Inngest Brownfield Audit

Use this skill when asked to inspect an existing codebase, add Inngest "where it makes sense", migrate fragile background work, or find durability gaps before making changes.

This is an agent-first workflow. Do the audit from evidence in the repo, name the specific files and call sites that drove each conclusion, and make small integration moves that preserve current behavior.

When to Trigger

Use this skill for requests like:

• "Audit this repo for Inngest opportunities"
• "Add Inngest to this codebase"
• "Make our webhooks / cron jobs / background tasks reliable"
• "Find places where work can be lost on deploy or process crash"
• "Replace fragile polling, delayed jobs, or fire-and-forget promises"
• "Make this AI workflow / agent durable"

If the user is starting from scratch instead of a brownfield repo, use inngest-setup, inngest-durable-functions, inngest-events, inngest-steps, and, for AI workflows, the agent patterns in this skill. Use inngest-agent-evals when the request includes scoring, sessions, experiments, deferred scorers, Insights, or outcome-based evaluation.

Audit Loop

1. Map the project shape.

   • Read package.json, workspace files, app/router structure, server entry points, deployment config, and test scripts.
   • Identify framework: Next.js App Router, Next.js Pages Router, Express, Hono, Fastify, Remix, SvelteKit, Astro, NestJS, worker-only service, or other.
   • Detect package manager and TypeScript conventions before adding files.

2. Find existing Inngest usage.

   • Search for inngest, createFunction, serve(, /api/inngest, INNGEST_, step.run, step.sleep, step.waitForEvent, step.sendEvent, step.invoke, step.ai, inngest.send, and @inngest/realtime.
   • If Inngest exists, inspect version, client config, serve endpoint, registered functions, event naming, env vars, and v3/v4 API shape before changing anything.

3. Find durability gaps.

   • Search for fire-and-forget work: void someAsync(), un-awaited promises, .then( chains, setTimeout, setInterval, detached jobs after HTTP response, and background work in route handlers.
   • Search for cron and schedulers: cron, node-cron, agenda, bull, bullmq, bee-queue, qstash, sqs, temporal, trigger.dev, deployment cron config, and scheduled API routes.
   • Search for webhooks and at-least-once producers: Stripe, Clerk, GitHub, Slack, Shopify, HubSpot, Linear, Svix, and generic webhook.
   • Search for long-running work: PDF generation, exports, video/image processing, embeddings, bulk email, imports, ETL, sync jobs, polling loops, retries, and external API calls.
   • Search for AI agent shapes: tool loops, LLM calls, streaming tokens, human approval, multi-step reasoning, vector search, eval loops, scoring, experiment assignment, user-feedback signals, and provider calls that need rate limits or retry-safe state.

4. Classify each candidate.

   • P0: user-visible loss, duplicate charge/email/action, timeout, missed webhook, or crash-prone workflow.
   • P1: fragile but recoverable background work, manual retry burden, noisy 429s, or poor observability.
   • P2: cleanup, ergonomics, or future migration opportunity.
   • For each candidate, record: file, current trigger, side effects, idempotency key, failure mode, recommended Inngest primitive, migration size, and confidence.

5. Choose the smallest safe integration.

   • Prefer one vertical slice over a wide rewrite.
   • Keep existing domain functions and data models where possible.
   • Add an Inngest client and serve endpoint only once.
   • Move side effects into step.run one boundary at a time.
   • Make event IDs and database writes idempotent before adding retries.
   • Add tests around existing behavior and the new event/function boundary.

Useful Discovery Commands

Run commands that fit the repo. Prefer rg; keep output focused.

rg -n "inngest|createFunction|step\\.|serve\\(|/api/inngest|INNGEST_" .
rg -n "setTimeout|setInterval|Promise\\.all|void [a-zA-Z0-9_]+\\(|\\.then\\(" .
rg -n "cron|node-cron|schedule|bull|bullmq|bee-queue|agenda|qstash|sqs" .
rg -n "webhook|stripe|svix|clerk|github|shopify|slack|hubspot|linear" .
rg -n "retry|backoff|poll|status|timeout|429|rate limit|rate-limit" .
rg -n "openai|anthropic|ai\\.|generateText|streamText|tool|agent|embedding" .

When the repo is large, narrow searches to app source directories and exclude generated/vendor folders.

Brownfield Decision Matrix

Existing shape	Inngest fit	Primary primitives
HTTP handler does slow side effects before responding	Emit event, return fast	inngest.send, event trigger, step.run
Webhook must acknowledge quickly but process reliably	Verify signature, emit idempotent event	Event ID, step.run, retries
Cron job loses progress midway	Cron-triggered durable function	Cron trigger, page-level step.run, flow control
Polling loop waits for external async work	Durable wait or durable poll	step.waitForEvent, step.sleep, step.run
Large fan-out exceeds request/serverless limits	Split orchestration and item work	step.sendEvent, per-item function, concurrency
External API hits 429s	Move limits to function config	throttle, rateLimit, concurrency
Human review can take days	Persist the wait in Inngest	step.waitForEvent, timeout, realtime
AI agent/tool loop needs retry-safe progress	One step per tool/model boundary	step.ai, step.run, step.sleep, realtime
AI workflow needs production evals	Attach outcome signals to durable runs	meta.sessions, step.score, createScorer, defer, group.experiment
Existing queue only hides fragile work	Replace queue boundary gradually	Event trigger, idempotency, function-level retries

Integration Plan Format

Before editing, summarize findings in this compact shape:

Inngest audit:
- Existing Inngest: none / partial / healthy / risky
- Framework: <framework and evidence>
- Best first slice: <file + workflow>
- Why: <loss/timeout/retry/idempotency failure>
- Proposed primitives: <event, steps, flow control, waits, realtime>
- Idempotency key: <source of truth>
- Files likely touched: <short list>
- Tests/checks: <commands or focused cases>

Then implement unless the user asked for audit-only.

Existing Inngest Checklist

If Inngest is already present, verify:

• A single shared client is exported from a stable module.
• The app id is a stable slug and is not derived from deploy-specific data.
• v4 local development uses INNGEST_DEV=1; production uses INNGEST_SIGNING_KEY.
• Serve endpoint path is discoverable, usually /api/inngest.
• The serve handler registers all functions that should sync.
• Side effects and non-deterministic work are inside steps.
• Step IDs are stable and descriptive.
• Event names follow domain/noun.verb.
• Events that may be replayed use deterministic IDs.
• Webhook handlers verify signatures before emitting events.
• Flow control is configured where external APIs have limits.
• Realtime uses v4 native inngest/realtime, not the v3 @inngest/realtime package.

Durable Agent Patterns

Use Inngest when an AI or agent workflow needs durable progress across model calls, tool calls, waits, approvals, or streaming UI updates.

Good candidates:

• Multi-step agent that calls tools or external APIs.
• LLM workflow that may exceed one HTTP request lifetime.
• Human-in-the-loop review, approval, correction, or escalation.
• Agent that must pause for an external event or scheduled follow-up.
• Bulk AI work that needs provider-level rate limits and cost protection.
• User-visible agent progress that should stream from durable execution.

Recommended shape:

1. HTTP/UI request stores the user intent and emits an event with a stable id.
2. Inngest function loads state inside step.run.
3. Each model call, tool call, vector search, and external side effect lives in its own step.ai or step.run boundary.
4. Human pauses use step.waitForEvent or step.waitForSignal with a timeout.
5. Progress updates use step.realtime.publish between steps, or inngest.realtime.publish inside an existing step.run.
6. Provider rate limits use concurrency, throttle, or rateLimit, not ad hoc in-process throttlers.

Avoid:

• Keeping agent state only in memory.
• Retrying whole agent loops after a single tool failure.
• Charging for repeated successful model calls because the result was not memoized.
• Using setTimeout or a cron poller for follow-ups and approvals.
• Streaming progress from a process-local WebSocket server when the workflow itself is durable elsewhere.

Implementation Guardrails

• Do not replace working queues, crons, or webhooks blindly. First preserve behavior with a thin Inngest slice.
• Do not create duplicate clients or serve endpoints if the repo already has them.
• Do not put database writes, API calls, random IDs, timestamps, or LLM calls outside steps in the new function.
• Do not hide missing idempotency behind retries. Retries require idempotent side effects.
• Do not hardcode secrets or dev-mode flags in source.
• Do not leave the app unable to sync: register new functions with the serve endpoint and run available type/tests.

Verification

Pick checks that prove the integration path:

• Typecheck/build/lint the touched app.
• Run existing tests around the migrated handler or workflow.
• Add focused tests for "handler emits event and returns fast" and "function calls the same domain operations in step boundaries" where the repo supports it.
• If local runtime is available, start the app and Inngest dev server, confirm the function syncs, then send a sample event.
• If only static checks are available, explicitly state that runtime sync was not verified.