span.setStatus({ code: SpanStatusCode.ERROR, message: err.message });
        throw error;
      } finally {
        span.end();
      }
    }
  );
}

(Go: tracer.Start(ctx, "cleanup-expired-sessions", trace.WithSpanKind(trace.SpanKindConsumer)).)

Step 6: AI Traces

If "Analyze the Architecture" found AI/LLM dependencies, instrument the model calls. A span becomes an AI Trace when it carries at least one gen_ai.* attribute and matched none of the earlier classification rules. In practice that means: keep gen_ai.* off the HTTP route span (it becomes an Endpoint), keep it off CONSUMER spans (they become Tasks), and give the model call its own child span with SpanKind.CLIENT rather than INTERNAL, so a failed call that records exception.* attributes still produces its AI Trace. Since calls happen inside a request or task, that child span stays linked to its Endpoint or Task by trace ID.

Boundaries: one span per model call (one provider API request = one AI Trace row). A multi-step agent run is multiple spans sharing a stable trace.name (same labeling discipline as task names: no IDs or timestamps). For streaming, end the span when the stream finishes.

Conversational product? If "Analyze the Architecture" flagged a chatbot, assistant, or tool-calling agent, follow ai-agent.md in this skill directory instead of stopping at the table below. It covers gen_ai.conversation.id (multi-turn grouping), what value belongs in user.id (a stable end-customer id, never a session id), tool-call payload shapes, and sub-agents. Without those the Users tab stays empty, and the Conversations tab fills with one throwaway single-turn conversation per request, because Traceway falls back to the OTel trace id. That is worse than empty: it looks like the feature is working.

Attributes Traceway reads (all optional, set what is available):

Conversation content is also read from trace.input/trace.output (or span.input/span.output) when gen_ai.prompt/gen_ai.completion are absent, and missing total_tokens/total_cost are computed from the input + output values. Token counts must be integer attributes and costs must be double or integer attributes. A number sent as a string is silently dropped and stored as 0, with no error on the export.

import { SpanKind } from "@opentelemetry/api";

return tracer.startActiveSpan("chat-completion", { kind: SpanKind.CLIENT }, async (span) => {
  const response = await openai.chat.completions.create({ model: "gpt-4o", messages });
  span.setAttributes({
    "gen_ai.system": "openai",
    "gen_ai.request.model": "gpt-4o",
    "gen_ai.usage.input_tokens": response.usage?.prompt_tokens ?? 0,
    "gen_ai.usage.output_tokens": response.usage?.completion_tokens ?? 0,
    "trace.name": "support-agent",
  });
  span.end();
  return response;
});

Zero-code path for OpenRouter users: in OpenRouter Settings -> Observability, add an OpenTelemetry Collector destination pointing at https://<instance>/api/otel/v1/traces with header {"Authorization": "Bearer <project-token>"}. Docs: https://docs.tracewayapp.com/client/openrouter

Step 7: Frontend and Mobile

Frontend and mobile projects do NOT use OTel; they use the Traceway SDKs reporting to /api/report with connection string <project-token>@https://<instance>/api/report.

Browser (React / Vue / Svelte / jQuery / plain JS), three pieces, all expected:

1. SDK: npm install @tracewayapp/react (or vue, svelte, jquery, frontend for plain JS) and initialize with the connection string (React: wrap the app in <TracewayProvider connectionString="...">). Captures errors, web vitals, and session replay.
2. Bundler plugin: npm install -D @tracewayapp/bundler-plugin, then add tracewayDebugIds() from @tracewayapp/bundler-plugin/vite (or /rollup, or TracewayDebugIdsWebpackPlugin from /webpack) to the bundler config, with source maps enabled (build.sourcemap: true / devtool: "source-map").
3. Source map upload: npm install -D @tracewayapp/sourcemap-upload, then run traceway-sourcemaps --url <instance> --token <source-map-upload-token> --directory ./dist as a postbuild or CI step (env vars: TRACEWAY_URL, TRACEWAY_SOURCEMAP_TOKEN). The upload token comes from the browser project and is a CI secret, never committed.

For the per-framework init code (plain JS, React, Vue, Svelte/SvelteKit, jQuery), the shared SDK options, error filtering, custom attributes, distributed tracing, and the full debug-ID + source map pipeline, read frontend-js.md in this skill directory. Online docs: https://docs.tracewayapp.com/client/react (or vue, svelte, jquery, js-sdk).

Full-stack JS (Next.js, SvelteKit, Remix): only when "Analyze the Architecture" confirmed the framework's server actually serves the app in production. A frontend-only deployment gets just the browser pieces above; a separate backend follows "Backend OTel Setup". When it is genuinely full-stack, integrate both sides under the two confirmed projects: server side follows "Backend OTel Setup" with the backend project's token, and browser side follows the three pieces above with the frontend project's token. For Next.js specifically, use the instrumentation.ts framework hook in the "Next.js server exception" above; never apply the generic Node preload to it.

For the Next.js browser build, @tracewayapp/bundler-plugin currently has a webpack entry point, not a Turbopack one. Set productionBrowserSourceMaps: true, add TracewayDebugIdsWebpackPlugin only when !isServer && !dev (the plugin is production-only and must not replace Next's development source-map mode), and make the production build run next build --webpack (Next.js 16 defaults to Turbopack, which otherwise ignores this plugin). After building, confirm client .js files contain //# debugId= and their .js.map siblings contain debugId, then upload .next/static/chunks with the React project's source-map upload token.

Mobile, always the platform SDK, never OTel:

• Flutter: flutter pub add traceway, then Traceway.run(connectionString: '<token>@https://<instance>/api/report', child: MyApp()). Then check whether the release build is obfuscated (--obfuscate --split-debug-info): if it is, production crash stack traces arrive obfuscated and stay unreadable until the build's .symbols files are uploaded, so use the mobile project's upload token and wire up the symbol upload. For options, platform permissions, the navigator observer, screen recording, privacy masking, the obfuscation check and symbol upload, and the Flutter web caveat, read flutter.md in this skill directory. Docs: https://docs.tracewayapp.com/client/flutter
• React Native: npm install @tracewayapp/react-native, wrap the app in TracewayProvider. Docs: https://docs.tracewayapp.com/client/react-native
• Android (native Kotlin/Java): add implementation("com.tracewayapp:traceway:1.0.1") from Maven Central and call Traceway.init(application = this, connectionString = "<token>@https://<instance>/api/report", options = TracewayOptions(version = "1.0.0")) from Application.onCreate() (register the Application class in the manifest; no permission entry is needed, the AAR's own manifest declares INTERNET and ACCESS_NETWORK_STATE and the merger folds them in). It captures every uncaught Java/Kotlin exception on every thread plus manual Traceway.captureException(...); errors and crashes only, no session replay. Release builds run R8, so production crashes arrive with renamed classes and rewritten line numbers and stay unreadable until the build's mapping.txt is uploaded: apply the com.tracewayapp.symbols Gradle plugin (version "1.0.1", resolved from mavenCentral()), which injects BuildConfig.TRACEWAY_PROGUARD_UUID and uploads the mapping with the mobile project's upload token; pass that UUID into TracewayOptions(proguardUuid = ...) so each crash matches its mapping. For init code, options, the manifest, and the full Gradle plugin setup, read android.md in this skill directory. Docs: https://docs.tracewayapp.com/client/android
• iOS / Swift (native SwiftUI or UIKit): add the Traceway iOS SDK via Swift Package Manager (https://github.com/tracewayapp/traceway-ios.git) and call Traceway.start(connectionString: "<token>@https://<instance>/api/report", options: TracewayOptions(version: "1.0.0")) as early as possible. It captures uncaught NSExceptions and fatal signals (hard crashes upload on the next launch) plus manual Traceway.capture(...); it reports errors and crashes only (no session replay). Release crashes arrive as bare addresses until the build's dSYMs are uploaded, so set up dSYM upload with the mobile project's upload token. For init code, options, the debugger caveat, and dSYM upload, read ios.md in this skill directory. If the app is NOT a Swift app (Objective-C only, a cross-platform stack with no Traceway mobile SDK, or a team standardized on OpenTelemetry), there is no native SDK: use an OTel distribution like Honeycomb with its exporter pointed at <instance>/api/otel and a Authorization: Bearer <project-token> header, exactly like a backend (see "Backend OTel Setup"). The non-Swift path is also in ios.md.

Step 8: Deployment and Server Metrics

Use the deployment and host-metrics choice confirmed in "Propose and Confirm the Project Map". If it remains unresolved, ask before changing deployment configuration:

1. How is this project deployed? Docker on a VM / directly on a VM or bare metal / Kubernetes / serverless or PaaS.
2. Do you want server (host) metrics tracked in Traceway? CPU, memory, disk, filesystem, network of the machine running the app.

On Kubernetes, stop here and read kubernetes.md in this skill directory. The four things that decide whether it works (service.name per node, the three *.utilization metrics that ship disabled, k8s.cluster.name, and root_path: /hostfs) are all easy to miss, and getting any of them wrong produces a plausible-looking config that reports nothing useful.

The agent is a tiny pre-configured OTel Collector that scrapes host metrics every 60s. It MUST use the same backend project token as the API, workers, tasks, and AI traces. Install it on the host (Linux systemd / macOS launchd; PowerShell installer exists for Windows):

curl -fsSL https://install.tracewayapp.com/install.sh | \
  TRACEWAY_TOKEN=<project-token> \
  TRACEWAY_ENDPOINT=https://<instance>/api/otel \
  TRACEWAY_SERVICE_NAME=<host-label, e.g. api-prod-eu-1> \
  bash

• TRACEWAY_ENDPOINT ends in /api/otel and is required for self-hosted instances; omit only for Traceway Cloud.
• Optional: TRACEWAY_LOG_PATHS (comma-separated globs to tail as logs), TRACEWAY_PROCESS_NAMES (per-process metrics).
• Re-running the installer upgrades in place, so the command is safe to keep in provisioning scripts.
• If the repo has host provisioning or deploy scripts (cloud-init, Ansible, Terraform user_data, deploy.sh), add the command there with the token referenced from a secret. Otherwise hand the operat

Truncated - read the full file at https://github.com/tracewayapp/traceway/blob/943dc67360831fe59c60e0b8cd6ab3c5531e1fab/skills/traceway-setup/SKILL.md.