@stigg/agent-sdk-js 0.0.0-stage → 1.0.1

package/README.md

@@ -1,3 +1,1148 @@
-# Temporary Holding Version
+# `@stigg/agent-sdk-js` — agent governance for the agent stack (TypeScript)
 
-This version is a temporary placeholder for this package. An operational version to replace this has been submitted for review and is awaiting a staged release.
+A thin SDK that gates and meters agent runs against [Stigg](https://stigg.io), plus runnable
+examples for the two layers it binds to: a client you wrap, and a gateway you call or operate.
+The examples are not part of the published package - they live beside the source in the Stigg
+SDK monorepo, and the paths below are relative to it.
+
+```bash
+npm install @stigg/agent-sdk-js      # or: yarn add @stigg/agent-sdk-js
+```
+
+Requires Node 22 or later. Ships CommonJS with type declarations, and a `./internal` subpath for
+the engine seam.
+
+Everything here runs **with no API keys and no Stigg account**: a built-in mock stands in for
+both the LLM vendors and the Stigg API.
+
+## What the SDK does
+
+One core lifecycle, several bindings, one swappable engine.
+
+```
+construct  →  register?  →  wrap  →  [ check? → execute → ingest ]  →  deregister?
+    |            |           |
+ per env,    optional:     the boundary you declare; the client goes through
+ one SDK     a write       Stigg, and check and ingest happen inside it
+             about the
+             agent
+```
+
+**There is no run to open.** What governs a call — who it is for, what gates it, what it reports —
+is a `stiggContext`, given once at `wrap` as the default for that client and again inside any
+single call's own parameters to override it. One wrapped client therefore serves every customer
+and every agent, which is what a multi-tenant server actually needs.
+
+`check` happens when the context names an `entitlement`. Omit it and the call meters only: the
+request still goes out, and it is still billed.
+
+`register` and `deregister` are optional. `register` hands back the agent id it was given — a
+convenience for the next line, not a handle the rest depends on. Skip them and the agent has no
+entity — calls still work, and checks and usage apply to the customer. That is deliberate:
+**wrapping a client must never depend on governance being available**, and a governance-only
+request (a budget, a parent) is the only thing that fails without it.
+
+- **The agent is the entity; attribution is a dimension.** An entity upsert is idempotent and
+  cheap, but it is still a write — which is why provisioning is `register()` and never something a
+  gated call carries. Naming an `agentId` in a context attributes to that agent; it does not
+  create it.
+- **A denial is a returned decision, not an exception** — it names which entity blocked and what
+  to do about it. Refusing is a commercial event (a top-up, an upgrade), not an incident. That is
+  `client.stigg.check()`. A **wrapped** call has nowhere to return one — the vendor's signature
+  owns the return type — so there it is raised as `AccessDeniedError`, and `stigg.denial(err)`
+  turns it back into the decision. A context that named no `entitlement` never gets there: it has
+  nothing to check against, so every call goes out and `check()` hands back an allow whose
+  `source` is `'unenforced'`.
+- **Metering never breaks the call.** Ingest failures are swallowed into the `onError` hook;
+  enforcement failures are never swallowed. They are deliberately separate code paths.
+
+```ts
+import { StiggAgentSDK } from '@stigg/agent-sdk-js';
+
+// One per environment. The key can be left out entirely to read STIGG_SERVER_API_KEY, or
+// replaced by a `{ baseUrl }` target that needs none - a local stand-in, or a self-hosted Stigg.
+// A key for somewhere other than the hosted API names that URL second:
+//   new StiggAgentSDK(apiKey, 'https://stigg.internal')
+// Everything after that is tuning, and every bit of it has a working default.
+const stigg = new StiggAgentSDK(apiKey);
+
+// Provisioning the agent: a write, so it is its own call. A *bare* registration is remembered per
+// process and skipped the second time; this one carries a parent and a budget, and a call that
+// changes placement always writes - two differing instructions cannot be collapsed into one.
+await stigg.register({
+  customerId: 'customer-acme',
+  agentId: 'agent-support-bot',
+  entityTypeId: 'agent',
+  entitlement: { currencyId: 'credits' },
+  parentId: 'team-eng',
+  usageLimitOnCreate: 5,
+  cadence: 'P1M',
+});
+
+// Governs the client and hands back a COPY whose calls are checked and metered. The client you
+// passed keeps working ungoverned.
+//
+// Wrapped with no customer, this client serves every tenant the process handles, and each call
+// names the one it is for. The types enforce that: `customerId` is **required** on every call
+// through a client wrapped this way.
+const govClient = stigg.wrap(new Anthropic(), {
+  stiggContext: {
+    // Naming an entitlement is the enforcement switch. Omit it and the calls meter without
+    // gating - every call still goes out, and nothing is ever refused.
+    entitlement: { currencyId: 'credits' },
+    // What the calls are attributed to. The type is required with it - a guessed attribution key
+    // resolves no entity, and the spend would land on the customer looking exactly like a call
+    // that was never governed. The parent needs no key: governance walks the chain up from here.
+    agentId: 'agent-support-bot',
+    agentType: { entityTypeId: 'agent', attributionKey: 'agentId' },
+    // Optional extra attribution, sent verbatim on every check and ingest.
+    // dimensions: { tenantId: 'acme-eu' },
+  },
+});
+
+// The call site is otherwise untouched; the context rides in the vendor's own parameter object
+// and is lifted back out before the vendor sees it.
+await govClient.messages.create({
+  model,
+  max_tokens: 1024,
+  messages,
+  stiggContext: { customerId: 'customer-acme' },
+});
+
+await govClient.messages.create({
+  model,
+  max_tokens: 1024,
+  messages,
+  stiggContext: { customerId: 'customer-globex' },
+});
+
+// Drains what is still in flight and reports what this client did.
+await govClient.stigg.flush();
+```
+
+**Nothing in the setup path touches the network.** The constructor is synchronous and so is
+`wrap` — resolving the engine is a probe that decides which transport serves which operation, and
+both start that work rather than waiting on it. The first call that needs an engine awaits the
+same probe; every call after it finds one already resolved. So there is no awaited setup step to
+forget, and no call that silently runs ungoverned because setup had not finished.
+
+The latency moves rather than disappearing: whoever makes the first call pays for the probe. If
+you would rather pay at boot — and have a bad key fail there instead of inside a request —
+`await stigg.init()` forces it early and hands back the resolution.
+
+Construct one per environment and share it. A second one for the same environment is wasted
+rather than wrong — and only partly a second client: `Stigg.initialize` keeps one global Node SDK
+instance per process and hands it back when the API key matches, so the entitlement cache and its
+socket are shared while the governance REST client is not. See
+[what the cache promises](#what-the-cache-promises) for what that means for `close()`.
+
+**A customer given at `wrap` is a default, not a scope.** One wrapped client can serve every
+tenant, so naming a customer there means every call that does not override it bills *that* one —
+correctly, silently, and forever if the client outlives the request. It is the right shape for a
+single-tenant process and the wrong one for a server, which is why the example above leaves it
+out and the types then require it per call.
+
+```ts
+// Single-tenant process: a default is fine, and calls need no context at all.
+const client = stigg.wrap(new Anthropic(), {
+  stiggContext: { customerId: 'customer-acme', entitlement: { currencyId: 'credits' } },
+});
+await client.messages.create({ model, max_tokens: 1024, messages });
+```
+
+`wrap()` also starts warming the entitlement cache when the context names both a customer and an
+entitlement, so the calls that follow read local state. That is an optimization and it is not
+awaited: a failure there is logged, never thrown, and the first check falls back to asking the
+server.
+
+**`wrap` returns a copy.** The object you built is never mutated and keeps working ungoverned,
+which is what makes the governed one safe to hand around. Wrapping an already-governed client
+throws rather than stacking a second gate on it — there is nothing a second wrap could mean now
+that the context travels with each call.
+
+`stigg.check(context)` and `stigg.report(context, metrics)` are the same two operations with no
+client to wrap, which is how a framework binding and the gateway reach them.
+
+`wrap()` recognises the vendor from the client's own shape — there is nothing there the SDK
+cannot see for itself — and throws, naming the vendors it knows, rather than handing back a
+client that quietly meters nothing. Pass `recipe` to name an adapter for a vendor it cannot
+identify.
+
+**Gemini chats are wrapped in their own right.** `new GoogleGenAI()` builds its chat module over
+the `models` module it holds — `new Chats(this.models, …)` — and every `Chat` keeps that
+reference privately, so a chat created from a wrapped client still calls the original, unpatched
+module. Wrapping the client governs `models.generateContent` and `models.generateContentStream`;
+wrap the chat to govern a conversation:
+
+```ts
+const ai = new GoogleGenAI({ apiKey });
+const models = stigg.wrap(ai, { stiggContext });     // one-shot calls
+const chat = stigg.wrap(ai.chats.create({ model }), { stiggContext }); // a conversation
+```
+
+Nothing is metered twice by doing both, because the chat never reaches the patched module.
+
+**Automatic function calling is metered in full only when you stream.** Hand Gemini a callable
+tool and its SDK drives the tool loop for you — up to ten model calls for one
+`generateContent` — and returns only the last response. Streamed, every round's chunks reach
+this SDK and are summed; measured on the wire, 122 + 170 = 292 of 292. Not streamed, the earlier
+rounds are discarded inside the vendor's SDK before it returns, so there is nothing left to read:
+a two-round call billed at 315 reports 181. That turn is still metered for what it can see, is
+tagged `gemini_afc_rounds`, and **raises** through `onError` with the shortfall named — it is
+counted in `summary().ingestFailures` rather than left for you to notice. To meter every round,
+stream the call, or set `automaticFunctionCalling: { disable: true }` and drive the loop
+yourself, which makes each round a call this SDK gates and meters on its own.
+
+Decisions describe themselves, so printing one takes no import: `decision.summary` is the log
+line, `decision.message` is what to tell the user. `stigg.denial(err)` turns a caught error back
+into a decision, or `undefined` when it was a real failure.
+
+### What a call checks, and what it reports
+
+They are different things, and keeping them apart is what lets Stigg do the pricing instead of
+your application.
+
+`entitlement` is what a call **checks** — a credit currency or a metered feature. Reporting goes
+somewhere else entirely: to `usageMapper`, one report per metric, named and valued as the vendor
+stated them. Stigg converts a credit-backed feature into the currency on its own side, which is
+what the API was built for and why `usage.report` answers with a credit balance.
+
+```
+[governance] ingest input_tokens=12 cache_read_input_tokens=3 output_tokens=46 ... credits=0.94
+```
+
+`entitlement` is also **the enforcement switch**. Name one and every call through the wrapped
+client is checked before it leaves. Leave it out and nothing is checked at all: no budget is
+pre-fetched, no call can be refused, and the calls meter exactly as gated ones do. There is no
+separate flag, because a context with nothing to check against cannot be enforced and a flag would
+only be a second place for the two to disagree — so the client says which it is, once, at `wrap`:
+
+```
+[stigg] this client named no entitlement, so nothing is checked and no call can be refused - it
+meters only. Pass `entitlement: { currencyId: 'credits' }` in `stiggContext`, at the wrap or on a
+call, to gate it.
+```
+
+The default is the vendor's own vocabulary, so **each recipe is where you read what that vendor
+meters** — `anthropicRecipe.metrics` is the list, typed as literals rather than as `string`.
+
+#### The events this SDK reports, and the meters to put behind them
+
+Reporting defaults to Stigg's event API (see the ingest table further down), and an event is
+matched by **name**. A wrapped client names each call after the call — the name someone reading
+that vendor's API would have guessed — so these are what your meters filter on:
+
+| Binding                   | Event name                                                          | Fields a meter can aggregate                                                              |
+| ------------------------- | ------------------------------------------------------------------- | ----------------------------------------------------------------------------------------- |
+| Anthropic                 | `anthropic.messages.create`                                         | `input_tokens`, `output_tokens`, `cache_read_input_tokens`, `cache_creation_input_tokens`, `total_tokens` |
+| OpenAI (chat)             | `openai.chat.completions.create`                                    | `prompt_tokens`, `completion_tokens`, `cached_tokens`, `total_tokens`                     |
+| OpenAI (responses)        | `openai.responses.create`                                           | `input_tokens`, `output_tokens`, `cached_tokens`, `total_tokens`                          |
+| OpenRouter                | `openrouter.chat.completions.create`, `openrouter.responses.create` | the OpenAI set, plus `cost`                                                               |
+| LiteLLM                   | `litellm.chat.completions.create`, `litellm.responses.create`       | the OpenAI set, plus `cost`                                                               |
+| TypeSafe                  | `typesafe.systemOne`                                                | `input_tokens`, `output_tokens`, `total_tokens`                                           |
+| Gemini                    | `gemini.models.generateContent`, `gemini.models.generateContentStream` | `promptTokenCount`, `candidatesTokenCount`, `cachedContentTokenCount`, `thoughtsTokenCount`, `toolUsePromptTokenCount`, `totalTokenCount` |
+| Gemini — a chat           | `gemini.chats.sendMessage`, `gemini.chats.sendMessageStream`        | the same five                                                                             |
+| A gateway, out of process | `<vendor>.proxied`                                                  | whatever the proxy declared its vendor to be                                              |
+
+Every metric rides the event under its own name, so **one event feeds as many features as you
+configure meters for** — that is why a turn costs one request however many counters it reported.
+The vendor's own usage block rides along too, under a `raw.` prefix and flattened to scalars, so a
+meter can read a counter this SDK does not name yet: `raw.cache_creation_input_tokens`, or a
+gateway's `raw.cost` before it was scaled into whole units. Your run's `dimensions` — `agentId`
+and anything else you set — are on the same event, and win any key the vendor also used.
+
+A meter to put behind one of them, in Stigg's own terms:
+
+```jsonc
+{
+  "aggregation": { "function": "SUM", "field": "output_tokens" },
+  "filters": [
+    { "conditions": [{ "field": "eventName", "operation": "EQUALS", "value": "anthropic.messages.create" }] },
+  ],
+}
+```
+
+> **An event reported against a name with no meter behind it returns 200 and meters nothing.** The
+> event API answers with an acknowledgement, not a balance, so nothing downstream can notice on its
+> own — configure the meter first, or run with `ingest: 'sync'`, which does answer.
+
+#### Pricing them yourself
+
+`weightedMetrics` turns a price-per-metric table into the mapper:
+
+```ts
+import { anthropicRecipe, weightedMetrics } from '@stigg/agent-sdk-js';
+
+const govClient = stigg.wrap(new Anthropic(), {
+  stiggContext: {
+    customerId,
+    entitlement: { currencyId: 'data_credit' },
+    // A whole credit per input token, half a credit per output token, cache reads free.
+    usageMapper: weightedMetrics(anthropicRecipe, 'data_credit', {
+      input_tokens: 1,
+      output_tokens: 0.5,
+    }),
+  },
+});
+```
+
+The table is keyed to that recipe's own names, so a typo does not compile. Naming a metric is
+optional — an unpriced one is free, and a table that prices nothing meters nothing.
+
+A weight is the price of **one unit** of that metric, so the example above is half a credit per
+output token. Pass `{ per: 1000 }` when the numbers read better per thousand.
+
+It also handles the three things a hand-written mapper has to get right: Stigg takes whole units
+so the total rounds up, a turn that measured nothing reports nothing rather than a zero, and a
+counter the response omitted is absent rather than zero. For one number without weights,
+`creditsPerTokens(featureId, n)` is the short form.
+
+#### Writing the mapper yourself
+
+A calculation a weights table cannot express is an ordinary function, and each binding exports the
+type of what it receives — `AnthropicMetrics` for Anthropic, and `OpenAIMetrics`,
+`OpenRouterMetrics`, `LiteLLMMetrics`, `GeminiMetrics` and `TypeSafeMetrics` for the rest:
+
+```ts
+import type { AnthropicMetrics, Metrics } from '@stigg/agent-sdk-js';
+
+const dataCredits = ({ metrics = {} }: { metrics?: AnthropicMetrics }): Metrics => {
+  const billable = (metrics.output_tokens ?? 0) * 5 + (metrics.input_tokens ?? 0);
+  return billable > 0 ? { data_credit: Math.ceil(billable / 100) } : {};
+};
+```
+
+Every counter is optional, because a vendor reports cache reads only when there were some. Use
+`AnthropicMetric` for one name rather than the table, and `MetricsOf<typeof someRecipe>` /
+`MetricsFor<typeof someRecipe>` for a recipe you built yourself.
+
+#### The overlap you have to know about
+
+Vendors disagree about whether cached tokens are already counted inside the input total, and the
+metrics are reported **verbatim** — because a number published under a vendor's own field name
+has to match that vendor's own dashboard.
+
+Every vendor also reports a **total**, which by definition overlaps everything beneath it. Where
+the vendor publishes one it is reported verbatim — `total_tokens`, or `totalTokenCount` for
+Gemini. Anthropic and TypeSafe publish none, so theirs is summed from counters that are disjoint,
+which is the only case where the SDK adds anything up. **Price the total or the breakdown, never
+both.**
+
+| Vendor           | Input-ish key   | Includes cached tokens?                                  |
+| ---------------- | --------------- | -------------------------------------------------------- |
+| Anthropic        | `input_tokens`  | **No** — cache keys are separate; the four sum correctly |
+| OpenAI Chat      | `prompt_tokens` | **Yes**, includes `cached_tokens`                        |
+| OpenAI Responses | `input_tokens`  | **Yes**, includes `cached_tokens`                        |
+| LangChain        | `input_tokens`  | **Yes**, includes both cache keys                        |
+| Gemini           | `promptTokenCount` | **Yes**, includes `cachedContentTokenCount`           |
+
+Gemini overlaps in both directions at once, so it is worth stating separately: pricing
+`promptTokenCount` **and** `cachedContentTokenCount` double-bills the cached tokens, while
+`thoughtsTokenCount` and `toolUsePromptTokenCount` sit *outside* the two headline counters —
+leave them out of a table and a reasoning model's thinking tokens, usually the expensive half of
+the answer, are billed at nothing. The four add back up to the vendor's own `totalTokenCount`,
+which is the arithmetic to check a table against.
+
+So if you price `prompt_tokens` **and** `cached_tokens` for OpenAI traffic, you bill the cached
+ones twice. Price one or the other, or use `normalizedMetrics` — four disjoint counters under
+SDK-owned names, comparable across every vendor:
+
+```ts
+import { normalizedMetrics } from '@stigg/agent-sdk-js';
+stigg.wrap(client, { stiggContext: { customerId, usageMapper: normalizedMetrics } });
+```
+
+OpenAI also has no raw name for cache creation, so its default set is three keys and a
+`cache_creation_input_tokens` feature stays at zero for OpenAI traffic.
+
+#### The failure that looks like success
+
+A feature only drains a credit budget when your catalog maps it to that currency. Until it does,
+every call succeeds, every ingest returns `ok`, and the budget never moves — so the call gates on a
+number nothing spends. Stigg answers those reports with no credit balance, so the first one says
+so:
+
+```
+[stigg] reported input_tokens but the response carried no credit balance - this feature is not
+mapped to credits, so the budget this run checks against will never drain. Map it in Stigg, or
+report to the currency directly.
+```
+
+### Bindings, by solution
+
+Organised by where the customer's business logic sits, not by vendor. Each has a generic core;
+vendor directories are thin recipes over it.
+
+The word "gateway" below does two jobs, so the directories do not: one you **call** is an ordinary
+client and lives in `bindings/client/gateways/`, beside `bindings/client/providers/` because both
+are recipes over the same wrapper. One you **operate** is the opposite binding and lives in
+`bindings/proxy/`.
+
+| Binding                                          | Reached as                                           | Mechanism                                                                                | Can refuse?                                |
+| ------------------------------------------------ | ---------------------------------------------------- | ---------------------------------------------------------------------------------------- | ------------------------------------------ |
+| Client                                           | `stigg.wrap(client, { stiggContext })`               | gates the recipe's methods on a copy, vendor matched by shape                            | Yes — when the context names an entitlement |
+| Client — a gateway you **call**                  | `wrap(client, { recipe: openrouterRecipe() })`       | same wrapper; the recipe bills `usage.cost` and attributes `response.model`              | Yes, same as any client                    |
+| Client — a gateway whose cost is in a **header** | `wrap(client, { recipe: litellmRecipe() })`          | same wrapper; the recipe bills `x-litellm-response-cost` and attributes `response.model` | Yes, same as any client                    |
+| Gateway you **operate** (`bindings/proxy`)       | `client.stigg.headers()` → `stigg.ingestFromHeaders(...)` | attribution travels as `x-stigg-*` headers; `vendor` names the shape it observed         | **No — ingest only**                       |
+
+A client run that names no `entitlement` lands in the same column as the gateway — metering
+without a gate — except that it is in-process, and naming one turns it back on.
+
+#### Providers that speak the OpenAI wire format
+
+A large part of the ecosystem exposes an OpenAI-compatible API rather than one of its own:
+OpenRouter, LiteLLM, Groq, Together, Fireworks, DeepSeek, xAI and Perplexity, plus self-hosted
+servers like vLLM and Ollama. There is no separate client to install for any of them. You point
+the OpenAI client at their base URL, and `wrap` matches on the **shape of the client**, not on who
+is at the other end of it:
+
+```ts
+// Groq, Together, Ollama, vLLM - anything OpenAI-compatible
+const client = stigg.wrap(
+  new OpenAI({ baseURL: 'https://api.groq.com/openai/v1', apiKey: process.env.GROQ_API_KEY }),
+  { stiggContext: { customerId, entitlement: { currencyId: 'credits' } } },
+);
+```
+
+So `OpenAI` in that line names the **wire format**, not the vendor being billed. What differs is
+which recipe matches, and that decides what you can meter:
+
+| Pointed at                           | Recipe matched     | Events it emits                   | Reports                           |
+| ------------------------------------ | ------------------ | --------------------------------- | --------------------------------- |
+| `https://openrouter.ai/api/v1`       | `openrouterRecipe` | `openrouter.*` (see below)        | tokens, `cost`, model that served |
+| the LiteLLM proxy URL you configured | `litellmRecipe`    | `litellm.*`                       | tokens, `cost`, model that served |
+| anything else OpenAI-compatible      | `openAIRecipe`     | `openai.*`                        | tokens only                       |
+
+Each of those covers **both OpenAI surfaces**, and a surface is its own event: Chat Completions
+reports `<vendor>.chat.completions.create` and the Responses API reports
+`<vendor>.responses.create`. A meter filtering on one sees nothing from the other, so configure the
+one your code actually calls - or both. The full list is in the event table earlier in this file.
+
+OpenRouter is recognised by its **hostname** and LiteLLM by the base URL you hand its recipe, which is
+why those two get a `cost` the token counters cannot recover. Everything else is OpenAI as far as
+this SDK can tell, because that is all the client reveals.
+
+> **One exception to that last row.** Given no base URL of its own, the LiteLLM recipe falls back to
+> looking for `litellm` anywhere in the client's URL - a self-hosted proxy has no canonical hostname
+> to compare against, the way OpenRouter's own hostname does. So an unrelated OpenAI-compatible
+> endpoint that happens to contain the string, like `https://api.acme.com/litellm-compat/v1`, is
+> claimed by it and reports under `litellm.*`. Pass `recipe: openAIRecipe` to settle it explicitly.
+
+**The consequence worth planning for:** every provider in that last row reports under `openai.*`.
+A raw-events meter cannot tell Groq from Ollama from OpenAI itself, and neither can a per-provider
+budget. If you need them separated, give the call a dimension of your own —
+
+```ts
+stiggContext: { customerId, dimensions: { provider: 'groq' } }
+```
+
+— and filter the meter on it, or pass a recipe with its own `vendor` name, which changes the event
+name to match.
+
+> **Not covered yet:** OpenRouter's own `@openrouter/sdk` package is a different shape — its method
+> is `chat.send` and its counters are camelCase — so `wrap` does not recognise it and will refuse
+> the client rather than meter it silently. Use the OpenAI-compatible endpoint above.
+
+Client recipes are data — method paths and usage extractors — so the factory wires the built-in
+ones with no vendor package loaded. Nothing in `src/index.ts` pulls a vendor package in at
+runtime, which is what keeps one import enough for every example.
+
+"Gateway" covers two deployments that want opposite bindings, and the split is worth naming.
+
+**A gateway you call** — OpenRouter as a service — returns its response into your process, and
+that response carries two things no token counter can recover: the credits it actually charged
+(`usage.cost`) and which model actually served after fallback routing (`response.model`). That is
+an ordinary client, so it takes the ordinary client binding: `openrouterRecipe()` reads both, and
+the call is gated before it ever leaves. It **is** registered by default and matched by base URL
+ahead of `openAIRecipe`, because reading a richer response is not a decision anyone needs to opt
+into — out of the box it reports the same token metrics the plain OpenAI recipe would, and adds
+`cost`, the model that served and the gateway's own request id.
+
+Tokens and money are different features, so both are reported and you price each in Stigg. When a
+response carries no cost, the key is simply absent rather than substituted with something else.
+`unitsPerCredit` (micro-credits by default) is the denomination of the `cost` feature.
+
+**A gateway you call whose price is in a header** — LiteLLM Proxy — is the same binding with one
+wrinkle worth knowing before you budget on it. LiteLLM prices every call itself, but writes the
+number into `x-litellm-response-cost` rather than into the body, so `litellmRecipe()` reads the
+response and not just the payload. Headers go on the wire before the model has produced a token,
+so a **streamed** call cannot carry a cost at all — tokens survive a stream fine, a price does
+not. So a non-streaming call reports its tokens **and** a `cost`, while a streamed one reports
+the tokens alone — the same features either way, one of them simply missing where the protocol
+cannot carry it.
+That is the protocol rather than a limit of the binding, and the gateway example prints both
+numbers side by side so the difference is not a surprise in an invoice.
+
+**A gateway you operate** — a proxy fronting services that may not even be Node — is the one that
+is not a binding in the full sense, and deliberately so: it is out-of-process and sees
+**requests, not runs**. It cannot provision, cannot check against run-scoped state, and cannot
+archive, and there is no client object out there to wrap. So the lifecycle splits across the
+boundary — the application checks in-process, attribution travels as headers,
+and the proxy only ingests. Governance never leaves the process.
+
+A recipe reports a `Measurement`, not just token counts: `metrics` carries the vendor's own field
+names and is what gets reported (a gateway puts its real spend in `metrics.cost` there, where a
+token mapper would only lose it), `raw` rides an event's dimensions under a `raw.` prefix,
+`dimensions` are merged over the context's own, and `idempotencyKey` lets the vendor's own request
+id — stable across _its_ retries, which a local step counter cannot be — replace the default
+`correlationId:step`. Returning
+a bare `TokenUsage` still works and stays the common case.
+
+### The engine: SDK first, REST to complete
+
+Two clients, composed per operation, and the SDK tells you which one it picked:
+
+```text
+[governance] mode=governance check=rest report=node-sdk provisioning=rest (governance enabled)
+```
+
+| Operation           | Client                                | Why                                                                      |
+| ------------------- | ------------------------------------- | ------------------------------------------------------------------------ |
+| `check`             | `@stigg/typescript` (REST, `v1-beta`) | only the REST surface takes `dimensions` and returns governance `chains` |
+| `ingest`            | `@stigg/node-server-sdk`              | it has the in-memory cache, and usage reporting is the hot path          |
+| provision / archive | `@stigg/typescript` (REST, `v1-beta`) | entities and assignments exist only on the governance surface            |
+
+**Ingest has two APIs, and the context picks one.** They are not variants of one call:
+
+|                        | `ingest: 'async'` — the default     | `ingest: 'sync'`                                     |
+| ---------------------- | ----------------------------------- | ---------------------------------------------------- |
+| Call                   | `events.report` / `reportEvent`     | `usage.report` / `reportUsageBulk`                   |
+| Latency and rate limit | low latency, much higher limit      | higher latency, lower limit                          |
+| The next check         | lags while Stigg aggregates         | already includes the spend                           |
+| What you send          | the call, as one event              | a `featureId` and a whole-number `value`, per metric |
+| Meter type             | **Raw events**, configured in Stigg | calculated / incremental                             |
+| What comes back        | an acknowledgement                  | a credit balance                                     |
+
+Async sends **one event for the whole turn**, every metric on its own dimension, because one event
+feeds any number of meters — fanning out per metric buys nothing and costs a request each. Sync
+sends one report per metric, batched into a single bulk call.
+
+A metric named after the currency the context checks is neither: it draws the wallet directly, since
+no raw event can reach a currency. It follows the mode only in whether it waits for the balance.
+
+**Governance is optional.** Not every account has it. When it is missing the factory falls back
+to customer-level enforcement on the Node SDK alone — every example still runs, denials just name
+the customer instead of an agent:
+
+```text
+[governance] mode=customer-level check=node-sdk report=node-sdk provisioning=unavailable
+             (403 GovernanceNotEnabled - customer-level fallback)
+```
+
+Nothing declares this. The client asks the account once at init and acts on the answer, so there
+is no second source of truth to disagree with the first. Without governance, checks and usage
+reporting still work at the customer level — and `register()` follows that split. A bare
+registration warns and hands the id back: it asked for an identity, and a call needs none. A call
+carrying a **budget or a parent** throws, because there is nowhere to write the assignment, and a
+run that looked governed with a budget that was never written is the failure worth refusing.
+
+### What the cache promises
+
+**This package has no cache.** Checks are served by the Node server SDK's entitlement cache, which
+is a far better one than a TTL over decisions could be:
+
+- it writes the server's own measurement response back into cached usage, so a report that moved
+  the budget is visible to the very next check;
+- it invalidates on realtime updates, so a plan change or a cancelled subscription lands without
+  waiting out a window;
+- it applies usage-period resets at read time, rather than serving a number that expired;
+- it falls back to configured entitlements when Stigg is unreachable;
+- and it can be backed by **Redis**, so sibling processes share one view instead of each holding
+  their own.
+
+All of that is configured through `cache`, which is the Node SDK's own configuration passed
+straight through — `redis`, `memoryCacheMaxSizeBytes`, `entitlementsFallback`,
+`creditEntitlementsFallback`, `realtimeUpdatesEnabled`, `enableEdge`, `offline`:
+
+```ts
+const stigg = new StiggAgentSDK(apiKey, {
+  cache: { redis: { host: 'redis', port: 6379, environmentPrefix: 'prod', ttl: -1 } },
+});
+```
+
+`wrap()` primes it for the customer in the context, loading their whole entitlement state rather
+than one decision — so a later check against a different entitlement is warm too, and each check
+is evaluated against the amount it actually asks for.
+
+#### If your application already has a Stigg client
+
+Pass it, and the SDK uses it instead of building one:
+
+```ts
+const stigg = new StiggAgentSDK(apiKey, { client: appStiggClient });
+```
+
+This is about **ownership**, not deduplication. `Stigg.initialize` already keeps one global
+instance per process and hands it back when the API key matches — so without `client` you may
+well be handed the application's own, and nothing observable can tell that apart from a client
+the SDK just built: the instance exposes neither its configuration nor its identity. Since
+`close()` nulls that global and tears the runtime down, `stigg.close()` would then take the
+application's cache and socket with it.
+
+A client passed here is **borrowed and never closed**; one the SDK builds is owned and closed. It
+replaces only the cached half — the governance REST client is still built from the API key,
+because the Node SDK has no governance surface — and it is mutually exclusive with `cache`, which
+configures a client the SDK is no longer building.
+
+An earlier design cached decisions here for a `cacheTtlMs` (default 5s). It duplicated the
+machinery above and degraded it: a cached allow outlived the report that spent the budget, could
+not hear an invalidation, and did not carry the requested amount in its key — so an allow fetched
+for one unit authorized a call asking for ten thousand. `cacheTtlMs` is now **ignored**, and
+warns when set.
+
+**The one path that caches nothing is an attributed governance check** — a check carrying
+dimensions, which is how per-agent budgets are enforced. It resolves an entity, which only the
+governance API can do, so it is a live call every time. That is why
+`RestGovernanceEngine.capabilities.strictCheck` is `true`: there is no window in which two
+processes can each approve the same budget. It costs a round trip per call, which is the right
+trade in front of a multi-second model call and the wrong one to optimize away with a stale allow.
+
+None of it applies to a context that names no `entitlement`: it warms nothing and checks nothing,
+because it never asks.
+
+#### How this SDK identifies itself
+
+**There is no telemetry.** Nothing is emitted on a schedule, nothing is batched in the background,
+and there is no second network path that could fail or leak. The SDK names itself on the requests
+it was already making, and that is all:
+
+| Half | What it sends |
+| --- | --- |
+| Cached reads and ingest (GraphQL) | `apollographql-client-name: agent-sdk-js`, `apollographql-client-version: <this package's version>` |
+| Governance (REST) | `X-SDK-CLIENT-NAME: agent-sdk-js`, `X-SDK-CLIENT-VERSION: <this package's version>` |
+
+No customer data, no prompts, no model output — the package name and its version, next to the
+process identity (`X-SDK-INSTANCE-ID` and friends) the Node SDK already sends.
+
+Two things it deliberately does **not** do. It does not send a distinct `source`: that header is a
+server-side enum with no member for this SDK, the API groups on it, so a new member is a server
+change that has to land first — the client name carries the distinction until then. And it does not
+rename a client you passed in through `client`, which keeps whatever name your application gave it;
+`Stigg.initialize` may also hand back a client your application built first, in which case the same
+applies. Both cases are correct — the name belongs to whoever owns the client.
+
+## 0. Try it with no setup at all
+
+```bash
+yarn workspace @stigg/agent-sdk-js run example-client-anthropic    # or -openai, -gemini, -openrouter, -litellm
+yarn workspace @stigg/agent-sdk-js run example-gateway             # or example-gateway-litellm
+```
+
+The full list, with what each one demonstrates, is in [section 6](#6-run-the-examples).
+
+The vendor mock speaks the real wire formats, so the vendor SDK parses a genuine response and
+every field the examples read (`usage`, `model`, `cost`, tool calls) is populated. The Stigg mock
+implements real entities, budgets and hierarchy, so a demo genuinely hits a denial and can name
+which entity caused it. Only the prose and the storage are fake.
+
+To force the mocks even when keys **are** present — useful when a key is valid but the account
+has no credits — set `AGENT_SDK_OFFLINE=1` in `.env`.
+
+## 1. Prerequisites
+
+- **Node 22** — the version in the repo's `.nvmrc`. Check with `node --version`.
+- **Yarn 4** via Corepack:
+
+  ```bash
+  corepack enable
+  ```
+
+## 2. Get an Anthropic API key (optional)
+
+Used by the client examples. Skip this to stay on the offline mock.
+
+1. Go to **<https://console.anthropic.com>** and sign up or sign in.
+2. Open **Billing** and add a payment method or buy credits. A brand-new org has no balance,
+   and requests fail with `400 credit balance is too low` until you do.
+3. Open **Settings → API Keys** and click **Create Key**. Give it a name.
+4. **Copy the key immediately** — it is shown only once. It starts with `sk-ant-`.
+5. Put it in `.env` as `AGENT_SDK_ANTHROPIC_API_KEY` (step 5 below).
+
+The examples default to the `claude-opus-5` model. Override with `AGENT_SDK_ANTHROPIC_MODEL` in `.env`.
+
+## 3. Gateway access (optional)
+
+Used by the gateway examples. Skip this entirely to stay on the offline mock, which speaks both
+wire formats — including LiteLLM's header-borne cost, and its absence on a stream.
+
+### OpenRouter
+
+1. Go to **<https://openrouter.ai>** and sign in with Google, GitHub, or email.
+2. Open **Settings → Credits** and add credits (a few dollars is plenty for these examples).
+3. Go to **<https://openrouter.ai/keys>** and click **Create Key**.
+4. Optionally set a **credit limit** on the key — a hard cap on what this key can ever spend.
+   Recommended while experimenting.
+5. Copy the key (starts with `sk-or-`) and put it in `.env` as `AGENT_SDK_OPENROUTER_API_KEY`.
+
+The example asks for `AGENT_SDK_OPENROUTER_MODEL` and falls back through
+`AGENT_SDK_OPENROUTER_FALLBACK_MODELS`.
+Both are model slugs — browse or verify them at **<https://openrouter.ai/models>**.
+
+### LiteLLM
+
+There is no signup: a LiteLLM proxy is something you host, so this is a URL rather than an account.
+
+1. Start one — `pip install 'litellm[proxy]'`, then `litellm --model gpt-4o --port 4000`. It needs
+   the upstream key for whatever it fronts (`OPENAI_API_KEY`, and so on) in its own environment.
+2. Put its URL in `.env` as `AGENT_SDK_LITELLM_BASE_URL`. The default is `http://localhost:4000` —
+   note a **local Stigg** uses port 4000 too, so move one of them if you run both.
+3. Put a virtual key in `AGENT_SDK_LITELLM_API_KEY` if the proxy requires one (`--master_key`, or
+   a key minted through `/key/generate`). Leave it blank and the examples use the offline mock.
+
+`AGENT_SDK_LITELLM_MODEL` is a **model group** from the proxy's own config, not a provider slug.
+A group is free to resolve to a different deployment, and that difference is the routing drift
+the examples flag.
+
+## 4. Point at a Stigg environment (optional)
+
+One key and one URL, pointed at whichever environment you mean. The key is environment-bound, so
+it and the URL always move together — there is no separate profile name to keep in sync with them.
+
+| Environment | `STIGG_BASE_URL`                                                |
+| ----------- | --------------------------------------------------------------- |
+| production  | `https://api.stigg.io` (the default when the variable is blank) |
+| staging     | `https://api-staging.stigg.io`                                  |
+| local       | `http://localhost:4000`                                         |
+
+Get a server key for that environment from **app.stigg.io → Integrations → API keys** and put it
+in `STIGG_API_KEY`.
+
+**Leave `STIGG_API_KEY` blank and everything still runs** against a built-in Stigg mock, with no
+account at all — the same rule the vendor keys follow. `AGENT_SDK_OFFLINE=1` forces the mocks even
+when keys are present, which is what you want when a key is valid but the account has no credits.
+
+Two things adjust themselves to the URL, so you do not configure them twice:
+
+- The mock answers both transports — REST and `/graphql` — so the engine resolves exactly as it
+  does against a real account: governance checks and provisioning over REST, ingest and
+  unattributed checks over the Node SDK. The line the SDK prints at startup says which it picked.
+- Against `localhost` the Node SDK's edge cache and realtime socket are switched off, because a
+  local API serves neither and the SDK maps a localhost base to edge port 3060.
+
+### Seed the environment
+
+```bash
+yarn workspace @stigg/agent-sdk-js run example-setup                    # 5-credit agent budget
+yarn workspace @stigg/agent-sdk-js run example-setup --agent-limit 6    # raise it
+yarn workspace @stigg/agent-sdk-js run example-setup --no-governance    # seed without entities
+yarn workspace @stigg/agent-sdk-js run example-setup --archive          # clean up, bottom-up
+```
+
+`--no-governance` describes this invocation only: it skips the entity and budget steps, so the
+account is seeded without a tree. It is not a mode the examples inherit — offline, each one starts
+its own mock, and that mock has governance enabled. To see customer-level enforcement end to end,
+use an account that genuinely has no governance.
+
+Idempotent per step, and safe to re-run — but not safe to run twice **at once**: the credit grant
+reads the balance and then grants, with nothing between the two, so two concurrent invocations can
+both decide the pool is short and both fund it. Run it once at a time. Steps are gated by what the
+account actually has:
+
+| Step                       | Needs governance? | What it does                                                                     |
+| -------------------------- | ----------------- | -------------------------------------------------------------------------------- |
+| 1. customer                | no                | creates it if missing                                                            |
+| 2. credit pool             | no                | grants credits when the balance is below `--pool-min`, guarded by a balance read |
+| 3. entity types + entities | **yes**           | `team` (attribution key `teamId`) and `agent` (`agentId`)                        |
+| 4. assignments             | **yes**           | One agent per example, each = 5 per `P1M`, all parented to `team-research`       |
+
+Each example registers **its own agent** — `agent-support-bot-anthropic`, `-openai`, `-gemini`, `-openrouter`,
+`-litellm`, and the base `agent-support-bot` for the gateway examples. Budgets attach to the agent, so
+sharing one would mean running the OpenAI example spends what the Anthropic example had left. The
+shared parent team keeps the ancestor rule demonstrable, and its budget is sized so one agent
+exhausting itself does not block the others:
+
+```text
+team-research (team) usage=0/30
+  agent-support-bot            usage=0/5
+  agent-support-bot-anthropic  usage=0/5
+  agent-support-bot-openai     usage=0/5
+  agent-support-bot-openrouter usage=0/5
+  agent-support-bot-litellm    usage=0/5
+```
+
+**The client examples provision their own agent** — the seed prepares the account (entity types,
+the parent team, a funded pool). That is the more realistic split: an admin sets up the account, an
+application provisions its own agents at runtime. `register()` carries it:
+
+One caveat about the cap, because the two halves can overlap. `usageLimitOnCreate` is exactly that
+— on create — so if the seed has already written an assignment for that agent, the application's
+cap is **not** applied and the seeded one stands. That is the point of the option (an admin who
+raises a budget out of band should not have it stamped back on every process start), but it does
+mean `--agent-limit` on the seed wins over the number below. Use `usageLimit` instead if the
+application is meant to own it.
+
+```ts
+await stigg.register({
+  customerId: 'customer-acme',
+  agentId: 'agent-support-bot-anthropic',
+  entityTypeId: 'agent',
+  // A budget is written over (entity x entitlement), so a call carrying one names it. `register`
+  // refuses a placement without it rather than provisioning an agent whose cap lands nowhere.
+  entitlement: { currencyId: 'credits' },
+  parentId: 'team-research',
+  usageLimit: 5,
+  cadence: 'P1M',
+});
+```
+
+A cap is **not** a field on the entity. An entity is just a node — id, type, display name,
+metadata — and a limit is an _assignment_ over `(entity × capability)`. So that one call is two
+idempotent upserts underneath: `PUT …/entities` then `PUT …/assignments`. It is also why
+`parentId` rides the assignment rather than the entity: the hierarchy is per-capability, so one
+agent can sit under different parents for different currencies.
+
+`usageLimit` is tri-state — a number caps it, `null` tracks without blocking (the observe-only
+rollout), and omitting it leaves whatever is already assigned untouched. Governance requires both
+it and `cadence` when it creates a budget, which is why both are passed above.
+
+The pool is deliberately much larger than the agent budget, so a demo denial comes from the
+agent's own cap (`BudgetExceeded`) rather than an empty wallet (`InsufficientCredits`). Those are
+different products — an upsell versus an escalation to whoever set the budget — and the SDK keeps
+them distinct all the way to the message the user sees.
+
+## 5. Configure and install
+
+Two files, deliberately separate. Both are git-ignored, and each has a committed `.template`
+beside it to copy from.
+
+**`.env`** (from `.env.template`) — credentials and SDK configuration. Everything in it is either a
+secret or a real setting you would also make in your own application.
+
+**`examples/demo.env`** (from `examples/demo.env.template`) — the demo's fixtures: the tenant, the
+team, the agent. Nothing in it is configuration — a real application knows its tenant from the
+request it is serving, never from an env file.
+
+Neither file is required **against the mocks**: every fixture default is compiled into the
+examples, so with no files and no keys at all they still run end to end. Those defaults are
+mock-only. Once `STIGG_API_KEY` is set the examples stop inventing ids for an environment they
+can actually write to — `AGENT_SDK_DEMO_CUSTOMER_ID` becomes required, and the rest degrade to a
+smaller demo (see the fixtures table below).
+
+`.env` is loaded first and is never overwritten, so setting one of the `AGENT_SDK_DEMO_*` names
+there (or exporting it in your shell) overrides the fixture without editing `demo.env`.
+
+### `.env` — credentials and endpoints
+
+| Variable                               | Default                                  | Purpose                                                                                                                              |
+| -------------------------------------- | ---------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ |
+| `AGENT_SDK_ANTHROPIC_API_KEY`          | — (falls back to `ANTHROPIC_API_KEY`)    | Layers 1 and 3                                                                                                                       |
+| `AGENT_SDK_ANTHROPIC_MODEL`            | `claude-opus-5`                          | Claude model for layers 1 and 3                                                                                                      |
+| `AGENT_SDK_OPENAI_API_KEY`             | — (falls back to `OPENAI_API_KEY`)       | The OpenAI client example                                                                                                            |
+| `AGENT_SDK_OPENAI_MODEL`               | `gpt-5`                                  | Model for the OpenAI client example                                                                                                  |
+| `AGENT_SDK_OPENROUTER_API_KEY`         | — (falls back to `OPENROUTER_API_KEY`)   | Layer 2                                                                                                                              |
+| `AGENT_SDK_GEMINI_API_KEY`             | —                                        | Falls back to `GEMINI_API_KEY` / `GOOGLE_API_KEY`. Absent, the Gemini example runs on the built-in mock                               |
+| `AGENT_SDK_GEMINI_MODEL`               | `gemini-3.8-flash`                       | Model for the Gemini client example                                                                                                  |
+| `AGENT_SDK_OPENROUTER_MODEL`           | `anthropic/claude-opus-5`                | Primary model requested from the gateway                                                                                             |
+| `AGENT_SDK_OPENROUTER_FALLBACK_MODELS` | `anthropic/claude-sonnet-5,openai/gpt-5` | Comma-separated fallbacks tried in order                                                                                             |
+| `AGENT_SDK_LITELLM_API_KEY`            | — (falls back to `LITELLM_API_KEY`)      | Layer 2. Blank ⇒ the offline mock                                                                                                    |
+| `AGENT_SDK_LITELLM_BASE_URL`           | `http://localhost:4000`                  | Your own proxy — self-hosted, so there is no public default                                                                          |
+| `AGENT_SDK_LITELLM_MODEL`              | `gpt-5`                                  | A model **group** from the proxy's config, not a provider slug                                                                       |
+| `STIGG_API_KEY`                        | —                                        | **Server** key (`server-…`). A publishable `client-…` key fails every call with `IdentityForbidden`. Blank ⇒ the built-in Stigg mock |
+| `STIGG_BASE_URL`                       | `https://api.stigg.io`                   | Which environment that key belongs to                                                                                                |
+
+### What the examples gate and meter against — in `examples/demo.env`
+
+These moved out of `.env`, because they fail the test that separates the two files: **if you would
+change it to play with an example, it is a fixture.** A real application decides every one of them
+in code — what it bills against is part of its pricing, not of its deployment.
+
+The SDK's own settings are arguments rather than variables, which is why none of them appear here:
+`new StiggAgentSDK(apiKey, baseUrl, { cache, client, timeoutMs, logger, onError })` for the
+environment, and `wrap(client, { stiggContext })` — overridable inside any call's own parameters —
+for what a call is governed by.
+
+| Variable                         | Default                 | Purpose                                                                            |
+| -------------------------------- | ----------------------- | ---------------------------------------------------------------------------------- |
+| `AGENT_SDK_DEMO_CURRENCY_ID`     | `credits` *(mock only)* | `stiggContext.entitlement` — credit currency to budget against. The default is the examples' own and applies only offline; the SDK has none. Naming neither this nor a feature is legal and means the calls **meter without gating** — `entitlement` is itself the enforcement switch. `two-agents` and the gateway examples refuse to run that way, since a budget draining is their whole subject |
+| `AGENT_SDK_DEMO_FEATURE_ID`      | —                       | `stiggContext.entitlement` — metered feature instead; exclusive with the above      |
+| `AGENT_SDK_DEMO_EVENT_NAME`      | —                       | `stiggContext.eventName` — overrides the name derived from the vendor call          |
+
+### `examples/demo.env` — demo fixtures
+
+Scaffolding, not configuration. Your own application supplies all of this from its own context.
+
+Each fixture has **three** states, because "no `demo.env` at all" and "I want no agent" are
+different things and an absent variable cannot mean both:
+
+| The variable is…      | Against the mock              | With a real `STIGG_API_KEY`                     |
+| --------------------- | ----------------------------- | ----------------------------------------------- |
+| **absent**            | the default below             | unset — the demo degrades, or says what it needs |
+| **set**               | that value                    | that value                                      |
+| **present but blank** | unset, deliberately           | unset, deliberately                             |
+
+The defaults are mock-only on purpose: against an environment you can write to, a guessed
+`customer-acme` is either someone else's tenant or one nobody asked for.
+
+| Variable                              | Default *(mock only)* | Unset ⇒                                                               |
+| ------------------------------------- | --------------------- | --------------------------------------------------------------------- |
+| `AGENT_SDK_DEMO_CUSTOMER_ID`          | `customer-acme`       | **Error.** Every governance call is scoped to a tenant, and the examples will not invent one |
+| `AGENT_SDK_DEMO_TEAM_ID`              | `team-research`       | The agents are **roots** — no ancestor budget binds them. `two-agents` needs one and says so |
+| `AGENT_SDK_DEMO_AGENT_ID`             | `agent-support-bot`   | **No agent at all**: checks and usage land on the customer's own pool. Blank it to see that path against the mock. Each client example appends its vendor (`…-anthropic`) |
+| `AGENT_SDK_DEMO_ENTITY_TYPE_ID` | `agent` *(mock only)* | `stiggContext.agentType.entityTypeId` — the entity type agents are provisioned as |
+| `AGENT_SDK_DEMO_ATTRIBUTION_KEY` | `agentId` *(mock only)* | `stiggContext.agentType.attributionKey` — the key that type identifies its agents under. **Both or neither**: the SDK defaults neither half, because a guessed key resolves no entity and the spend lands on the customer looking exactly like an ungoverned call |
+| `AGENT_SDK_OFFLINE`                   | empty                 | Set to `1` to force the mocks even when keys are present              |
+| `AGENT_SDK_DEMO_MOCK_GOVERNANCE_FAIL` | empty                 | Mock only: make the runtime check return `503`                        |
+
+Install dependencies from the **repo root**:
+
+```bash
+corepack enable
+yarn install
+```
+
+## 6. Run the examples
+
+All commands work from anywhere in the repo, and all of them run with no keys at all — except
+`example-client-typesafe`, whose vendor the built-in mock does not speak. That one still needs no
+Stigg account.
+
+| Command                     | Runs                                                                  | What it shows                                                                                               |
+| --------------------------- | --------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------- |
+| `example-client-anthropic`  | [`client/anthropic.ts`](./examples/client/anthropic.ts)               | A governed chat. One `wrap()` gates and meters every turn                       |
+| `example-client-openai`     | [`client/openai.ts`](./examples/client/openai.ts)                     | The same script, a different vendor — governance is not vendor-specific work                                |
+| `example-client-gemini`     | [`client/gemini.ts`](./examples/client/gemini.ts)                     | The same script again, on a vendor that counts in its own names and streams running totals                  |
+| `example-client-openrouter` | [`client/openrouter.ts`](./examples/client/openrouter.ts)             | `openrouterRecipe()` billing the credits actually charged instead of a token count                          |
+| `example-client-litellm`    | [`client/litellm.ts`](./examples/client/litellm.ts)                   | The same script against a self-hosted proxy. Streams, so it meters tokens — see the gateway example for why |
+| `example-client-typesafe`   | [`client/typesafe.ts`](./examples/client/typesafe.ts)                 | A non-chat client: typed questions, batched or one per call. **Needs a real TypeSafe key**                  |
+| `example-two-agents`        | [`client/two-agents.ts`](./examples/client/two-agents.ts)             | Two agents, one team: a sibling denied by its own cap, then the survivor denied by the team above it        |
+| `example-gateway-openrouter`| [`gateway/openrouter.ts`](./examples/gateway/openrouter.ts)           | The two gateway topologies: calling one gates, operating one only ingests. `example-gateway` runs this one  |
+| `example-gateway-litellm`   | [`gateway/litellm.ts`](./examples/gateway/litellm.ts)                 | The same two topologies when the cost is in a header: real spend, then tokens, from one request             |
+| `example-setup`             | [`shared/setup-governance.ts`](./examples/shared/setup-governance.ts) | Seeds a customer, a pool and per-entity budgets. Idempotent                                                 |
+
+`example-client` is an alias for the Anthropic one. Prefix any of them with `AGENT_SDK_OFFLINE=1`
+to force the mocks even when keys are present.
+
+### Layer 1 — client abstraction (interactive chat)
+
+An interactive, multi-turn chat, governed end to end, against any of five vendors. Type `exit`,
+`quit`, an empty line, or press **Ctrl-D** / **Ctrl-C** to end it.
+
+```bash
+yarn workspace @stigg/agent-sdk-js run example-client-anthropic     # also: example-client
+yarn workspace @stigg/agent-sdk-js run example-client-openai
+yarn workspace @stigg/agent-sdk-js run example-client-gemini
+yarn workspace @stigg/agent-sdk-js run example-client-openrouter
+yarn workspace @stigg/agent-sdk-js run example-client-litellm
+```
+
+All five run [the same script](./examples/client/core/chat.ts) — same governance wiring, same
+output, not one line of it vendor-specific. What differs is a helper per vendor
+([anthropic](./examples/client/anthropic.ts), [openai](./examples/client/openai.ts),
+[gemini](./examples/client/gemini.ts), [openrouter](./examples/client/openrouter.ts),
+[litellm](./examples/client/litellm.ts)) holding the four things that genuinely differ:
+which key to read, how to build the client, how to stream a turn, and where usage sits on the
+response. That is the same split the SDK makes internally, for the same reason.
+
+All five report the names their own vendor used, which is why the ingest lines differ between
+them: Anthropic states `input_tokens`/`output_tokens`, OpenAI states `prompt_tokens`/
+`completion_tokens`, and Gemini states `promptTokenCount`/`candidatesTokenCount` — plus a
+`thoughtsTokenCount` that is billed as output and counted beside the candidates rather than
+inside them. Run `example-gateway-litellm` to watch a gateway add `cost` beside them.
+
+```text
+[governance] stigg=in-process mock customer=customer-acme currency=credits
+[stigg] mode=governance check=rest (node-sdk when unattributed) report=node-sdk provisioning=rest (governance enabled)
+[stigg] agent agent-support-bot-anthropic registered under customer customer-acme
+
+you> what is metering?
+[governance] check hasAccess=true remaining=5 source=rest
+assistant> Metering is the practice of recording how much of a service ...
+  [usage:turn 1] input=12 output=46 cache_read=3 cache_creation=0
+[governance] ingest input_tokens=12 cache_read_input_tokens=3 output_tokens=46 key=9d63...:1 ok=true credits=0.94
+
+...
+
+you> one more
+[governance] check hasAccess=false remaining=0.5742 deniedBy=agent-support-bot-anthropic reason=BudgetExceeded remedy=raise-budget source=rest
+  Budget exhausted for agent-support-bot-anthropic. Ask whoever set this budget to raise it, or wait for the next reset.
+  Raise the budget in another shell, then just keep typing - the next message re-checks:
+    yarn workspace @stigg/agent-sdk-js run example-setup --agent-limit 6
+```
+
+The chat stays open after a denial. Run the suggested command in another shell and keep typing:
+because denials are never cached, the next message goes through. That is the whole point of
+making a refusal a commercial event rather than an error.
+
+Every check reads `source=rest`, and `remaining` falls as the ingests land. This is an attributed
+check — the agent travels as a dimension, so it resolves an entity and only the governance API can
+answer it. Nothing is served from a decision cache, which is why the denial arrives on the first
+call that actually exceeds the budget rather than one window later.
+
+### Layer 2 — gateways, in both directions
+
+The same prompt through OpenRouter three times: wrapped with `openrouterRecipe()` (non-streaming,
+then streaming), then once more down the header path, so the two topologies sit side by side.
+
+```bash
+yarn workspace @stigg/agent-sdk-js run example-gateway
+```
+
+```text
+--- wrapped, non-streaming ---
+requested model : anthropic/claude-opus-5
+served model    : anthropic/claude-sonnet-5  <- routing drift
+cost (credits)  : 0.00108
+prompt tokens   : 21 (cached: 7)
+completion tok. : 39 (reasoning: 4)
+[governance] hasAccess=true remaining=25000 source=rest
+[governance] ingest prompt_tokens=21 completion_tokens=39 cached_tokens=7 cost=1080 key=gen-mock... ok=true
+
+--- header path, non-streaming ---
+...
+[governance] proxy ingest prompt_tokens=21 completion_tokens=39 cached_tokens=7 key=gen-mock... ok=true
+(ingest-only: no check on this path, and tokens not cost - the mapper cannot see usage.cost)
+```
+
+The contrast is the whole point. Both report the same three token counters; only the wrapped call
+also reports **`cost=1080`** — 0.00108 credits in micro-credits, what the gateway really charged —
+attributed to the model that actually served. The header path carries a usage report rather than a
+price, so the one feature that says what the call cost in money is simply missing from it. Same
+request, same seams; the difference is a feature, not a denomination.
+
+If the primary model is down, rate-limited, or refuses, OpenRouter serves a fallback slug instead
+and `served model` differs — flagged as `<- routing drift`, which silently changes COGS. The
+recipe puts that slug on the ingest as a `model` dimension, so the drift is visible in Stigg and
+not just in the log.
+
+The `cost` feature is denominated by the recipe: `openrouterRecipe()` reports micro-credits by
+default, because OpenRouter charges fractional credits and Stigg ingests whole units. Set
+`unitsPerCredit` to match however the Stigg feature is defined. When a response carries no cost at
+all — BYOK, and a few upstreams — the `cost` key is simply absent; the token metrics were never
+conditional on it.
+
+#### When the price is in a header
+
+LiteLLM makes the same comparison and splits the wrapped call in two, because for this gateway
+_how_ you called it decides whether the price is knowable at all:
+
+```bash
+yarn workspace @stigg/agent-sdk-js run example-gateway-litellm
+```
+
+```text
+--- wrapped, non-streaming ---
+requested group : gpt-5
+served model    : gpt-5-2025-08-07  <- routing drift
+cost (USD)      : 0.001005
+[governance] ingest prompt_tokens=21 completion_tokens=36 cached_tokens=7 cost=1005 key=chatcmpl-... ok=true
+
+--- wrapped, streaming ---
+cost (USD)      : (not reported - the header precedes the body)
+[governance] ingest prompt_tokens=21 completion_tokens=36 cached_tokens=7 key=chatcmpl-... ok=true
+
+--- header path, non-streaming ---
+[governance] proxy ingest prompt_tokens=21 completion_tokens=36 cached_tokens=7 key=chatcmpl-... ok=true
+(ingest-only: no check on this path, and tokens not cost - what crosses the boundary is a usage
+report, so the price the proxy computed has nowhere to go)
+```
+
+Same proxy, same recipe, same prompt. The tokens are identical in all three; what differs is
+whether `cost` is there at all. The proxy knew the price every time — it computed it — but a
+header is flushed before generation starts, so the streamed call has nothing to read
+([BerriAI/litellm#12689](https://github.com/BerriAI/litellm/issues/12689)), and the ingest-only
+boundary carries a usage report rather than a price. Gating is unaffected throughout: both wrapped
+legs are refused before the request leaves, whatever they end up reporting.
+
+So if you budget in spend against LiteLLM, price the `cost` feature and know that streamed turns
+will not contribute to it — the token features still will. `litellmRecipe()` reports micro-dollars
+by default; set `unitsPerDollar` to match however the Stigg feature is defined.
+
+### Seeing the fail-closed path
+
+The governance check is fail-closed server-side: it returns `503` rather than silently allowing.
+
+```bash
+AGENT_SDK_OFFLINE=1 AGENT_SDK_DEMO_MOCK_GOVERNANCE_FAIL=1 \
+  yarn workspace @stigg/agent-sdk-js run example-client-anthropic
+# [stigg] check failed: check: 503 governance cache unavailable
+# [governance] check hasAccess=false reason=EngineError remedy=retry source=policy
+
+# The explicit opt-in - refuse the call when the check itself could not be made. An SDK option,
+# not an env var, and overridable per call through `stiggContext`:
+#   new StiggAgentSDK(apiKey, baseUrl, { checkErrorAction: 'deny' })
+```
+
+`allow` is the default: a soft limit, so an unreachable Stigg does not take the application down
+with it. `deny` mirrors the server's own fail-closed behaviour and is a decision worth making
+deliberately — "never break the customer's run" and "never spend past a budget" are not the same
+promise, and only you know which one this call is for.
+
+## 7. Typecheck, lint and test
+
+```bash
+yarn workspace @stigg/agent-sdk-js run build   # typecheck, then emit dist/ - what gets published
+yarn workspace @stigg/agent-sdk-js run lint
+yarn workspace @stigg/agent-sdk-js run test    # node:test via tsx; no network, literal fakes
+```
+
+`build` is two passes over two configs, and the split is the point. `tsconfig.json` is the
+typechecker's view: it covers `src`, `examples`, `test` and `testing`, and emits nothing.
+`tsconfig.build.json` compiles the two entry points - `src/index.ts` and `src/internal.ts` - and
+whatever they import, so only `src/` reaches `dist/`. It clears `dist/` first, because a module
+that was renamed or deleted would otherwise sit there and ship.
+
+Either pass resolves workspace dependencies to their built `.d.ts` (see `paths` in
+`tsconfig.json`) rather than to their source, which is compiled with different flags; Nx builds
+them first.
+
+## 8. Troubleshooting
+
+| Symptom                                     | Cause and fix                                                                                                                                                                                                                                                                                                                                        |
+| ------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `403 GovernanceNotEnabled`                  | The account has no governance. Checks and usage still work; `register()` throws, because there is nowhere to create the agent entity. Contact Stigg to enable it.                                                                                                                                                                                    |
+| `No STIGG_API_KEY`                          | Set it to a server key for the environment in `STIGG_BASE_URL`, or leave it blank to run on the built-in mock.                                                                                                                                                                                                                                       |
+| `503` and every turn denied                 | The governance cache is unavailable and `checkErrorAction` is `deny`. The default is `allow`, so this is a policy that was asked for; drop it to keep running through the outage, or relax a single call with `stiggContext: { checkErrorAction: 'allow' }`. |
+| `usage value must be a whole number`        | Stigg rejects fractional usage under `'sync'`. Raise the unit size in your `usageMapper` — `creditsPerTokens(id, n)` — or map to a coarser unit.                                                                                                                                                                                                                                                    |
+| `... is backed by a raw-events meter`       | A `sync` run reported to a feature only a meter can fill. Report that run with `ingest: 'async'` (the default), or point it at a feature whose value your application computes.                                                                                                                                                                      |
+| Ingest reports `ok=true` but usage stays 0  | Under `async` an acknowledgement is all the event API returns. Check a raw-events meter filters on the event name (see the event table above); against a **local** Stigg, also that the `ingest-usage-stream-processor` and `calculate-usage-stream-processor` lambdas are running. `ingest: 'sync'` answers with a balance and will say so instead. |
+| Still denied right after a top-up           | Budgets reset per cadence window, so a spent agent stays spent. Re-run `example-setup --agent-limit N` with a higher limit, or use a new `AGENT_SDK_DEMO_AGENT_ID`.                                                                                                                                                                                  |
+| Denial names the customer, not the agent    | Customer-level mode (no governance), or the context named no `agentId` and carried no attribution dimensions of its own.                                                                                                                                                                                                                             |
+| Denial names an agent you did not pass      | A `dimensions` key resolves an entity that already exists in the account. The agent belongs in `stiggContext.agentId`; `dimensions` gates on whatever it names, registered here or not.                                                                                                                                                              |
+| `401` / `authentication_error`              | The key is present but wrong. Clear it from `.env` to fall back to the offline mock, or fix it.                                                                                                                                                                                                                                                      |
+| `400 credit balance is too low` (Anthropic) | Add credits under Billing, or set `AGENT_SDK_OFFLINE=1` to run on the mock.                                                                                                                                                                                                                                                                          |
+| `402` / insufficient credits (OpenRouter)   | Add credits under Settings → Credits, or set `AGENT_SDK_OFFLINE=1` to run on the mock.                                                                                                                                                                                                                                                               |
+| `400` naming a model (OpenRouter)           | A slug in `AGENT_SDK_OPENROUTER_MODEL` / `AGENT_SDK_OPENROUTER_FALLBACK_MODELS` does not exist. Check <https://openrouter.ai/models>.                                                                                                                                                                                                                |
+| `ECONNREFUSED` on port 4000 (LiteLLM)       | No proxy is running there. Start one (`litellm --model gpt-4o --port 4000`), fix `AGENT_SDK_LITELLM_BASE_URL`, or clear `AGENT_SDK_LITELLM_API_KEY` to fall back to the mock.                                                                                                                                                                        |
+| `400` / unknown model group (LiteLLM)       | `AGENT_SDK_LITELLM_MODEL` must name a group from the proxy's own config, not a provider slug. List them with `curl $AGENT_SDK_LITELLM_BASE_URL/v1/models`.                                                                                                                                                                                           |
+| LiteLLM turns bill `units=1`, not spend     | Expected on a streamed call: `x-litellm-response-cost` is flushed before the body, so there is no cost to read. Call non-streaming to bill real spend.                                                                                                                                                                                               |
+| `Cannot find module`                        | Run `yarn install` from the repo root, not from this directory.                                                                                                                                                                                                                                                                                      |
+| Chat prints nothing and exits               | Input is not a TTY (piped input hits end-of-file). Try `printf 'hi\nexit\n' \| yarn workspace @stigg/agent-sdk-js run example-client-anthropic`.                                                                                                                                                                                                     |
+| A CLI hangs after the last line             | The Node SDK keeps a background cache alive. Call `gov.flush({ close: true })`, as the examples do.                                                                                                                                                                                                                                                  |
+
+## 9. Layout
+
+```
+src/
+  index.ts                                            the whole public surface, one import
+  stigg-agent-sdk.ts                                 the composition root: the one file that knows both sides
+  types.ts                                            the vocabulary, and every call's argument list
+  config.ts errors.ts logger.ts                       validated per-environment config, flat errors, structural logger
+  decision-text.ts headers.ts                         how a decision reads; the cross-process attribution convention
+  engine/     engine.ts decision.ts                   the four-operation seam; chains -> decision (pure)
+              node-sdk-engine.ts                      the only file importing @stigg/node-server-sdk
+              rest-governance-engine.ts               the only file importing @stigg/typescript
+              composite-engine.ts create-engine.ts    routing per operation, and who picked what
+  core/       run-handle.ts                           the handle every binding consumes
+  bindings/   client/core/                           the recipe contract, the registry, the wrapper
+              client/providers/ client/gateways/     a model provider, or a gateway you call
+              proxy/                                 a gateway you operate: headers in, ingest only
+              shared/                                usage shapes and metric tables both halves read
+
+docs/         instances.md                            what exists at runtime, and what is shared
+
+testing/      mock-stigg-server.ts                    the Stigg stand-in the examples and the tests share
+
+examples/
+  shared/     demo-env, the two mocks, setup-governance    scaffolding every example uses
+  client/     core/{chat,vendor} + anthropic|openai|gemini|openrouter|litellm|typesafe   a helper per vendor
+              two-agents                                   two siblings, one team budget between them
+  gateway/    core/{gateway,vendor,proxy} + openrouter|litellm  calling one; operating one on a real socket
+```
+
+The examples mirror the bindings directory deliberately — a generic core plus a thin per-vendor
+file is the shape the problem has, at both layers.
+
+Core never imports a binding or a concrete engine, and no binding imports another — only
+`stigg-agent-sdk.ts` knows both, which is what lets `wrap()` live on the SDK object
+without core learning a vendor's name. Adding a provider is a recipe file; adding an engine is a
+leaf file plus a branch in the factory.