Home / Add-ons / Migration Service
Migration service — move providers without breaking production
A compatibility review, a shadow-traffic run and a quality diff before cutover, so you change the base URL with evidence instead of hope.
Who it is for
Teams moving from another provider and unwilling to discover the differences in production.
The problem it removes
- The new provider returns different output for the same prompt.
- Streaming, tools and JSON mode behave differently.
- Nobody has a rollback plan if quality drops.
What you get
- Compatibility review: models, parameters, streaming, tools, JSON mode
- Shadow traffic: mirror requests, compare outputs
- Quality diff report with the disagreements listed
- Cutover and rollback plan
- One week of post-cutover monitoring notes
How to use it
1. Scope call
We list what your code sends today and what must keep working.
2. Shadow run
Mirrored traffic for a fixed window; you get the diff, not an opinion.
3. Cutover
You switch with a rollback step written down.
What the compatibility review checks
Two providers can both claim an OpenAI-compatible endpoint and still differ in nine places that break code. The review lists each one against your actual request shape rather than against the provider's documentation, because documentation describes what the API accepts, not what your code sends.
| Surface | What usually differs | What the review records |
|---|---|---|
| Chat completions | Message roles accepted, system-message handling | Whether every message your code sends is accepted unchanged |
| Streaming | Chunk shape, keep-alive behaviour, termination signal | Where your parser would need a branch |
| Tool calls | Argument encoding, parallel calls, finish reasons | The exact shape of a tool response on each side |
| JSON mode | Schema enforcement, malformed-output behaviour | Whether invalid JSON is retried or returned |
| Token counting | Tokenizer differences change prompt cost | Measured token delta on your own samples |
| Error codes | Status codes for the same failure | A mapping so retry logic stops guessing |
| Rate-limit headers | Names, units, reset semantics | The header your client should read |
| Embeddings | Dimensions and normalisation | Whether an existing index must be rebuilt |
| Context limits | Maximum input and output | Which prompts must be shortened before cutover |
How shadow traffic works without production access
Shadow traffic does not require credentials. You export a sample of representative requests — 200 to 2,000, redacted — and we replay them against the new endpoint with the same parameters. The two response sets are then compared offline. Your production system is never involved, and no customer data leaves your environment unless you choose to send it.
The sample matters more than the size. A thousand requests from one customer proves one customer's workload works. Two hundred requests spread across your input classes — short and long, structured and free-form, tool-using and plain — finds the differences that matter. We ask for the spread, not the volume.
Replay is run at a fixed temperature where the parameter is supported, because comparing outputs sampled at random measures noise. Where determinism is not available, each case is run three times and the comparison uses the distribution rather than a single pair.
What the shadow run needs from you
Three things, none of them credentials. A request export in JSONL with one request object per line, a note on which fields are sensitive so they can be redacted before they leave your environment, and a list of the input classes you consider important. The export is usually produced by a logging middleware you already have; if you do not log requests today, the review includes a short patch that turns logging on for a sample window and off again afterwards.
Redaction happens on your side. Names, account numbers and free text can all be replaced with fixed tokens without changing the shape of the request, and shape is what the compatibility review measures. If a redacted request behaves differently from the original, that is itself a finding about how much the model depends on the content.
What the quality diff report contains
- Pass rate per input class — not one aggregate number that hides a failing class.
- Every disagreement, listed — truncated outputs, refusals, changed formatting, dropped tool calls, different JSON shapes.
- Severity per disagreement — cosmetic, needs a parser change, or changes the answer.
- Token delta — measured on your samples, because a cheaper per-token price can lose to a longer tokeniser.
- Latency distribution — median and p95 per class, measured from the shadow run, not quoted from a marketing page.
- Recommendation — migrate, migrate with changes, or stay. A report that always recommends migrating is a sales document.
Reading the diff without overreacting
Not every difference is a defect. The report separates three kinds, and only the third should stop a cutover.
| Kind | Example | Verdict |
|---|---|---|
| Cosmetic | Different whitespace, a reordered JSON key, a different opening phrase | Ignore if your parser is stable |
| Structural | Tool arguments as a string instead of an object; a missing finish reason | Patch the parser before cutover |
| Semantic | A refusal where the old model answered; a truncated answer; a wrong extraction | Do not cut over for that input class |
Most failed migrations are structural differences discovered in production because the evaluation only compared prose. The report exists to move that discovery into a spreadsheet.
The rollback plan
Rollback is written before cutover, not after something goes wrong. It has five steps and a named owner for each.
- One switch — the base URL and key live in a single configuration value, not scattered across services.
- A named trigger — the error rate or quality threshold that means revert, agreed in advance rather than argued about during an incident.
- A time box — how long the old credentials stay valid after cutover. Thirty days is typical.
- A data check — what to compare after rollback to confirm the old path still works.
- A written record — what was tried, what failed, and what the next attempt should change.
When migration is not worth it
If your usage sits comfortably inside your current plan, the saving from a cheaper per-token price is usually smaller than the engineering time the migration consumes. The same is true if your application depends on a feature the new provider implements differently — a migration that requires rewriting your tool-call layer is a project, not a switch.
The review will say this plainly when the numbers point that way. A $250 report that recommends staying is a successful engagement; the alternative is discovering the same thing after two weeks of engineering.
Timeline
| Day | Activity | Output |
|---|---|---|
| 1 | Scope call and request export | Written scope, sample received |
| 2–3 | Compatibility review | Surface-by-surface findings |
| 4–6 | Shadow replay and comparison | Raw comparison data |
| 7–8 | Report writing | Quality diff, rollback plan |
| 9–14 | Your cutover window | Post-cutover notes for one week |
What we do not do
- No production access. The shadow run uses an export, not your live traffic.
- No code changes in your repository. Patch guidance and a diff, applied by your team.
- No guarantee of identical output. Different models produce different text; the report tells you where that matters.
- No migration of stored embeddings unless the dimensions match. A rebuild is a separate job and the report says when one is required.
Questions
What does it cost?
$250 one-time. The shadow run is included.
How long does it take?
One to two weeks depending on how many models and features you use.
Do you need production access?
No. We need representative requests, not credentials.
What if quality is worse?
The diff shows where. You can stop before cutover; the report is still yours.
Does it include code changes?
We provide the patch guidance; your team applies it.
Sources
- OpenAI chat completions reference — checked 21 November 2026
- Plugsky model catalogue — checked 21 November 2026
- Plugsky pricing — plan limits and fair use — checked 21 November 2026
Related add-ons
Guided Onboarding
Two live sessions with a Plugsky engineer: keys, first integration, an evaluation set and a cost baseline. Delivered in one week.
Compliance Audit
A scoped technical assessment: what applies to you, what is missing, and the order to fix it. A report and an evidence appendix, not legal advice.
One intake form, then two sessions in the same week. Paid once through Stripe; nothing renews.
Book the engagementCompare add-ons