@semiont/inference 0.5.29 → 0.5.31

package/README.md

@@ -90,22 +90,47 @@ interface InferenceClient {
   readonly type: string;     // 'anthropic' | 'ollama' | 'mock'
   readonly modelId: string;  // configured model name
 
+  // Declared per-provider capabilities — consumers read these instead of
+  // switching on provider identity:
+  readonly maxConcurrency: number;        // independent calls that gain from running at once
+                                          // (Anthropic 4; Ollama 1 — one GPU, no parallel gain)
+  readonly verifyDetectionYield: boolean; // whether detection count-verifies extractions
+                                          // (true for real providers; Mock false — tests opt in)
+
   limits(): Promise<InferenceLimits>;
-  generateText(prompt, maxTokens, temperature, options?): Promise<string>;
-  generateTextWithMetadata(prompt, maxTokens, temperature, options?): Promise<InferenceResponse>;
+  generateText(prompt, maxTokens, temperature, signal?): Promise<string>;
+  generateTextWithMetadata(prompt, maxTokens, temperature, signal?): Promise<InferenceResponse>;
+  generateStructured<T>(prompt, maxTokens, temperature, elementSchema, signal?): Promise<StructuredResponse<T>>;
 }
+```
+
+Whatever varies by provider is a **capability declared here**, hard-coded per implementation — downstream packages are written purely in terms of this contract and never sniff provider identity. A new provider must take a position on each capability (tests pin the declarations).
+
+Every generation method takes a trailing optional `AbortSignal`: aborting tears down the underlying transport (and, on Anthropic, the SDK's internal retry loop) so a cancelled call rejects promptly instead of surviving as a billed background request. Implementations must honor it — accepting and ignoring the signal is a defect.
+
+```typescript
 
 interface InferenceLimits {
-  contextTokens: number;     // context window (Anthropic: max input; Ollama: shared input+output)
-  maxOutputTokens: number;   // max output per generation (Ollama mirrors the shared window here)
+  contextTokens: number;         // context window (Anthropic: max input; Ollama: shared input+output)
+  maxOutputTokens: number;       // max output per generation (Ollama mirrors the shared window here)
+  outputTokensPerHour?: number;  // provider's worst-case output-rate model, when it
+                                 // publishes one (Anthropic: 128_000; absent for Ollama)
 }
 
 interface InferenceResponse {
   text: string;
   stopReason: 'end_turn' | 'max_tokens' | 'stop_sequence' | string;
+  usage?: TokenUsage;   // the PROVIDER's own token counts — never estimated;
+}                       // absent means unreported, not zero
+
+interface TokenUsage {
+  inputTokens: number;
+  outputTokens: number;
 }
 ```
 
+`StructuredResponse<T>` carries the same optional `usage`, so a consumer can pair what a call yielded with what it cost from one return value.
+
 ### Structured generation
 
 `generateStructured(prompt, maxTokens, temperature, elementSchema)` returns **parsed array elements**, not text — the JSON guarantee lives in the return type (`StructuredResponse<T> = { items: T[]; stopReason }`), not in a comment:
@@ -123,7 +148,7 @@ Each implementation honors the contract with its provider's mechanism:
 - **Ollama**: grammar-constrained sampling — the request's `format` field carries the caller's element schema wrapped in an array schema.
 - **Anthropic**: response-level structured output — `output_config.format` carries the caller's element schema under an **array root** (accepted on both live-config models; `.plans/spikes/output-config-array-root.md`), so the response text IS the schema-conforming JSON. No tools, no wrapper, no unwrap.
 
-A response that cannot be read as an array — the SDK delivering unparsed tool input as a string, a missing array, an unhonoured grammar — **throws** (`Structured response could not be read: …`). It is never coerced to `[]`: an empty extraction is a legitimate, distinct outcome, and conflating the two silently discards real data (STRUCTURED-INFERENCE).
+A response that cannot be read as an array — unparseable output, a missing array, an unhonoured grammar, or an **empty response** (which throws with the stop reason that produced it, so thinking-exhaustion classifies as the truncation it is) — **throws a typed `StructuredReadError`** carrying the provider's `stopReason`, because the cause classifies differently downstream: `max_tokens` is truncation (an identical retry truncates identically, so detection subdivides), `'unknown'` (no stop reason at all) is measured size-correlated on real documents (detection subdivides that too, while keeping it retryable), and everything else is model misbehavior a retry may fix. It is never coerced to `[]`: an empty extraction is a legitimate, distinct outcome, and conflating the two silently discards real data.
 
 Current callers all expect arrays (entity extraction, motivation detection). If an object-emitting caller appears, `generateStructured` grows a sibling, not an option — see the notes in [src/interface.ts](src/interface.ts).
 
@@ -131,18 +156,20 @@ Current callers all expect arrays (entity extraction, motivation detection). If
 
 `limits()` publishes the provider's **actual** context/output ceilings for the configured model — discovered from the provider itself, never hand-maintained constants:
 
-- **Anthropic**: the Models API (`models.retrieve`) — `max_input_tokens` / `max_tokens`.
-- **Ollama**: `POST /api/show` — the model's context window. Input and output share that window, so it is published as both fields (`maxOutputTokens === contextTokens` signals a shared window).
+- **Anthropic**: the Models API (`models.retrieve`) — `max_input_tokens` / `max_tokens` — plus `outputTokensPerHour: 128_000`, the SDK's own worst-case rate model (the `calculateNonstreamingTimeout` constant): the one duration statement the provider surface makes, which detection's duration-safe budgets derive from.
+- **Ollama**: `POST /api/show` — the model's context window. Input and output share that window, so it is published as both fields (`maxOutputTokens === contextTokens` signals a shared window). No rate is published — local hardware's rate is unknowable a priori. Absence does **not** mean no duration bound: the detection consumer applies its own conservative assumed floor rate instead, because an unbounded output budget turned model repetition loops into hour-long transient burns.
 
 Discovery is lazy and cached per client; a failed discovery is **not** cached — the next call retries. When ceilings cannot be determined (unknown model, endpoint unreachable), `limits()` **throws**: fail-loud, never a guessed floor.
 
 Two request-time behaviors ride on the limits:
 - **Ollama sets `num_ctx` explicitly** on every generate request — sized to the prompt estimate + output budget, capped at the model window. Without it, Ollama's model-*default* window silently clips large prompts. A request that genuinely cannot fit **throws** instead of being clipped.
+- **The Ollama adapter owns its transport timeouts.** With `stream: false`, Ollama sends no response headers until generation completes, and Node's default fetch would kill any call generating longer than ~5 minutes (undici's `headersTimeout`) — a ceiling below every deliberate bound, owned by nobody. Generate requests run on a per-request undici@7 dispatcher with those timeouts disabled; the caller's `AbortSignal` is the one bound. The undici `^7` pin is load-bearing (the built-in fetch rejects an undici@8 Agent) and test-gated.
+- **Cloud-routed Ollama models are reported honestly, not corrected**: hidden thinking returned despite `think: false` is surfaced on the response and warned (it inflates `eval_count`, which is documented at the field), and the structured `format` is advisory rather than grammar-enforced on that path — violations surface as `StructuredReadError`.
 - **Anthropic streams internally** above the SDK's non-streaming output ceiling (≈21K tokens) — same interface, same response shape.
 
 ### `MockInferenceClient`
 
-A scripted test double ([src/implementations/mock.ts](src/implementations/mock.ts)): construct it with a list of canned responses, then inspect `calls` (recorded prompt/maxTokens/temperature/options per invocation). `reset()` and `setResponses()` helpers included. An optional third constructor argument injects `InferenceLimits` for chunking/budget tests; the default is generous (1M/1M) so ordinary tests never trip window guards.
+A scripted test double ([src/implementations/mock.ts](src/implementations/mock.ts)): construct it with a list of canned responses, then inspect `calls` (recorded prompt/maxTokens/temperature/options per invocation). `reset()` and `setResponses()` helpers included. An optional third constructor argument injects `InferenceLimits` for chunking/budget tests; the default is generous (1M/1M window plus a generous published rate) so ordinary tests never trip window guards, duration caps, or the count-verifier. Its capabilities are deterministic-test defaults — `maxConcurrency: 1`, `verifyDetectionYield: false` — so a test exercising concurrency or verification declares its own client rather than paying a surprise call.
 
 ```typescript
 import { MockInferenceClient } from '@semiont/inference';
@@ -195,10 +222,14 @@ Every generation records a usage metric through `@semiont/observability`'s `reco
 
 ### Adding a New Provider
 
-1. Implement `InferenceClient` interface in `src/implementations/`
+1. Implement the `InferenceClient` interface in `src/implementations/` — including a
+   position on each declared capability (`maxConcurrency`: does this provider gain from
+   concurrent independent calls? `verifyDetectionYield`: should detection count-verify
+   its extractions?). The capability pins in `factory.test.ts` fail until you take one.
 2. Add type to `InferenceClientType` union in `src/factory.ts`
 3. Add case in `createInferenceClient()` switch
-4. Application code in `@semiont/make-meaning` requires no changes
+4. Application code in `@semiont/make-meaning` and `@semiont/jobs` requires no changes —
+   consumers read capabilities off the contract instead of switching on provider identity
 
 ## Dependencies
 

package/dist/index.d.ts

@@ -1,14 +1,25 @@
 import { Logger } from '@semiont/core';
 
+/**
+ * What the call actually cost, as the PROVIDER counted it — never estimated
+ * here. Optional because a provider may not report it (and a call that fails
+ * before generating has nothing to report); absent means unknown, and a
+ * consumer must treat it as unknown rather than substituting a guess.
+ */
+interface TokenUsage {
+    inputTokens: number;
+    outputTokens: number;
+}
 interface InferenceResponse {
     text: string;
     stopReason: 'end_turn' | 'max_tokens' | 'stop_sequence' | string;
+    usage?: TokenUsage;
 }
 /**
  * Raw JSON Schema for ONE array element of a structured generation — a plain
  * object, not a TS type and not a validator instance. Both providers consume
- * JSON Schema directly (Anthropic nests it under the forced tool's
- * `input_schema`; Ollama sends it as `format: { type: 'array', items: … }`),
+ * JSON Schema directly (Anthropic as the array-root schema under
+ * `output_config.format`; Ollama as `format: { type: 'array', items: … }`),
  * so anything richer would abstract one shape with two consumers.
  *
  * Constrain it to what both providers enforce: objects,
@@ -35,6 +46,7 @@ type ElementSchema = Record<string, unknown>;
 interface StructuredResponse<T> {
     items: T[];
     stopReason: 'end_turn' | 'max_tokens' | 'stop_sequence' | string;
+    usage?: TokenUsage;
 }
 /**
  * A provider's actual ceilings for the configured model, discovered from the
@@ -52,12 +64,80 @@ interface InferenceLimits {
     contextTokens: number;
     /** Maximum output tokens per generation. */
     maxOutputTokens: number;
+    /**
+     * The provider's own worst-case output-rate model, in output tokens per
+     * hour, when it publishes one. Anthropic's SDK projects a call's maximum
+     * duration as `max_tokens / rate` (client.js `calculateNonstreamingTimeout`,
+     * 128_000/hour) and refuses non-streaming calls projected past 10 minutes —
+     * the one duration statement that provider surface makes. Consumers with
+     * their own call deadline derive a duration-safe output budget from it
+     * (ABANDONED-INFERENCE P4). Absent for providers whose rates are
+     * unknowable a priori (Ollama — local hardware). Absence does NOT mean no
+     * duration bound: the detection consumer applies its own conservative
+     * assumed floor rate instead (OLLAMA-DETECTION-TESTING P3b) — an unbounded
+     * budget turned model repetition loops into hour-long transient burns.
+     */
+    outputTokensPerHour?: number;
+}
+/**
+ * Thrown when a structured generation's response cannot be read as the
+ * requested array — never coerced to `[]` (empty is a legitimate, distinct
+ * outcome). One class for every implementation, because the message shape
+ * and the classification contract must not diverge between providers.
+ *
+ * Carries the provider's stop reason because the cause classifies
+ * differently downstream: `max_tokens` means the JSON was cut off by the
+ * output budget — the same input truncates the same way, so a retry is
+ * guaranteed waste — while any other reason is model misbehavior a retry
+ * may legitimately fix.
+ *
+ * Also thrown — on either generation path — when a response arrives EMPTY:
+ * a thinking model can exhaust the whole output budget on hidden reasoning
+ * before its first response character (measured live, gpt-oss:120b-cloud
+ * 2026-09-05). Truncated-to-nothing is still truncation, and it needs the
+ * same stop-reason ride to classify correctly.
+ */
+declare class StructuredReadError extends Error {
+    readonly stopReason: string;
+    readonly name = "StructuredReadError";
+    constructor(detail: string, stopReason: string, options?: ErrorOptions);
 }
 interface InferenceClient {
     /** Provider type identifier (e.g. 'anthropic', 'ollama') */
     readonly type: string;
     /** Model identifier used for generation (e.g. 'claude-opus-4-6', 'llama3') */
     readonly modelId: string;
+    /**
+     * How many INDEPENDENT inference calls a caller should run concurrently
+     * against this provider for a throughput gain (DETECTION-QUALITY-THROUGHPUT
+     * P6 — detection's per-type fan-out reads this).
+     *
+     * This is a property of the provider's economics, which is why it lives on
+     * the provider and not in the caller. A HOSTED API whose per-account rate
+     * limit sits far above one job's usage has real spare capacity, so >1
+     * genuinely parallelizes. A LOCAL single-model server (Ollama) is 1: its
+     * throughput is hardware-bound, so concurrent requests only queue or split
+     * one GPU — no aggregate speedup, and N live KV-cache contexts is memory
+     * pressure that can OOM. There is no honest default across those two worlds,
+     * so this is required, not optional.
+     *
+     * Hard-coded per implementation for now; the natural seam for future
+     * per-provider or admin tuning (a value that later comes from config changes
+     * only where this is SET, not the callers).
+     */
+    readonly maxConcurrency: number;
+    /**
+     * Whether detection should run the count-verifier against this provider's
+     * extractions (OLLAMA-DETECTION-TESTING P3c; universalized to every real
+     * provider by user ruling 2026-09-05 — unverified completeness is not a
+     * savings). Declared HERE, per implementation, because @semiont/jobs does no
+     * provider-specific switching (architecture ruling, same date): whatever
+     * varies by provider is a capability on this contract, like
+     * `maxConcurrency`. The mock alone defaults false, so deterministic tests
+     * opt in explicitly rather than paying a queue-popping count call by
+     * surprise.
+     */
+    readonly verifyDetectionYield: boolean;
     /**
      * The provider's actual context/output ceilings for `modelId`. Discovered
      * lazily on first call and cached for the client's lifetime; a failed
@@ -67,13 +147,23 @@ interface InferenceClient {
      */
     limits(): Promise<InferenceLimits>;
     /**
-     * Generate text from a prompt (simple interface)
+     * Generate text from a prompt (simple interface).
+     *
+     * `signal` (here and on every generation method — a trailing optional
+     * parameter, deliberately not an options bag; STRUCTURED-INFERENCE removed
+     * that shape on purpose): true cancellation, ABANDONED-INFERENCE P1.
+     * Implementations MUST thread it to their transport so an abort tears down
+     * the underlying request — and, for SDKs with internal retry loops, ends
+     * those too — rejecting promptly. Accepting the parameter and ignoring it
+     * is a defect worse than not having it: cancellation tests pass against
+     * such an adapter while zombie requests keep running (and billing) in
+     * production.
      */
-    generateText(prompt: string, maxTokens: number, temperature: number): Promise<string>;
+    generateText(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<string>;
     /**
      * Generate text with detailed response information
      */
-    generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number): Promise<InferenceResponse>;
+    generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<InferenceResponse>;
     /**
      * Generate a JSON array whose elements satisfy `elementSchema`, as parsed
      * values — the structured counterpart of `generateTextWithMetadata`, and
@@ -87,7 +177,7 @@ interface InferenceClient {
      * "Structured response could not be read" error — they never coerce to
      * `[]`, because empty is a legitimate, distinct outcome.
      */
-    generateStructured<T>(prompt: string, maxTokens: number, temperature: number, elementSchema: ElementSchema): Promise<StructuredResponse<T>>;
+    generateStructured<T>(prompt: string, maxTokens: number, temperature: number, elementSchema: ElementSchema, signal?: AbortSignal): Promise<StructuredResponse<T>>;
 }
 
 type InferenceClientType = 'anthropic' | 'ollama';
@@ -102,6 +192,8 @@ declare function createInferenceClient(config: InferenceClientConfig, logger?: L
 
 declare class AnthropicInferenceClient implements InferenceClient {
     readonly type: "anthropic";
+    readonly maxConcurrency = 4;
+    readonly verifyDetectionYield = true;
     readonly modelId: string;
     private client;
     private logger?;
@@ -111,9 +203,9 @@ declare class AnthropicInferenceClient implements InferenceClient {
     private discover;
     private discoverModel;
     private requestMessage;
-    generateText(prompt: string, maxTokens: number, temperature: number): Promise<string>;
-    generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number): Promise<InferenceResponse>;
-    generateStructured<T>(prompt: string, maxTokens: number, temperature: number, elementSchema: ElementSchema): Promise<StructuredResponse<T>>;
+    generateText(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<string>;
+    generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<InferenceResponse>;
+    generateStructured<T>(prompt: string, maxTokens: number, temperature: number, elementSchema: ElementSchema, signal?: AbortSignal): Promise<StructuredResponse<T>>;
     /** Issue the request, recording an error metric if the transport throws. */
     private recordedRequest;
     private recordError;
@@ -121,6 +213,8 @@ declare class AnthropicInferenceClient implements InferenceClient {
 
 declare class OllamaInferenceClient implements InferenceClient {
     readonly type: "ollama";
+    readonly maxConcurrency = 1;
+    readonly verifyDetectionYield = true;
     readonly modelId: string;
     private baseURL;
     private logger?;
@@ -128,15 +222,17 @@ declare class OllamaInferenceClient implements InferenceClient {
     constructor(model: string, baseURL?: string, logger?: Logger);
     limits(): Promise<InferenceLimits>;
     private discoverLimits;
-    generateText(prompt: string, maxTokens: number, temperature: number): Promise<string>;
-    generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number): Promise<InferenceResponse>;
-    generateStructured<T>(prompt: string, maxTokens: number, temperature: number, elementSchema: ElementSchema): Promise<StructuredResponse<T>>;
+    generateText(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<string>;
+    generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<InferenceResponse>;
+    generateStructured<T>(prompt: string, maxTokens: number, temperature: number, elementSchema: ElementSchema, signal?: AbortSignal): Promise<StructuredResponse<T>>;
     private generate;
 }
 
 declare class MockInferenceClient implements InferenceClient {
     readonly type: "mock";
     readonly modelId: "mock-model";
+    readonly maxConcurrency = 1;
+    readonly verifyDetectionYield = false;
     private responses;
     private responseIndex;
     private stopReasons;
@@ -149,8 +245,8 @@ declare class MockInferenceClient implements InferenceClient {
     }>;
     constructor(responses?: string[], stopReasons?: string[], limits?: InferenceLimits);
     limits(): Promise<InferenceLimits>;
-    generateText(prompt: string, maxTokens: number, temperature: number): Promise<string>;
-    generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number): Promise<InferenceResponse>;
+    generateText(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<string>;
+    generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<InferenceResponse>;
     /**
      * Structured surface: pops the same responses queue and PARSES the entry,
      * mirroring the real contract — a queued string that is not a JSON array
@@ -158,11 +254,11 @@ declare class MockInferenceClient implements InferenceClient {
      * queuing it (`setResponses(['not json'])`). The element schema is recorded
      * on `calls` so tests can assert what the caller declared.
      */
-    generateStructured<T>(prompt: string, maxTokens: number, temperature: number, elementSchema: ElementSchema): Promise<StructuredResponse<T>>;
+    generateStructured<T>(prompt: string, maxTokens: number, temperature: number, elementSchema: ElementSchema, signal?: AbortSignal): Promise<StructuredResponse<T>>;
     private nextResponse;
     reset(): void;
     setResponses(responses: string[], stopReasons?: string[]): void;
 }
 
-export { AnthropicInferenceClient, MockInferenceClient, OllamaInferenceClient, createInferenceClient };
-export type { ElementSchema, InferenceClient, InferenceClientConfig, InferenceClientType, InferenceLimits, InferenceResponse, StructuredResponse };
+export { AnthropicInferenceClient, MockInferenceClient, OllamaInferenceClient, StructuredReadError, createInferenceClient };
+export type { ElementSchema, InferenceClient, InferenceClientConfig, InferenceClientType, InferenceLimits, InferenceResponse, StructuredResponse, TokenUsage };

package/dist/index.js

@@ -2,9 +2,32 @@
 import Anthropic from "@anthropic-ai/sdk";
 import { isObject } from "@semiont/core";
 import { recordInferenceUsage } from "@semiont/observability";
-var NONSTREAMING_MAX_OUTPUT_TOKENS = Math.floor(128e3 / 6);
+
+// src/interface.ts
+var StructuredReadError = class extends Error {
+  constructor(detail, stopReason, options) {
+    super(`Structured response could not be read: ${detail} (stop_reason: ${stopReason})`, options);
+    this.stopReason = stopReason;
+  }
+  stopReason;
+  name = "StructuredReadError";
+};
+
+// src/implementations/anthropic.ts
+var OUTPUT_TOKENS_PER_HOUR = 128e3;
+var NONSTREAMING_MAX_OUTPUT_TOKENS = Math.floor(OUTPUT_TOKENS_PER_HOUR / 6);
 var AnthropicInferenceClient = class {
   type = "anthropic";
+  // Hosted API: a single detection job uses a sliver of the account rate limit
+  // (measured 2026-09-04: ~1 request / 72 s, zero 429s at 4 concurrent types),
+  // so independent calls genuinely parallelize. Conservative until an 8-way run
+  // measures the next step (DETECTION-QUALITY-THROUGHPUT P6).
+  maxConcurrency = 4;
+  // Universal for real providers (user ruling 2026-09-05): "no observed
+  // collapse" here was absence-of-looking, and the unexplained ~2× yield gap
+  // vs gemma on the same document is exactly what verification answers. The
+  // ~2× billed input is the accepted cost.
+  verifyDetectionYield = true;
   modelId;
   client;
   logger;
@@ -42,21 +65,25 @@ var AnthropicInferenceClient = class {
     const raw = info;
     const structuredOutputsSupported = isObject(raw) && isObject(raw["capabilities"]) && isObject(raw["capabilities"]["structured_outputs"]) && raw["capabilities"]["structured_outputs"]["supported"] === true;
     return {
-      limits: { contextTokens: info.max_input_tokens, maxOutputTokens: info.max_tokens },
+      limits: {
+        contextTokens: info.max_input_tokens,
+        maxOutputTokens: info.max_tokens,
+        outputTokensPerHour: OUTPUT_TOKENS_PER_HOUR
+      },
       structuredOutputsSupported
     };
   }
-  requestMessage(params) {
+  requestMessage(params, signal) {
     if (params.max_tokens > NONSTREAMING_MAX_OUTPUT_TOKENS) {
-      return this.client.messages.stream(params).finalMessage();
+      return this.client.messages.stream(params, { signal }).finalMessage();
     }
-    return this.client.messages.create(params);
+    return this.client.messages.create(params, { signal });
   }
-  async generateText(prompt, maxTokens, temperature) {
-    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature);
+  async generateText(prompt, maxTokens, temperature, signal) {
+    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature, signal);
     return response.text;
   }
-  async generateTextWithMetadata(prompt, maxTokens, temperature) {
+  async generateTextWithMetadata(prompt, maxTokens, temperature, signal) {
     this.logger?.debug("Generating text with inference client", {
       model: this.modelId,
       promptLength: prompt.length,
@@ -70,7 +97,7 @@ var AnthropicInferenceClient = class {
       messages: [{ role: "user", content: prompt }]
     };
     const start = performance.now();
-    const response = await this.recordedRequest(params, start);
+    const response = await this.recordedRequest(params, start, signal);
     const textContent = response.content.find((c) => c.type === "text");
     if (!textContent || textContent.type !== "text") {
       this.recordError(start, response);
@@ -92,14 +119,15 @@ var AnthropicInferenceClient = class {
     this.logger?.info("Text generation completed", {
       model: this.modelId,
       textLength: text.length,
-      stopReason: response.stop_reason
+      stopReason: response.stop_reason,
+      requestId: requestIdOf(response)
     });
     return {
       text,
       stopReason: response.stop_reason || "unknown"
     };
   }
-  async generateStructured(prompt, maxTokens, temperature, elementSchema) {
+  async generateStructured(prompt, maxTokens, temperature, elementSchema, signal) {
     const discovery = await this.discover();
     if (!discovery.structuredOutputsSupported) {
       throw new Error(
@@ -127,7 +155,7 @@ var AnthropicInferenceClient = class {
       }
     };
     const start = performance.now();
-    const response = await this.recordedRequest(params, start);
+    const response = await this.recordedRequest(params, start, signal);
     const textContent = response.content.find((c) => c.type === "text");
     if (!textContent || textContent.type !== "text") {
       this.recordError(start, response);
@@ -147,10 +175,7 @@ var AnthropicInferenceClient = class {
         textLength: textContent.text.length,
         stopReason: response.stop_reason
       });
-      throw new Error(
-        `Structured response could not be read: response is not valid JSON (stop_reason: ${response.stop_reason})`,
-        { cause: err }
-      );
+      throw new StructuredReadError("response is not valid JSON", response.stop_reason || "unknown", { cause: err });
     }
     if (!Array.isArray(parsed)) {
       this.recordError(start, response);
@@ -159,9 +184,7 @@ var AnthropicInferenceClient = class {
         parsedType: typeof parsed,
         stopReason: response.stop_reason
       });
-      throw new Error(
-        `Structured response could not be read: parsed to ${typeof parsed}, not an array (stop_reason: ${response.stop_reason})`
-      );
+      throw new StructuredReadError(`parsed to ${typeof parsed}, not an array`, response.stop_reason || "unknown");
     }
     recordInferenceUsage({
       provider: this.type,
@@ -174,17 +197,19 @@ var AnthropicInferenceClient = class {
     this.logger?.info("Structured generation completed", {
       model: this.modelId,
       items: parsed.length,
-      stopReason: response.stop_reason
+      stopReason: response.stop_reason,
+      requestId: requestIdOf(response)
     });
     return {
       items: parsed,
-      stopReason: response.stop_reason || "unknown"
+      stopReason: response.stop_reason || "unknown",
+      ...usageOf(response)
     };
   }
   /** Issue the request, recording an error metric if the transport throws. */
-  async recordedRequest(params, start) {
+  async recordedRequest(params, start, signal) {
     try {
-      return await this.requestMessage(params);
+      return await this.requestMessage(params, signal);
     } catch (err) {
       recordInferenceUsage({
         provider: this.type,
@@ -206,14 +231,34 @@ var AnthropicInferenceClient = class {
     });
   }
 };
+function requestIdOf(response) {
+  if (isObject(response) && typeof response["_request_id"] === "string") {
+    return response["_request_id"];
+  }
+  return void 0;
+}
+function usageOf(response) {
+  const { input_tokens, output_tokens } = response.usage ?? {};
+  if (input_tokens === void 0 || output_tokens === void 0) return {};
+  return { usage: { inputTokens: input_tokens, outputTokens: output_tokens } };
+}
 
 // src/implementations/ollama.ts
+import { Agent } from "undici";
 import { estimateTokens, isNumber, isObject as isObject2 } from "@semiont/core";
 import { recordInferenceUsage as recordInferenceUsage2 } from "@semiont/observability";
+var unboundedTransport = new Agent({ headersTimeout: 0, bodyTimeout: 0 });
 var NUM_CTX_ESTIMATE_SLACK = 0.2;
 var NUM_CTX_TEMPLATE_ALLOWANCE = 64;
 var OllamaInferenceClient = class {
   type = "ollama";
+  // Local single model: generation throughput is hardware-bound, so concurrent
+  // requests queue or split one GPU for no aggregate speedup — and each live
+  // context costs KV-cache memory. Detection runs its types sequentially here
+  // (DETECTION-QUALITY-THROUGHPUT P6).
+  maxConcurrency = 1;
+  // Where the collapse risk was MEASURED (F7) — the verifier's original home.
+  verifyDetectionYield = true;
   modelId;
   baseURL;
   logger;
@@ -251,15 +296,15 @@ var OllamaInferenceClient = class {
     }
     return { contextTokens, maxOutputTokens: contextTokens };
   }
-  async generateText(prompt, maxTokens, temperature) {
-    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature);
+  async generateText(prompt, maxTokens, temperature, signal) {
+    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature, signal);
     return response.text;
   }
-  async generateTextWithMetadata(prompt, maxTokens, temperature) {
-    return this.generate(prompt, maxTokens, temperature, void 0);
+  async generateTextWithMetadata(prompt, maxTokens, temperature, signal) {
+    return this.generate(prompt, maxTokens, temperature, void 0, signal);
   }
-  async generateStructured(prompt, maxTokens, temperature, elementSchema) {
-    const response = await this.generate(prompt, maxTokens, temperature, elementSchema);
+  async generateStructured(prompt, maxTokens, temperature, elementSchema, signal) {
+    const response = await this.generate(prompt, maxTokens, temperature, elementSchema, signal);
     let parsed;
     try {
       parsed = JSON.parse(response.text);
@@ -269,10 +314,7 @@ var OllamaInferenceClient = class {
         textLength: response.text.length,
         stopReason: response.stopReason
       });
-      throw new Error(
-        `Structured response could not be read: response is not valid JSON (stop_reason: ${response.stopReason})`,
-        { cause: err }
-      );
+      throw new StructuredReadError("response is not valid JSON", response.stopReason, { cause: err });
     }
     if (!Array.isArray(parsed)) {
       this.logger?.error("Structured response could not be read", {
@@ -280,13 +322,11 @@ var OllamaInferenceClient = class {
         parsedType: typeof parsed,
         stopReason: response.stopReason
       });
-      throw new Error(
-        `Structured response could not be read: parsed to ${typeof parsed}, not an array (stop_reason: ${response.stopReason})`
-      );
+      throw new StructuredReadError(`parsed to ${typeof parsed}, not an array`, response.stopReason);
     }
-    return { items: parsed, stopReason: response.stopReason };
+    return { items: parsed, stopReason: response.stopReason, ...response.usage ? { usage: response.usage } : {} };
   }
-  async generate(prompt, maxTokens, temperature, elementSchema) {
+  async generate(prompt, maxTokens, temperature, elementSchema, signal) {
     this.logger?.debug("Generating text with Ollama", {
       model: this.modelId,
       promptLength: prompt.length,
@@ -326,7 +366,9 @@ var OllamaInferenceClient = class {
       res = await fetch(url, {
         method: "POST",
         headers: { "Content-Type": "application/json" },
-        body: JSON.stringify(body)
+        body: JSON.stringify(body),
+        signal,
+        dispatcher: unboundedTransport
       });
     } catch (err) {
       recordInferenceUsage2({
@@ -353,6 +395,13 @@ var OllamaInferenceClient = class {
       throw new Error(`Ollama API error (${res.status}): ${body2}`);
     }
     const data = await res.json();
+    const stopReason = mapStopReason(data.done_reason);
+    if (data.thinking) {
+      this.logger?.warn("Model produced hidden thinking despite think:false", {
+        model: this.modelId,
+        thinkingChars: data.thinking.length
+      });
+    }
     if (!data.response) {
       recordInferenceUsage2({
         provider: this.type,
@@ -362,8 +411,12 @@ var OllamaInferenceClient = class {
         inputTokens: data.prompt_eval_count,
         outputTokens: data.eval_count
       });
-      this.logger?.error("Empty response from Ollama", { model: this.modelId });
-      throw new Error("Empty response from Ollama");
+      this.logger?.error("Empty response from Ollama", {
+        model: this.modelId,
+        stopReason,
+        thinkingChars: data.thinking?.length
+      });
+      throw new StructuredReadError("response is empty", stopReason);
     }
     recordInferenceUsage2({
       provider: this.type,
@@ -373,7 +426,6 @@ var OllamaInferenceClient = class {
       inputTokens: data.prompt_eval_count,
       outputTokens: data.eval_count
     });
-    const stopReason = mapStopReason(data.done_reason);
     this.logger?.info("Text generation completed", {
       model: this.modelId,
       textLength: data.response.length,
@@ -381,7 +433,8 @@ var OllamaInferenceClient = class {
     });
     return {
       text: data.response,
-      stopReason
+      stopReason,
+      ...data.prompt_eval_count !== void 0 && data.eval_count !== void 0 ? { usage: { inputTokens: data.prompt_eval_count, outputTokens: data.eval_count } } : {}
     };
   }
 };
@@ -439,11 +492,22 @@ function createInferenceClient(config, logger) {
 // src/implementations/mock.ts
 var GENEROUS_LIMITS = {
   contextTokens: 1e6,
-  maxOutputTokens: 1e6
+  maxOutputTokens: 1e6,
+  // Generous RATE too: without one the mock reads as rate-silent, which since
+  // OLLAMA-DETECTION-TESTING P3b/P3c opts consumers into the assumed duration
+  // floor and the count-verifier. Tests exercise those by injecting
+  // rate-silent limits deliberately, never by the default.
+  outputTokensPerHour: 36e8
 };
 var MockInferenceClient = class {
   type = "mock";
   modelId = "mock-model";
+  // Deterministic default for tests; a test exercising concurrency sets its own.
+  maxConcurrency = 1;
+  // Deterministic default: a count call pops the shared response queue, so
+  // tests exercising the verifier declare their own true rather than every
+  // consumer paying it by surprise.
+  verifyDetectionYield = false;
   responses = [];
   responseIndex = 0;
   stopReasons = [];
@@ -457,11 +521,12 @@ var MockInferenceClient = class {
   async limits() {
     return this.injectedLimits;
   }
-  async generateText(prompt, maxTokens, temperature) {
-    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature);
+  async generateText(prompt, maxTokens, temperature, signal) {
+    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature, signal);
     return response.text;
   }
-  async generateTextWithMetadata(prompt, maxTokens, temperature) {
+  async generateTextWithMetadata(prompt, maxTokens, temperature, signal) {
+    throwIfAborted(signal);
     this.calls.push({ prompt, maxTokens, temperature });
     return this.nextResponse();
   }
@@ -472,22 +537,18 @@ var MockInferenceClient = class {
    * queuing it (`setResponses(['not json'])`). The element schema is recorded
    * on `calls` so tests can assert what the caller declared.
    */
-  async generateStructured(prompt, maxTokens, temperature, elementSchema) {
+  async generateStructured(prompt, maxTokens, temperature, elementSchema, signal) {
+    throwIfAborted(signal);
     this.calls.push({ prompt, maxTokens, temperature, elementSchema });
     const { text, stopReason } = this.nextResponse();
     let parsed;
     try {
       parsed = JSON.parse(text);
     } catch (err) {
-      throw new Error(
-        `Structured response could not be read: response is not valid JSON (stop_reason: ${stopReason})`,
-        { cause: err }
-      );
+      throw new StructuredReadError("response is not valid JSON", stopReason, { cause: err });
     }
     if (!Array.isArray(parsed)) {
-      throw new Error(
-        `Structured response could not be read: parsed to ${typeof parsed}, not an array (stop_reason: ${stopReason})`
-      );
+      throw new StructuredReadError(`parsed to ${typeof parsed}, not an array`, stopReason);
     }
     return { items: parsed, stopReason };
   }
@@ -510,10 +571,16 @@ var MockInferenceClient = class {
     this.responseIndex = 0;
   }
 };
+function throwIfAborted(signal) {
+  if (signal?.aborted) {
+    throw new DOMException("This operation was aborted", "AbortError");
+  }
+}
 export {
   AnthropicInferenceClient,
   MockInferenceClient,
   OllamaInferenceClient,
+  StructuredReadError,
   createInferenceClient
 };
 //# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/implementations/anthropic.ts","../src/implementations/ollama.ts","../src/factory.ts","../src/implementations/mock.ts"],"sourcesContent":["// Anthropic Claude implementation of InferenceClient interface\n\nimport Anthropic from '@anthropic-ai/sdk';\nimport { isObject, type Logger } from '@semiont/core';\nimport { recordInferenceUsage } from '@semiont/observability';\nimport { ElementSchema, InferenceClient, InferenceLimits, InferenceResponse, StructuredResponse } from '../interface.js';\n\n// The SDK refuses non-streaming create() calls whose projected duration\n// exceeds its 10-minute timeout: it throws when\n// (60min × max_tokens) / 128_000 > 10min, i.e. above 128_000/6 ≈ 21,333\n// output tokens (client.js, calculateNonstreamingTimeout). Above that we\n// stream internally and assemble the final message — same request shape,\n// same response handling, same interface.\nconst NONSTREAMING_MAX_OUTPUT_TOKENS = Math.floor(128_000 / 6);\n\n// Structured generation rides `output_config.format` — response-level\n// structured output: the response TEXT is the schema-conforming JSON, with a\n// top-level ARRAY root (accepted on both live-config models — spike\n// 2026-08-06, `.plans/spikes/output-config-array-root.md`). This replaced\n// the pre-structured-outputs scaffolding: a forced `emit_json_array` tool\n// whose object-only input required an `items` wrapper and an unwrap — and\n// the unwrap was the exact line that silently coerced an unreadable payload\n// to `[]` (STRUCTURED-INFERENCE §Problem). There is no tool-input\n// accumulation step left for the SDK to hand over unparsed; the read path is\n// now the same parse-and-verify shape as Ollama's.\n\n/**\n * Everything one Models API call teaches us about the configured model. The\n * capability stays PRIVATE to this client: its only consumer is the gate in\n * `generateStructured`, so it does not cross the `InferenceClient` interface\n * (no speculative surface — widen `InferenceLimits` only when an external\n * consumer exists).\n */\ninterface ModelDiscovery {\n  limits: InferenceLimits;\n  structuredOutputsSupported: boolean;\n}\n\nexport class AnthropicInferenceClient implements InferenceClient {\n  readonly type = 'anthropic' as const;\n  readonly modelId: string;\n  private client: Anthropic;\n  private logger?: Logger;\n  private discoveryPromise?: Promise<ModelDiscovery>;\n\n  constructor(apiKey: string, model: string, baseURL?: string, logger?: Logger) {\n    this.client = new Anthropic({\n      apiKey,\n      baseURL: baseURL || 'https://api.anthropic.com',\n    });\n    this.modelId = model;\n    this.logger = logger;\n  }\n\n  limits(): Promise<InferenceLimits> {\n    return this.discover().then(d => d.limits);\n  }\n\n  private discover(): Promise<ModelDiscovery> {\n    if (!this.discoveryPromise) {\n      this.discoveryPromise = this.discoverModel().catch((err: unknown) => {\n        // Never cache a failed discovery — a transient outage would otherwise\n        // pin every future call to the same rejection.\n        this.discoveryPromise = undefined;\n        throw err;\n      });\n    }\n    return this.discoveryPromise;\n  }\n\n  private async discoverModel(): Promise<ModelDiscovery> {\n    // The Models API publishes the actual ceilings AND capabilities per\n    // model — no hand-maintained table to go stale when a new model ships,\n    // and the API's own metadata outranks documentation prose when the two\n    // disagree (measured: the docs' supported-model list was stale while\n    // `capabilities.structured_outputs.supported` was correct).\n    const info = await this.client.models.retrieve(this.modelId).catch((err: unknown) => {\n      throw new Error(\n        `Failed to discover model limits for '${this.modelId}' from the Models API`,\n        { cause: err },\n      );\n    });\n    if (info.max_input_tokens == null || info.max_tokens == null) {\n      throw new Error(`Models API reports no context/output ceilings for '${this.modelId}'`);\n    }\n    // `capabilities` is not declared on the SDK's ModelInfo type — narrow\n    // through the core guards rather than casting. Absent metadata reads as\n    // unsupported: the gate then refuses loudly (D4), never guesses.\n    const raw: unknown = info;\n    const structuredOutputsSupported =\n      isObject(raw) &&\n      isObject(raw['capabilities']) &&\n      isObject(raw['capabilities']['structured_outputs']) &&\n      raw['capabilities']['structured_outputs']['supported'] === true;\n    return {\n      limits: { contextTokens: info.max_input_tokens, maxOutputTokens: info.max_tokens },\n      structuredOutputsSupported,\n    };\n  }\n\n  private requestMessage(params: Anthropic.MessageCreateParamsNonStreaming): Promise<Anthropic.Message> {\n    if (params.max_tokens > NONSTREAMING_MAX_OUTPUT_TOKENS) {\n      return this.client.messages.stream(params).finalMessage();\n    }\n    return this.client.messages.create(params);\n  }\n\n  async generateText(prompt: string, maxTokens: number, temperature: number): Promise<string> {\n    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature);\n    return response.text;\n  }\n\n  async generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number): Promise<InferenceResponse> {\n    this.logger?.debug('Generating text with inference client', {\n      model: this.modelId,\n      promptLength: prompt.length,\n      maxTokens,\n      temperature,\n    });\n\n    const params: Anthropic.MessageCreateParamsNonStreaming = {\n      model: this.modelId,\n      max_tokens: maxTokens,\n      temperature,\n      messages: [{ role: 'user', content: prompt }],\n    };\n\n    const start = performance.now();\n    const response = await this.recordedRequest(params, start);\n\n    const textContent = response.content.find(c => c.type === 'text');\n    if (!textContent || textContent.type !== 'text') {\n      this.recordError(start, response);\n      this.logger?.error('No text content in inference response', {\n        model: this.modelId,\n        contentTypes: response.content.map(c => c.type)\n      });\n      throw new Error('No text content in inference response');\n    }\n    const text = textContent.text;\n\n    recordInferenceUsage({\n      provider: this.type,\n      model: this.modelId,\n      durationMs: performance.now() - start,\n      outcome: 'success',\n      inputTokens: response.usage?.input_tokens,\n      outputTokens: response.usage?.output_tokens,\n    });\n\n    this.logger?.info('Text generation completed', {\n      model: this.modelId,\n      textLength: text.length,\n      stopReason: response.stop_reason\n    });\n\n    return {\n      text,\n      stopReason: response.stop_reason || 'unknown'\n    };\n  }\n\n  async generateStructured<T>(\n    prompt: string,\n    maxTokens: number,\n    temperature: number,\n    elementSchema: ElementSchema,\n  ): Promise<StructuredResponse<T>> {\n    // Capability gate (D3/D4): model choice is deployment config, so the\n    // client asks the provider whether the configured model can honour\n    // strictness — and REFUSES when it cannot. Silent fallback to\n    // unconstrained tool use is exactly the behaviour that turned 202 real\n    // entities into a green empty job. The discovery is the same cached\n    // Models API call `limits()` uses; no extra round trip.\n    const discovery = await this.discover();\n    if (!discovery.structuredOutputsSupported) {\n      throw new Error(\n        `Model '${this.modelId}' does not report support for strict structured outputs ` +\n        `(Models API capabilities.structured_outputs) — refusing rather than degrading to ` +\n        `unconstrained tool use, which silently discards unreadable results. Re-point the ` +\n        `inference.model key that pins this worker/actor in .semiont/semiontconfig/*.toml ` +\n        `(e.g. environments.<env>.workers.<job-type>.inference.model) at a model that ` +\n        `reports supported: true.`,\n      );\n    }\n\n    this.logger?.debug('Generating structured output with inference client', {\n      model: this.modelId,\n      promptLength: prompt.length,\n      maxTokens,\n      temperature,\n    });\n\n    const params: Anthropic.MessageCreateParamsNonStreaming = {\n      model: this.modelId,\n      max_tokens: maxTokens,\n      temperature,\n      messages: [{ role: 'user', content: prompt }],\n      // Response-level structured output with an ARRAY root: the response\n      // text IS the schema-conforming JSON. No tools, no prefill.\n      output_config: {\n        format: {\n          type: 'json_schema',\n          schema: { type: 'array', items: elementSchema },\n        },\n      },\n    };\n\n    const start = performance.now();\n    const response = await this.recordedRequest(params, start);\n\n    const textContent = response.content.find(c => c.type === 'text');\n    if (!textContent || textContent.type !== 'text') {\n      this.recordError(start, response);\n      this.logger?.error('No text content in structured inference response', {\n        model: this.modelId,\n        contentTypes: response.content.map(c => c.type)\n      });\n      throw new Error('No text content in structured inference response');\n    }\n\n    // Anything that does not read as an array is a THROW, never a coerced\n    // `[]` — \"we could not read the model\" must never be conflated with\n    // \"the model found nothing\": that conflation is what silently discarded\n    // 202 real entities as a green empty result. A truncated (`max_tokens`)\n    // response surfaces here too, as unparseable JSON naming its stop_reason.\n    let parsed: unknown;\n    try {\n      parsed = JSON.parse(textContent.text);\n    } catch (err) {\n      this.recordError(start, response);\n      this.logger?.error('Structured response could not be read', {\n        model: this.modelId,\n        textLength: textContent.text.length,\n        stopReason: response.stop_reason,\n      });\n      throw new Error(\n        `Structured response could not be read: response is not valid JSON (stop_reason: ${response.stop_reason})`,\n        { cause: err },\n      );\n    }\n    if (!Array.isArray(parsed)) {\n      this.recordError(start, response);\n      this.logger?.error('Structured response could not be read', {\n        model: this.modelId,\n        parsedType: typeof parsed,\n        stopReason: response.stop_reason,\n      });\n      throw new Error(\n        `Structured response could not be read: parsed to ${typeof parsed}, not an array (stop_reason: ${response.stop_reason})`,\n      );\n    }\n\n    recordInferenceUsage({\n      provider: this.type,\n      model: this.modelId,\n      durationMs: performance.now() - start,\n      outcome: 'success',\n      inputTokens: response.usage?.input_tokens,\n      outputTokens: response.usage?.output_tokens,\n    });\n\n    this.logger?.info('Structured generation completed', {\n      model: this.modelId,\n      items: parsed.length,\n      stopReason: response.stop_reason\n    });\n\n    return {\n      items: parsed as T[],\n      stopReason: response.stop_reason || 'unknown',\n    };\n  }\n\n  /** Issue the request, recording an error metric if the transport throws. */\n  private async recordedRequest(params: Anthropic.MessageCreateParamsNonStreaming, start: number): Promise<Anthropic.Message> {\n    try {\n      return await this.requestMessage(params);\n    } catch (err) {\n      recordInferenceUsage({\n        provider: this.type,\n        model: this.modelId,\n        durationMs: performance.now() - start,\n        outcome: 'error',\n      });\n      throw err;\n    }\n  }\n\n  private recordError(start: number, response: Anthropic.Message): void {\n    recordInferenceUsage({\n      provider: this.type,\n      model: this.modelId,\n      durationMs: performance.now() - start,\n      outcome: 'error',\n      inputTokens: response.usage?.input_tokens,\n      outputTokens: response.usage?.output_tokens,\n    });\n  }\n}\n","// Ollama implementation of InferenceClient interface\n// Uses native Ollama HTTP API (no SDK dependency)\n\nimport { estimateTokens, isNumber, isObject } from '@semiont/core';\nimport type { Logger } from '@semiont/core';\nimport { recordInferenceUsage } from '@semiont/observability';\nimport { ElementSchema, InferenceClient, InferenceLimits, InferenceResponse, StructuredResponse } from '../interface.js';\n\n// Slack added to the chars/4 prompt estimate when sizing `num_ctx`:\n// proportional to the estimate (the heuristic's error grows with prompt size)\n// plus a small fixed allowance for the model's chat template. The risk profile\n// is asymmetric — an undersized window silently clips input (the exact hole\n// managed num_ctx exists to close) while an oversized one only costs memory —\n// so the slack leans generous. Always capped at the model's real window.\nconst NUM_CTX_ESTIMATE_SLACK = 0.2;\nconst NUM_CTX_TEMPLATE_ALLOWANCE = 64;\n\ninterface OllamaGenerateResponse {\n  response: string;\n  done: boolean;\n  done_reason?: string;\n  /** Number of prompt tokens evaluated. Available on most Ollama versions. */\n  prompt_eval_count?: number;\n  /** Number of tokens generated. */\n  eval_count?: number;\n}\n\nexport class OllamaInferenceClient implements InferenceClient {\n  readonly type = 'ollama' as const;\n  readonly modelId: string;\n  private baseURL: string;\n  private logger?: Logger;\n\n  private limitsPromise?: Promise<InferenceLimits>;\n\n  constructor(model: string, baseURL?: string, logger?: Logger) {\n    this.baseURL = (baseURL || 'http://localhost:11434').replace(/\\/+$/, '');\n    this.modelId = model;\n    this.logger = logger;\n  }\n\n  limits(): Promise<InferenceLimits> {\n    if (!this.limitsPromise) {\n      this.limitsPromise = this.discoverLimits().catch((err: unknown) => {\n        // Never cache a failed discovery — a transient outage would otherwise\n        // pin every future call to the same rejection.\n        this.limitsPromise = undefined;\n        throw err;\n      });\n    }\n    return this.limitsPromise;\n  }\n\n  private async discoverLimits(): Promise<InferenceLimits> {\n    const res = await fetch(`${this.baseURL}/api/show`, {\n      method: 'POST',\n      headers: { 'Content-Type': 'application/json' },\n      body: JSON.stringify({ model: this.modelId }),\n    });\n    if (!res.ok) {\n      throw new Error(\n        `Failed to discover model limits: /api/show returned ${res.status} for '${this.modelId}'`,\n      );\n    }\n    const data: unknown = await res.json();\n    const modelInfo = isObject(data) && isObject(data['model_info']) ? data['model_info'] : undefined;\n    const contextTokens = readContextLength(modelInfo);\n    if (contextTokens === undefined) {\n      throw new Error(`/api/show reports no context length for '${this.modelId}'`);\n    }\n    // Shared window: input and output draw from the same context — there is\n    // no separate output ceiling, so the window is published as both (the\n    // `maxOutputTokens === contextTokens` shape consumers key the split on).\n    return { contextTokens, maxOutputTokens: contextTokens };\n  }\n\n  async generateText(prompt: string, maxTokens: number, temperature: number): Promise<string> {\n    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature);\n    return response.text;\n  }\n\n  async generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number): Promise<InferenceResponse> {\n    return this.generate(prompt, maxTokens, temperature, undefined);\n  }\n\n  async generateStructured<T>(\n    prompt: string,\n    maxTokens: number,\n    temperature: number,\n    elementSchema: ElementSchema,\n  ): Promise<StructuredResponse<T>> {\n    // Grammar-constrained sampling: the schema goes to Ollama's `format`\n    // parameter, which constrains generation itself — same mechanism as the\n    // old bare array schema, now element-typed. The response text is then\n    // parsed here, and anything that does not read as an array is a THROW,\n    // never a coerced [] — \"could not read the model\" must stay distinct\n    // from \"the model found nothing.\"\n    const response = await this.generate(prompt, maxTokens, temperature, elementSchema);\n\n    let parsed: unknown;\n    try {\n      parsed = JSON.parse(response.text);\n    } catch (err) {\n      this.logger?.error('Structured response could not be read', {\n        model: this.modelId,\n        textLength: response.text.length,\n        stopReason: response.stopReason,\n      });\n      throw new Error(\n        `Structured response could not be read: response is not valid JSON (stop_reason: ${response.stopReason})`,\n        { cause: err },\n      );\n    }\n    if (!Array.isArray(parsed)) {\n      this.logger?.error('Structured response could not be read', {\n        model: this.modelId,\n        parsedType: typeof parsed,\n        stopReason: response.stopReason,\n      });\n      throw new Error(\n        `Structured response could not be read: parsed to ${typeof parsed}, not an array (stop_reason: ${response.stopReason})`,\n      );\n    }\n\n    return { items: parsed as T[], stopReason: response.stopReason };\n  }\n\n  private async generate(\n    prompt: string,\n    maxTokens: number,\n    temperature: number,\n    elementSchema: ElementSchema | undefined,\n  ): Promise<InferenceResponse> {\n    this.logger?.debug('Generating text with Ollama', {\n      model: this.modelId,\n      promptLength: prompt.length,\n      maxTokens,\n      temperature,\n      structured: elementSchema !== undefined,\n    });\n\n    // Managed context window: size num_ctx to cover this request, capped at\n    // the model's discovered window. Without an explicit num_ctx Ollama uses\n    // the model's *default* window and SILENTLY CLIPS any prompt beyond it —\n    // input loss with no error (found 2026-07-30).\n    const limits = await this.limits();\n    const promptTokens = estimateTokens(prompt);\n    if (promptTokens + maxTokens > limits.contextTokens) {\n      throw new Error(\n        `Prompt (~${promptTokens} tokens) + output budget (${maxTokens}) exceed the ` +\n        `'${this.modelId}' context window (${limits.contextTokens} tokens)`,\n      );\n    }\n    const numCtx = Math.min(\n      limits.contextTokens,\n      promptTokens + maxTokens\n        + Math.ceil(promptTokens * NUM_CTX_ESTIMATE_SLACK) + NUM_CTX_TEMPLATE_ALLOWANCE,\n    );\n\n    const url = `${this.baseURL}/api/generate`;\n    const start = performance.now();\n\n    // Ollama's `format` parameter accepts either the literal string\n    // `\"json\"` (any valid JSON, including objects, numbers, etc.) or a\n    // JSON schema (constrains the top-level shape). The structured contract\n    // is \"an array of elements matching the caller's schema,\" so we pass an\n    // array schema wrapping it — the bare `\"json\"` string would let the\n    // model satisfy \"valid JSON\" with `{\"entities\": [...]}` and break every\n    // consumer that maps over the top-level value.\n    const body: Record<string, unknown> = {\n      model: this.modelId,\n      prompt,\n      stream: false,\n      think: false,\n      options: {\n        num_predict: maxTokens,\n        num_ctx: numCtx,\n        temperature,\n      },\n    };\n    if (elementSchema !== undefined) {\n      body['format'] = { type: 'array', items: elementSchema };\n    }\n\n    let res: Response;\n    try {\n      res = await fetch(url, {\n        method: 'POST',\n        headers: { 'Content-Type': 'application/json' },\n        body: JSON.stringify(body),\n      });\n    } catch (err) {\n      recordInferenceUsage({\n        provider: this.type,\n        model: this.modelId,\n        durationMs: performance.now() - start,\n        outcome: 'error',\n      });\n      throw err;\n    }\n\n    if (!res.ok) {\n      recordInferenceUsage({\n        provider: this.type,\n        model: this.modelId,\n        durationMs: performance.now() - start,\n        outcome: 'error',\n      });\n      const body = await res.text();\n      this.logger?.error('Ollama API error', {\n        model: this.modelId,\n        status: res.status,\n        body,\n      });\n      throw new Error(`Ollama API error (${res.status}): ${body}`);\n    }\n\n    const data = await res.json() as OllamaGenerateResponse;\n\n    if (!data.response) {\n      recordInferenceUsage({\n        provider: this.type,\n        model: this.modelId,\n        durationMs: performance.now() - start,\n        outcome: 'error',\n        inputTokens: data.prompt_eval_count,\n        outputTokens: data.eval_count,\n      });\n      this.logger?.error('Empty response from Ollama', { model: this.modelId });\n      throw new Error('Empty response from Ollama');\n    }\n\n    recordInferenceUsage({\n      provider: this.type,\n      model: this.modelId,\n      durationMs: performance.now() - start,\n      outcome: 'success',\n      inputTokens: data.prompt_eval_count,\n      outputTokens: data.eval_count,\n    });\n\n    const stopReason = mapStopReason(data.done_reason);\n\n    this.logger?.info('Text generation completed', {\n      model: this.modelId,\n      textLength: data.response.length,\n      stopReason,\n    });\n\n    return {\n      text: data.response,\n      stopReason,\n    };\n  }\n}\n\n/**\n * The context length lives in `model_info` under an architecture-prefixed key\n * (e.g. `llama.context_length`); `general.architecture` names the prefix.\n * Falls back to any `*.context_length` key for models whose metadata omits\n * the architecture field.\n */\nfunction readContextLength(modelInfo: Record<string, unknown> | undefined): number | undefined {\n  if (!modelInfo) return undefined;\n  const arch = modelInfo['general.architecture'];\n  if (typeof arch === 'string') {\n    const direct = modelInfo[`${arch}.context_length`];\n    if (isNumber(direct) && direct > 0) return direct;\n  }\n  const fallbackKey = Object.keys(modelInfo).find(k => k.endsWith('.context_length'));\n  if (fallbackKey !== undefined) {\n    const fallback = modelInfo[fallbackKey];\n    if (isNumber(fallback) && fallback > 0) return fallback;\n  }\n  return undefined;\n}\n\nfunction mapStopReason(doneReason: string | undefined): string {\n  switch (doneReason) {\n    case 'stop': return 'end_turn';\n    case 'length': return 'max_tokens';\n    default: return doneReason || 'unknown';\n  }\n}\n","// Factory for creating inference client instances based on configuration\n\nimport type { Logger } from '@semiont/core';\nimport { InferenceClient } from './interface.js';\nimport { AnthropicInferenceClient } from './implementations/anthropic.js';\nimport { OllamaInferenceClient } from './implementations/ollama.js';\n\nexport type InferenceClientType = 'anthropic' | 'ollama';\n\nexport interface InferenceClientConfig {\n  type: InferenceClientType;\n  apiKey?: string;\n  model: string;\n  endpoint?: string;\n  baseURL?: string;\n}\n\nexport function createInferenceClient(config: InferenceClientConfig, logger?: Logger): InferenceClient {\n  switch (config.type) {\n    case 'anthropic': {\n      if (!config.apiKey || config.apiKey.trim() === '') {\n        throw new Error('apiKey is required for Anthropic inference client');\n      }\n      return new AnthropicInferenceClient(\n        config.apiKey,\n        config.model,\n        config.endpoint || config.baseURL,\n        logger\n      );\n    }\n\n    case 'ollama': {\n      return new OllamaInferenceClient(\n        config.model,\n        config.endpoint || config.baseURL,\n        logger\n      );\n    }\n\n    default:\n      throw new Error(`Unsupported inference client type: ${config.type}`);\n  }\n}\n","// Mock implementation of InferenceClient for testing\n\nimport { ElementSchema, InferenceClient, InferenceLimits, InferenceResponse, StructuredResponse } from '../interface.js';\n\n// Generous defaults so existing consumers never trip chunking or window\n// guards unless a test injects tighter limits deliberately.\nconst GENEROUS_LIMITS: InferenceLimits = {\n  contextTokens: 1_000_000,\n  maxOutputTokens: 1_000_000,\n};\n\nexport class MockInferenceClient implements InferenceClient {\n  readonly type = 'mock' as const;\n  readonly modelId = 'mock-model' as const;\n  private responses: string[] = [];\n  private responseIndex: number = 0;\n  private stopReasons: string[] = [];\n  private injectedLimits: InferenceLimits;\n  public calls: Array<{ prompt: string; maxTokens: number; temperature: number; elementSchema?: ElementSchema }> = [];\n\n  constructor(responses: string[] = ['Mock response'], stopReasons?: string[], limits?: InferenceLimits) {\n    this.responses = responses;\n    this.stopReasons = stopReasons || responses.map(() => 'end_turn');\n    this.injectedLimits = limits ?? GENEROUS_LIMITS;\n  }\n\n  async limits(): Promise<InferenceLimits> {\n    return this.injectedLimits;\n  }\n\n  async generateText(prompt: string, maxTokens: number, temperature: number): Promise<string> {\n    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature);\n    return response.text;\n  }\n\n  async generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number): Promise<InferenceResponse> {\n    this.calls.push({ prompt, maxTokens, temperature });\n    return this.nextResponse();\n  }\n\n  /**\n   * Structured surface: pops the same responses queue and PARSES the entry,\n   * mirroring the real contract — a queued string that is not a JSON array\n   * throws \"could not be read\", so tests inject the malformed shape simply by\n   * queuing it (`setResponses(['not json'])`). The element schema is recorded\n   * on `calls` so tests can assert what the caller declared.\n   */\n  async generateStructured<T>(\n    prompt: string,\n    maxTokens: number,\n    temperature: number,\n    elementSchema: ElementSchema,\n  ): Promise<StructuredResponse<T>> {\n    this.calls.push({ prompt, maxTokens, temperature, elementSchema });\n    const { text, stopReason } = this.nextResponse();\n\n    let parsed: unknown;\n    try {\n      parsed = JSON.parse(text);\n    } catch (err) {\n      throw new Error(\n        `Structured response could not be read: response is not valid JSON (stop_reason: ${stopReason})`,\n        { cause: err },\n      );\n    }\n    if (!Array.isArray(parsed)) {\n      throw new Error(\n        `Structured response could not be read: parsed to ${typeof parsed}, not an array (stop_reason: ${stopReason})`,\n      );\n    }\n    return { items: parsed as T[], stopReason };\n  }\n\n  private nextResponse(): InferenceResponse {\n    const text = this.responses[this.responseIndex];\n    const stopReason = this.stopReasons[this.responseIndex] || 'end_turn';\n\n    if (this.responseIndex < this.responses.length - 1) {\n      this.responseIndex++;\n    }\n\n    return { text, stopReason };\n  }\n\n  // Test helper methods\n  reset(): void {\n    this.calls = [];\n    this.responseIndex = 0;\n  }\n\n  setResponses(responses: string[], stopReasons?: string[]): void {\n    this.responses = responses;\n    this.stopReasons = stopReasons || responses.map(() => 'end_turn');\n    this.responseIndex = 0;\n  }\n}\n"],"mappings":";AAEA,OAAO,eAAe;AACtB,SAAS,gBAA6B;AACtC,SAAS,4BAA4B;AASrC,IAAM,iCAAiC,KAAK,MAAM,QAAU,CAAC;AAyBtD,IAAM,2BAAN,MAA0D;AAAA,EACtD,OAAO;AAAA,EACP;AAAA,EACD;AAAA,EACA;AAAA,EACA;AAAA,EAER,YAAY,QAAgB,OAAe,SAAkB,QAAiB;AAC5E,SAAK,SAAS,IAAI,UAAU;AAAA,MAC1B;AAAA,MACA,SAAS,WAAW;AAAA,IACtB,CAAC;AACD,SAAK,UAAU;AACf,SAAK,SAAS;AAAA,EAChB;AAAA,EAEA,SAAmC;AACjC,WAAO,KAAK,SAAS,EAAE,KAAK,OAAK,EAAE,MAAM;AAAA,EAC3C;AAAA,EAEQ,WAAoC;AAC1C,QAAI,CAAC,KAAK,kBAAkB;AAC1B,WAAK,mBAAmB,KAAK,cAAc,EAAE,MAAM,CAAC,QAAiB;AAGnE,aAAK,mBAAmB;AACxB,cAAM;AAAA,MACR,CAAC;AAAA,IACH;AACA,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,gBAAyC;AAMrD,UAAM,OAAO,MAAM,KAAK,OAAO,OAAO,SAAS,KAAK,OAAO,EAAE,MAAM,CAAC,QAAiB;AACnF,YAAM,IAAI;AAAA,QACR,wCAAwC,KAAK,OAAO;AAAA,QACpD,EAAE,OAAO,IAAI;AAAA,MACf;AAAA,IACF,CAAC;AACD,QAAI,KAAK,oBAAoB,QAAQ,KAAK,cAAc,MAAM;AAC5D,YAAM,IAAI,MAAM,sDAAsD,KAAK,OAAO,GAAG;AAAA,IACvF;AAIA,UAAM,MAAe;AACrB,UAAM,6BACJ,SAAS,GAAG,KACZ,SAAS,IAAI,cAAc,CAAC,KAC5B,SAAS,IAAI,cAAc,EAAE,oBAAoB,CAAC,KAClD,IAAI,cAAc,EAAE,oBAAoB,EAAE,WAAW,MAAM;AAC7D,WAAO;AAAA,MACL,QAAQ,EAAE,eAAe,KAAK,kBAAkB,iBAAiB,KAAK,WAAW;AAAA,MACjF;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,eAAe,QAA+E;AACpG,QAAI,OAAO,aAAa,gCAAgC;AACtD,aAAO,KAAK,OAAO,SAAS,OAAO,MAAM,EAAE,aAAa;AAAA,IAC1D;AACA,WAAO,KAAK,OAAO,SAAS,OAAO,MAAM;AAAA,EAC3C;AAAA,EAEA,MAAM,aAAa,QAAgB,WAAmB,aAAsC;AAC1F,UAAM,WAAW,MAAM,KAAK,yBAAyB,QAAQ,WAAW,WAAW;AACnF,WAAO,SAAS;AAAA,EAClB;AAAA,EAEA,MAAM,yBAAyB,QAAgB,WAAmB,aAAiD;AACjH,SAAK,QAAQ,MAAM,yCAAyC;AAAA,MAC1D,OAAO,KAAK;AAAA,MACZ,cAAc,OAAO;AAAA,MACrB;AAAA,MACA;AAAA,IACF,CAAC;AAED,UAAM,SAAoD;AAAA,MACxD,OAAO,KAAK;AAAA,MACZ,YAAY;AAAA,MACZ;AAAA,MACA,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,OAAO,CAAC;AAAA,IAC9C;AAEA,UAAM,QAAQ,YAAY,IAAI;AAC9B,UAAM,WAAW,MAAM,KAAK,gBAAgB,QAAQ,KAAK;AAEzD,UAAM,cAAc,SAAS,QAAQ,KAAK,OAAK,EAAE,SAAS,MAAM;AAChE,QAAI,CAAC,eAAe,YAAY,SAAS,QAAQ;AAC/C,WAAK,YAAY,OAAO,QAAQ;AAChC,WAAK,QAAQ,MAAM,yCAAyC;AAAA,QAC1D,OAAO,KAAK;AAAA,QACZ,cAAc,SAAS,QAAQ,IAAI,OAAK,EAAE,IAAI;AAAA,MAChD,CAAC;AACD,YAAM,IAAI,MAAM,uCAAuC;AAAA,IACzD;AACA,UAAM,OAAO,YAAY;AAEzB,yBAAqB;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,OAAO,KAAK;AAAA,MACZ,YAAY,YAAY,IAAI,IAAI;AAAA,MAChC,SAAS;AAAA,MACT,aAAa,SAAS,OAAO;AAAA,MAC7B,cAAc,SAAS,OAAO;AAAA,IAChC,CAAC;AAED,SAAK,QAAQ,KAAK,6BAA6B;AAAA,MAC7C,OAAO,KAAK;AAAA,MACZ,YAAY,KAAK;AAAA,MACjB,YAAY,SAAS;AAAA,IACvB,CAAC;AAED,WAAO;AAAA,MACL;AAAA,MACA,YAAY,SAAS,eAAe;AAAA,IACtC;AAAA,EACF;AAAA,EAEA,MAAM,mBACJ,QACA,WACA,aACA,eACgC;AAOhC,UAAM,YAAY,MAAM,KAAK,SAAS;AACtC,QAAI,CAAC,UAAU,4BAA4B;AACzC,YAAM,IAAI;AAAA,QACR,UAAU,KAAK,OAAO;AAAA,MAMxB;AAAA,IACF;AAEA,SAAK,QAAQ,MAAM,sDAAsD;AAAA,MACvE,OAAO,KAAK;AAAA,MACZ,cAAc,OAAO;AAAA,MACrB;AAAA,MACA;AAAA,IACF,CAAC;AAED,UAAM,SAAoD;AAAA,MACxD,OAAO,KAAK;AAAA,MACZ,YAAY;AAAA,MACZ;AAAA,MACA,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,OAAO,CAAC;AAAA;AAAA;AAAA,MAG5C,eAAe;AAAA,QACb,QAAQ;AAAA,UACN,MAAM;AAAA,UACN,QAAQ,EAAE,MAAM,SAAS,OAAO,cAAc;AAAA,QAChD;AAAA,MACF;AAAA,IACF;AAEA,UAAM,QAAQ,YAAY,IAAI;AAC9B,UAAM,WAAW,MAAM,KAAK,gBAAgB,QAAQ,KAAK;AAEzD,UAAM,cAAc,SAAS,QAAQ,KAAK,OAAK,EAAE,SAAS,MAAM;AAChE,QAAI,CAAC,eAAe,YAAY,SAAS,QAAQ;AAC/C,WAAK,YAAY,OAAO,QAAQ;AAChC,WAAK,QAAQ,MAAM,oDAAoD;AAAA,QACrE,OAAO,KAAK;AAAA,QACZ,cAAc,SAAS,QAAQ,IAAI,OAAK,EAAE,IAAI;AAAA,MAChD,CAAC;AACD,YAAM,IAAI,MAAM,kDAAkD;AAAA,IACpE;AAOA,QAAI;AACJ,QAAI;AACF,eAAS,KAAK,MAAM,YAAY,IAAI;AAAA,IACtC,SAAS,KAAK;AACZ,WAAK,YAAY,OAAO,QAAQ;AAChC,WAAK,QAAQ,MAAM,yCAAyC;AAAA,QAC1D,OAAO,KAAK;AAAA,QACZ,YAAY,YAAY,KAAK;AAAA,QAC7B,YAAY,SAAS;AAAA,MACvB,CAAC;AACD,YAAM,IAAI;AAAA,QACR,mFAAmF,SAAS,WAAW;AAAA,QACvG,EAAE,OAAO,IAAI;AAAA,MACf;AAAA,IACF;AACA,QAAI,CAAC,MAAM,QAAQ,MAAM,GAAG;AAC1B,WAAK,YAAY,OAAO,QAAQ;AAChC,WAAK,QAAQ,MAAM,yCAAyC;AAAA,QAC1D,OAAO,KAAK;AAAA,QACZ,YAAY,OAAO;AAAA,QACnB,YAAY,SAAS;AAAA,MACvB,CAAC;AACD,YAAM,IAAI;AAAA,QACR,oDAAoD,OAAO,MAAM,gCAAgC,SAAS,WAAW;AAAA,MACvH;AAAA,IACF;AAEA,yBAAqB;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,OAAO,KAAK;AAAA,MACZ,YAAY,YAAY,IAAI,IAAI;AAAA,MAChC,SAAS;AAAA,MACT,aAAa,SAAS,OAAO;AAAA,MAC7B,cAAc,SAAS,OAAO;AAAA,IAChC,CAAC;AAED,SAAK,QAAQ,KAAK,mCAAmC;AAAA,MACnD,OAAO,KAAK;AAAA,MACZ,OAAO,OAAO;AAAA,MACd,YAAY,SAAS;AAAA,IACvB,CAAC;AAED,WAAO;AAAA,MACL,OAAO;AAAA,MACP,YAAY,SAAS,eAAe;AAAA,IACtC;AAAA,EACF;AAAA;AAAA,EAGA,MAAc,gBAAgB,QAAmD,OAA2C;AAC1H,QAAI;AACF,aAAO,MAAM,KAAK,eAAe,MAAM;AAAA,IACzC,SAAS,KAAK;AACZ,2BAAqB;AAAA,QACnB,UAAU,KAAK;AAAA,QACf,OAAO,KAAK;AAAA,QACZ,YAAY,YAAY,IAAI,IAAI;AAAA,QAChC,SAAS;AAAA,MACX,CAAC;AACD,YAAM;AAAA,IACR;AAAA,EACF;AAAA,EAEQ,YAAY,OAAe,UAAmC;AACpE,yBAAqB;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,OAAO,KAAK;AAAA,MACZ,YAAY,YAAY,IAAI,IAAI;AAAA,MAChC,SAAS;AAAA,MACT,aAAa,SAAS,OAAO;AAAA,MAC7B,cAAc,SAAS,OAAO;AAAA,IAChC,CAAC;AAAA,EACH;AACF;;;ACxSA,SAAS,gBAAgB,UAAU,YAAAA,iBAAgB;AAEnD,SAAS,wBAAAC,6BAA4B;AASrC,IAAM,yBAAyB;AAC/B,IAAM,6BAA6B;AAY5B,IAAM,wBAAN,MAAuD;AAAA,EACnD,OAAO;AAAA,EACP;AAAA,EACD;AAAA,EACA;AAAA,EAEA;AAAA,EAER,YAAY,OAAe,SAAkB,QAAiB;AAC5D,SAAK,WAAW,WAAW,0BAA0B,QAAQ,QAAQ,EAAE;AACvE,SAAK,UAAU;AACf,SAAK,SAAS;AAAA,EAChB;AAAA,EAEA,SAAmC;AACjC,QAAI,CAAC,KAAK,eAAe;AACvB,WAAK,gBAAgB,KAAK,eAAe,EAAE,MAAM,CAAC,QAAiB;AAGjE,aAAK,gBAAgB;AACrB,cAAM;AAAA,MACR,CAAC;AAAA,IACH;AACA,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,iBAA2C;AACvD,UAAM,MAAM,MAAM,MAAM,GAAG,KAAK,OAAO,aAAa;AAAA,MAClD,QAAQ;AAAA,MACR,SAAS,EAAE,gBAAgB,mBAAmB;AAAA,MAC9C,MAAM,KAAK,UAAU,EAAE,OAAO,KAAK,QAAQ,CAAC;AAAA,IAC9C,CAAC;AACD,QAAI,CAAC,IAAI,IAAI;AACX,YAAM,IAAI;AAAA,QACR,uDAAuD,IAAI,MAAM,SAAS,KAAK,OAAO;AAAA,MACxF;AAAA,IACF;AACA,UAAM,OAAgB,MAAM,IAAI,KAAK;AACrC,UAAM,YAAYD,UAAS,IAAI,KAAKA,UAAS,KAAK,YAAY,CAAC,IAAI,KAAK,YAAY,IAAI;AACxF,UAAM,gBAAgB,kBAAkB,SAAS;AACjD,QAAI,kBAAkB,QAAW;AAC/B,YAAM,IAAI,MAAM,4CAA4C,KAAK,OAAO,GAAG;AAAA,IAC7E;AAIA,WAAO,EAAE,eAAe,iBAAiB,cAAc;AAAA,EACzD;AAAA,EAEA,MAAM,aAAa,QAAgB,WAAmB,aAAsC;AAC1F,UAAM,WAAW,MAAM,KAAK,yBAAyB,QAAQ,WAAW,WAAW;AACnF,WAAO,SAAS;AAAA,EAClB;AAAA,EAEA,MAAM,yBAAyB,QAAgB,WAAmB,aAAiD;AACjH,WAAO,KAAK,SAAS,QAAQ,WAAW,aAAa,MAAS;AAAA,EAChE;AAAA,EAEA,MAAM,mBACJ,QACA,WACA,aACA,eACgC;AAOhC,UAAM,WAAW,MAAM,KAAK,SAAS,QAAQ,WAAW,aAAa,aAAa;AAElF,QAAI;AACJ,QAAI;AACF,eAAS,KAAK,MAAM,SAAS,IAAI;AAAA,IACnC,SAAS,KAAK;AACZ,WAAK,QAAQ,MAAM,yCAAyC;AAAA,QAC1D,OAAO,KAAK;AAAA,QACZ,YAAY,SAAS,KAAK;AAAA,QAC1B,YAAY,SAAS;AAAA,MACvB,CAAC;AACD,YAAM,IAAI;AAAA,QACR,mFAAmF,SAAS,UAAU;AAAA,QACtG,EAAE,OAAO,IAAI;AAAA,MACf;AAAA,IACF;AACA,QAAI,CAAC,MAAM,QAAQ,MAAM,GAAG;AAC1B,WAAK,QAAQ,MAAM,yCAAyC;AAAA,QAC1D,OAAO,KAAK;AAAA,QACZ,YAAY,OAAO;AAAA,QACnB,YAAY,SAAS;AAAA,MACvB,CAAC;AACD,YAAM,IAAI;AAAA,QACR,oDAAoD,OAAO,MAAM,gCAAgC,SAAS,UAAU;AAAA,MACtH;AAAA,IACF;AAEA,WAAO,EAAE,OAAO,QAAe,YAAY,SAAS,WAAW;AAAA,EACjE;AAAA,EAEA,MAAc,SACZ,QACA,WACA,aACA,eAC4B;AAC5B,SAAK,QAAQ,MAAM,+BAA+B;AAAA,MAChD,OAAO,KAAK;AAAA,MACZ,cAAc,OAAO;AAAA,MACrB;AAAA,MACA;AAAA,MACA,YAAY,kBAAkB;AAAA,IAChC,CAAC;AAMD,UAAM,SAAS,MAAM,KAAK,OAAO;AACjC,UAAM,eAAe,eAAe,MAAM;AAC1C,QAAI,eAAe,YAAY,OAAO,eAAe;AACnD,YAAM,IAAI;AAAA,QACR,YAAY,YAAY,6BAA6B,SAAS,iBAC1D,KAAK,OAAO,qBAAqB,OAAO,aAAa;AAAA,MAC3D;AAAA,IACF;AACA,UAAM,SAAS,KAAK;AAAA,MAClB,OAAO;AAAA,MACP,eAAe,YACX,KAAK,KAAK,eAAe,sBAAsB,IAAI;AAAA,IACzD;AAEA,UAAM,MAAM,GAAG,KAAK,OAAO;AAC3B,UAAM,QAAQ,YAAY,IAAI;AAS9B,UAAM,OAAgC;AAAA,MACpC,OAAO,KAAK;AAAA,MACZ;AAAA,MACA,QAAQ;AAAA,MACR,OAAO;AAAA,MACP,SAAS;AAAA,QACP,aAAa;AAAA,QACb,SAAS;AAAA,QACT;AAAA,MACF;AAAA,IACF;AACA,QAAI,kBAAkB,QAAW;AAC/B,WAAK,QAAQ,IAAI,EAAE,MAAM,SAAS,OAAO,cAAc;AAAA,IACzD;AAEA,QAAI;AACJ,QAAI;AACF,YAAM,MAAM,MAAM,KAAK;AAAA,QACrB,QAAQ;AAAA,QACR,SAAS,EAAE,gBAAgB,mBAAmB;AAAA,QAC9C,MAAM,KAAK,UAAU,IAAI;AAAA,MAC3B,CAAC;AAAA,IACH,SAAS,KAAK;AACZ,MAAAC,sBAAqB;AAAA,QACnB,UAAU,KAAK;AAAA,QACf,OAAO,KAAK;AAAA,QACZ,YAAY,YAAY,IAAI,IAAI;AAAA,QAChC,SAAS;AAAA,MACX,CAAC;AACD,YAAM;AAAA,IACR;AAEA,QAAI,CAAC,IAAI,IAAI;AACX,MAAAA,sBAAqB;AAAA,QACnB,UAAU,KAAK;AAAA,QACf,OAAO,KAAK;AAAA,QACZ,YAAY,YAAY,IAAI,IAAI;AAAA,QAChC,SAAS;AAAA,MACX,CAAC;AACD,YAAMC,QAAO,MAAM,IAAI,KAAK;AAC5B,WAAK,QAAQ,MAAM,oBAAoB;AAAA,QACrC,OAAO,KAAK;AAAA,QACZ,QAAQ,IAAI;AAAA,QACZ,MAAAA;AAAA,MACF,CAAC;AACD,YAAM,IAAI,MAAM,qBAAqB,IAAI,MAAM,MAAMA,KAAI,EAAE;AAAA,IAC7D;AAEA,UAAM,OAAO,MAAM,IAAI,KAAK;AAE5B,QAAI,CAAC,KAAK,UAAU;AAClB,MAAAD,sBAAqB;AAAA,QACnB,UAAU,KAAK;AAAA,QACf,OAAO,KAAK;AAAA,QACZ,YAAY,YAAY,IAAI,IAAI;AAAA,QAChC,SAAS;AAAA,QACT,aAAa,KAAK;AAAA,QAClB,cAAc,KAAK;AAAA,MACrB,CAAC;AACD,WAAK,QAAQ,MAAM,8BAA8B,EAAE,OAAO,KAAK,QAAQ,CAAC;AACxE,YAAM,IAAI,MAAM,4BAA4B;AAAA,IAC9C;AAEA,IAAAA,sBAAqB;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,OAAO,KAAK;AAAA,MACZ,YAAY,YAAY,IAAI,IAAI;AAAA,MAChC,SAAS;AAAA,MACT,aAAa,KAAK;AAAA,MAClB,cAAc,KAAK;AAAA,IACrB,CAAC;AAED,UAAM,aAAa,cAAc,KAAK,WAAW;AAEjD,SAAK,QAAQ,KAAK,6BAA6B;AAAA,MAC7C,OAAO,KAAK;AAAA,MACZ,YAAY,KAAK,SAAS;AAAA,MAC1B;AAAA,IACF,CAAC;AAED,WAAO;AAAA,MACL,MAAM,KAAK;AAAA,MACX;AAAA,IACF;AAAA,EACF;AACF;AAQA,SAAS,kBAAkB,WAAoE;AAC7F,MAAI,CAAC,UAAW,QAAO;AACvB,QAAM,OAAO,UAAU,sBAAsB;AAC7C,MAAI,OAAO,SAAS,UAAU;AAC5B,UAAM,SAAS,UAAU,GAAG,IAAI,iBAAiB;AACjD,QAAI,SAAS,MAAM,KAAK,SAAS,EAAG,QAAO;AAAA,EAC7C;AACA,QAAM,cAAc,OAAO,KAAK,SAAS,EAAE,KAAK,OAAK,EAAE,SAAS,iBAAiB,CAAC;AAClF,MAAI,gBAAgB,QAAW;AAC7B,UAAM,WAAW,UAAU,WAAW;AACtC,QAAI,SAAS,QAAQ,KAAK,WAAW,EAAG,QAAO;AAAA,EACjD;AACA,SAAO;AACT;AAEA,SAAS,cAAc,YAAwC;AAC7D,UAAQ,YAAY;AAAA,IAClB,KAAK;AAAQ,aAAO;AAAA,IACpB,KAAK;AAAU,aAAO;AAAA,IACtB;AAAS,aAAO,cAAc;AAAA,EAChC;AACF;;;AC1QO,SAAS,sBAAsB,QAA+B,QAAkC;AACrG,UAAQ,OAAO,MAAM;AAAA,IACnB,KAAK,aAAa;AAChB,UAAI,CAAC,OAAO,UAAU,OAAO,OAAO,KAAK,MAAM,IAAI;AACjD,cAAM,IAAI,MAAM,mDAAmD;AAAA,MACrE;AACA,aAAO,IAAI;AAAA,QACT,OAAO;AAAA,QACP,OAAO;AAAA,QACP,OAAO,YAAY,OAAO;AAAA,QAC1B;AAAA,MACF;AAAA,IACF;AAAA,IAEA,KAAK,UAAU;AACb,aAAO,IAAI;AAAA,QACT,OAAO;AAAA,QACP,OAAO,YAAY,OAAO;AAAA,QAC1B;AAAA,MACF;AAAA,IACF;AAAA,IAEA;AACE,YAAM,IAAI,MAAM,sCAAsC,OAAO,IAAI,EAAE;AAAA,EACvE;AACF;;;ACpCA,IAAM,kBAAmC;AAAA,EACvC,eAAe;AAAA,EACf,iBAAiB;AACnB;AAEO,IAAM,sBAAN,MAAqD;AAAA,EACjD,OAAO;AAAA,EACP,UAAU;AAAA,EACX,YAAsB,CAAC;AAAA,EACvB,gBAAwB;AAAA,EACxB,cAAwB,CAAC;AAAA,EACzB;AAAA,EACD,QAA0G,CAAC;AAAA,EAElH,YAAY,YAAsB,CAAC,eAAe,GAAG,aAAwB,QAA0B;AACrG,SAAK,YAAY;AACjB,SAAK,cAAc,eAAe,UAAU,IAAI,MAAM,UAAU;AAChE,SAAK,iBAAiB,UAAU;AAAA,EAClC;AAAA,EAEA,MAAM,SAAmC;AACvC,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,aAAa,QAAgB,WAAmB,aAAsC;AAC1F,UAAM,WAAW,MAAM,KAAK,yBAAyB,QAAQ,WAAW,WAAW;AACnF,WAAO,SAAS;AAAA,EAClB;AAAA,EAEA,MAAM,yBAAyB,QAAgB,WAAmB,aAAiD;AACjH,SAAK,MAAM,KAAK,EAAE,QAAQ,WAAW,YAAY,CAAC;AAClD,WAAO,KAAK,aAAa;AAAA,EAC3B;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,mBACJ,QACA,WACA,aACA,eACgC;AAChC,SAAK,MAAM,KAAK,EAAE,QAAQ,WAAW,aAAa,cAAc,CAAC;AACjE,UAAM,EAAE,MAAM,WAAW,IAAI,KAAK,aAAa;AAE/C,QAAI;AACJ,QAAI;AACF,eAAS,KAAK,MAAM,IAAI;AAAA,IAC1B,SAAS,KAAK;AACZ,YAAM,IAAI;AAAA,QACR,mFAAmF,UAAU;AAAA,QAC7F,EAAE,OAAO,IAAI;AAAA,MACf;AAAA,IACF;AACA,QAAI,CAAC,MAAM,QAAQ,MAAM,GAAG;AAC1B,YAAM,IAAI;AAAA,QACR,oDAAoD,OAAO,MAAM,gCAAgC,UAAU;AAAA,MAC7G;AAAA,IACF;AACA,WAAO,EAAE,OAAO,QAAe,WAAW;AAAA,EAC5C;AAAA,EAEQ,eAAkC;AACxC,UAAM,OAAO,KAAK,UAAU,KAAK,aAAa;AAC9C,UAAM,aAAa,KAAK,YAAY,KAAK,aAAa,KAAK;AAE3D,QAAI,KAAK,gBAAgB,KAAK,UAAU,SAAS,GAAG;AAClD,WAAK;AAAA,IACP;AAEA,WAAO,EAAE,MAAM,WAAW;AAAA,EAC5B;AAAA;AAAA,EAGA,QAAc;AACZ,SAAK,QAAQ,CAAC;AACd,SAAK,gBAAgB;AAAA,EACvB;AAAA,EAEA,aAAa,WAAqB,aAA8B;AAC9D,SAAK,YAAY;AACjB,SAAK,cAAc,eAAe,UAAU,IAAI,MAAM,UAAU;AAChE,SAAK,gBAAgB;AAAA,EACvB;AACF;","names":["isObject","recordInferenceUsage","body"]}
+{"version":3,"sources":["../src/implementations/anthropic.ts","../src/interface.ts","../src/implementations/ollama.ts","../src/factory.ts","../src/implementations/mock.ts"],"sourcesContent":["// Anthropic Claude implementation of InferenceClient interface\n\nimport Anthropic from '@anthropic-ai/sdk';\nimport { isObject, type Logger } from '@semiont/core';\nimport { recordInferenceUsage } from '@semiont/observability';\nimport { ElementSchema, InferenceClient, InferenceLimits, InferenceResponse, StructuredReadError, StructuredResponse, TokenUsage } from '../interface.js';\n\n// The SDK's worst-case output-rate model: client.js's\n// calculateNonstreamingTimeout projects a call's maximum duration as\n// (60min × max_tokens) / 128_000 and refuses non-streaming create() calls\n// projected past its 10-minute timeout. ONE constant, two derivations: the\n// streaming switch below, and `limits().outputTokensPerHour` — the single\n// duration statement this provider surface makes, which detection's\n// duration budget derives from (ABANDONED-INFERENCE P4). Invalidated if\n// the SDK revises its rate model: check calculateNonstreamingTimeout on\n// SDK upgrades.\nconst OUTPUT_TOKENS_PER_HOUR = 128_000;\n\n// Above ~21,333 output tokens (the 10-minute projection) we stream\n// internally and assemble the final message — same request shape, same\n// response handling, same interface.\nconst NONSTREAMING_MAX_OUTPUT_TOKENS = Math.floor(OUTPUT_TOKENS_PER_HOUR / 6);\n\n// Structured generation rides `output_config.format` — response-level\n// structured output: the response TEXT is the schema-conforming JSON, with a\n// top-level ARRAY root (accepted on both live-config models — spike\n// 2026-08-06, `.plans/spikes/output-config-array-root.md`). This replaced\n// the pre-structured-outputs scaffolding: a forced `emit_json_array` tool\n// whose object-only input required an `items` wrapper and an unwrap — and\n// the unwrap was the exact line that silently coerced an unreadable payload\n// to `[]` (STRUCTURED-INFERENCE §Problem). There is no tool-input\n// accumulation step left for the SDK to hand over unparsed; the read path is\n// now the same parse-and-verify shape as Ollama's.\n\n/**\n * Everything one Models API call teaches us about the configured model. The\n * capability stays PRIVATE to this client: its only consumer is the gate in\n * `generateStructured`, so it does not cross the `InferenceClient` interface\n * (no speculative surface — widen `InferenceLimits` only when an external\n * consumer exists).\n */\ninterface ModelDiscovery {\n  limits: InferenceLimits;\n  structuredOutputsSupported: boolean;\n}\n\nexport class AnthropicInferenceClient implements InferenceClient {\n  readonly type = 'anthropic' as const;\n  // Hosted API: a single detection job uses a sliver of the account rate limit\n  // (measured 2026-09-04: ~1 request / 72 s, zero 429s at 4 concurrent types),\n  // so independent calls genuinely parallelize. Conservative until an 8-way run\n  // measures the next step (DETECTION-QUALITY-THROUGHPUT P6).\n  readonly maxConcurrency = 4;\n  // Universal for real providers (user ruling 2026-09-05): \"no observed\n  // collapse\" here was absence-of-looking, and the unexplained ~2× yield gap\n  // vs gemma on the same document is exactly what verification answers. The\n  // ~2× billed input is the accepted cost.\n  readonly verifyDetectionYield = true;\n  readonly modelId: string;\n  private client: Anthropic;\n  private logger?: Logger;\n  private discoveryPromise?: Promise<ModelDiscovery>;\n\n  constructor(apiKey: string, model: string, baseURL?: string, logger?: Logger) {\n    this.client = new Anthropic({\n      apiKey,\n      baseURL: baseURL || 'https://api.anthropic.com',\n    });\n    this.modelId = model;\n    this.logger = logger;\n  }\n\n  limits(): Promise<InferenceLimits> {\n    return this.discover().then(d => d.limits);\n  }\n\n  private discover(): Promise<ModelDiscovery> {\n    if (!this.discoveryPromise) {\n      this.discoveryPromise = this.discoverModel().catch((err: unknown) => {\n        // Never cache a failed discovery — a transient outage would otherwise\n        // pin every future call to the same rejection.\n        this.discoveryPromise = undefined;\n        throw err;\n      });\n    }\n    return this.discoveryPromise;\n  }\n\n  private async discoverModel(): Promise<ModelDiscovery> {\n    // The Models API publishes the actual ceilings AND capabilities per\n    // model — no hand-maintained table to go stale when a new model ships,\n    // and the API's own metadata outranks documentation prose when the two\n    // disagree (measured: the docs' supported-model list was stale while\n    // `capabilities.structured_outputs.supported` was correct).\n    const info = await this.client.models.retrieve(this.modelId).catch((err: unknown) => {\n      throw new Error(\n        `Failed to discover model limits for '${this.modelId}' from the Models API`,\n        { cause: err },\n      );\n    });\n    if (info.max_input_tokens == null || info.max_tokens == null) {\n      throw new Error(`Models API reports no context/output ceilings for '${this.modelId}'`);\n    }\n    // `capabilities` is not declared on the SDK's ModelInfo type — narrow\n    // through the core guards rather than casting. Absent metadata reads as\n    // unsupported: the gate then refuses loudly (D4), never guesses.\n    const raw: unknown = info;\n    const structuredOutputsSupported =\n      isObject(raw) &&\n      isObject(raw['capabilities']) &&\n      isObject(raw['capabilities']['structured_outputs']) &&\n      raw['capabilities']['structured_outputs']['supported'] === true;\n    return {\n      limits: {\n        contextTokens: info.max_input_tokens,\n        maxOutputTokens: info.max_tokens,\n        outputTokensPerHour: OUTPUT_TOKENS_PER_HOUR,\n      },\n      structuredOutputsSupported,\n    };\n  }\n\n  private requestMessage(params: Anthropic.MessageCreateParamsNonStreaming, signal?: AbortSignal): Promise<Anthropic.Message> {\n    // The signal rides the SDK's RequestOptions: an abort tears down the live\n    // attempt AND is checked between the SDK's internal retries, so a\n    // cancelled call cannot survive as a background zombie inside the SDK's\n    // own retry/backoff loop (ABANDONED-INFERENCE P0 caught one completing\n    // 24–34 minutes after abandonment). For visibility into those internal\n    // retries themselves, the SDK's ANTHROPIC_LOG=debug env knob logs every\n    // attempt — nothing to re-implement here.\n    if (params.max_tokens > NONSTREAMING_MAX_OUTPUT_TOKENS) {\n      return this.client.messages.stream(params, { signal }).finalMessage();\n    }\n    return this.client.messages.create(params, { signal });\n  }\n\n  async generateText(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<string> {\n    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature, signal);\n    return response.text;\n  }\n\n  async generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<InferenceResponse> {\n    this.logger?.debug('Generating text with inference client', {\n      model: this.modelId,\n      promptLength: prompt.length,\n      maxTokens,\n      temperature,\n    });\n\n    const params: Anthropic.MessageCreateParamsNonStreaming = {\n      model: this.modelId,\n      max_tokens: maxTokens,\n      temperature,\n      messages: [{ role: 'user', content: prompt }],\n    };\n\n    const start = performance.now();\n    const response = await this.recordedRequest(params, start, signal);\n\n    const textContent = response.content.find(c => c.type === 'text');\n    if (!textContent || textContent.type !== 'text') {\n      this.recordError(start, response);\n      this.logger?.error('No text content in inference response', {\n        model: this.modelId,\n        contentTypes: response.content.map(c => c.type)\n      });\n      throw new Error('No text content in inference response');\n    }\n    const text = textContent.text;\n\n    recordInferenceUsage({\n      provider: this.type,\n      model: this.modelId,\n      durationMs: performance.now() - start,\n      outcome: 'success',\n      inputTokens: response.usage?.input_tokens,\n      outputTokens: response.usage?.output_tokens,\n    });\n\n    this.logger?.info('Text generation completed', {\n      model: this.modelId,\n      textLength: text.length,\n      stopReason: response.stop_reason,\n      requestId: requestIdOf(response),\n    });\n\n    return {\n      text,\n      stopReason: response.stop_reason || 'unknown'\n    };\n  }\n\n  async generateStructured<T>(\n    prompt: string,\n    maxTokens: number,\n    temperature: number,\n    elementSchema: ElementSchema,\n    signal?: AbortSignal,\n  ): Promise<StructuredResponse<T>> {\n    // Capability gate (D3/D4): model choice is deployment config, so the\n    // client asks the provider whether the configured model can honour\n    // strictness — and REFUSES when it cannot. Silent fallback to\n    // unconstrained tool use is exactly the behaviour that turned 202 real\n    // entities into a green empty job. The discovery is the same cached\n    // Models API call `limits()` uses; no extra round trip.\n    const discovery = await this.discover();\n    if (!discovery.structuredOutputsSupported) {\n      throw new Error(\n        `Model '${this.modelId}' does not report support for strict structured outputs ` +\n        `(Models API capabilities.structured_outputs) — refusing rather than degrading to ` +\n        `unconstrained tool use, which silently discards unreadable results. Re-point the ` +\n        `inference.model key that pins this worker/actor in .semiont/semiontconfig/*.toml ` +\n        `(e.g. environments.<env>.workers.<job-type>.inference.model) at a model that ` +\n        `reports supported: true.`,\n      );\n    }\n\n    this.logger?.debug('Generating structured output with inference client', {\n      model: this.modelId,\n      promptLength: prompt.length,\n      maxTokens,\n      temperature,\n    });\n\n    const params: Anthropic.MessageCreateParamsNonStreaming = {\n      model: this.modelId,\n      max_tokens: maxTokens,\n      temperature,\n      messages: [{ role: 'user', content: prompt }],\n      // Response-level structured output with an ARRAY root: the response\n      // text IS the schema-conforming JSON. No tools, no prefill.\n      output_config: {\n        format: {\n          type: 'json_schema',\n          schema: { type: 'array', items: elementSchema },\n        },\n      },\n    };\n\n    const start = performance.now();\n    const response = await this.recordedRequest(params, start, signal);\n\n    const textContent = response.content.find(c => c.type === 'text');\n    if (!textContent || textContent.type !== 'text') {\n      this.recordError(start, response);\n      this.logger?.error('No text content in structured inference response', {\n        model: this.modelId,\n        contentTypes: response.content.map(c => c.type)\n      });\n      throw new Error('No text content in structured inference response');\n    }\n\n    // Anything that does not read as an array is a THROW, never a coerced\n    // `[]` — \"we could not read the model\" must never be conflated with\n    // \"the model found nothing\": that conflation is what silently discarded\n    // 202 real entities as a green empty result. A truncated (`max_tokens`)\n    // response surfaces here too, as unparseable JSON naming its stop_reason.\n    let parsed: unknown;\n    try {\n      parsed = JSON.parse(textContent.text);\n    } catch (err) {\n      this.recordError(start, response);\n      this.logger?.error('Structured response could not be read', {\n        model: this.modelId,\n        textLength: textContent.text.length,\n        stopReason: response.stop_reason,\n      });\n      throw new StructuredReadError('response is not valid JSON', response.stop_reason || 'unknown', { cause: err });\n    }\n    if (!Array.isArray(parsed)) {\n      this.recordError(start, response);\n      this.logger?.error('Structured response could not be read', {\n        model: this.modelId,\n        parsedType: typeof parsed,\n        stopReason: response.stop_reason,\n      });\n      throw new StructuredReadError(`parsed to ${typeof parsed}, not an array`, response.stop_reason || 'unknown');\n    }\n\n    recordInferenceUsage({\n      provider: this.type,\n      model: this.modelId,\n      durationMs: performance.now() - start,\n      outcome: 'success',\n      inputTokens: response.usage?.input_tokens,\n      outputTokens: response.usage?.output_tokens,\n    });\n\n    this.logger?.info('Structured generation completed', {\n      model: this.modelId,\n      items: parsed.length,\n      stopReason: response.stop_reason,\n      requestId: requestIdOf(response),\n    });\n\n    return {\n      items: parsed as T[],\n      stopReason: response.stop_reason || 'unknown',\n      ...usageOf(response),\n    };\n  }\n\n  /** Issue the request, recording an error metric if the transport throws. */\n  private async recordedRequest(params: Anthropic.MessageCreateParamsNonStreaming, start: number, signal?: AbortSignal): Promise<Anthropic.Message> {\n    try {\n      return await this.requestMessage(params, signal);\n    } catch (err) {\n      recordInferenceUsage({\n        provider: this.type,\n        model: this.modelId,\n        durationMs: performance.now() - start,\n        outcome: 'error',\n      });\n      throw err;\n    }\n  }\n\n  private recordError(start: number, response: Anthropic.Message): void {\n    recordInferenceUsage({\n      provider: this.type,\n      model: this.modelId,\n      durationMs: performance.now() - start,\n      outcome: 'error',\n      inputTokens: response.usage?.input_tokens,\n      outputTokens: response.usage?.output_tokens,\n    });\n  }\n}\n\n/**\n * The provider's request id, for correlating our logs with Anthropic's and\n * telling one attempt from another. The SDK attaches `_request_id` to the\n * returned message at runtime but does not declare it on the `Message` type,\n * so it is read through a guard rather than a cast.\n */\nfunction requestIdOf(response: unknown): string | undefined {\n  if (isObject(response) && typeof response['_request_id'] === 'string') {\n    return response['_request_id'];\n  }\n  return undefined;\n}\n\n/**\n * The provider's own token counts, shaped for `TokenUsage`. Absent when the\n * SDK reports none — never zero-filled: a zero would read as \"this call cost\n * nothing\", which is a different claim from \"we do not know\".\n */\nfunction usageOf(response: { usage?: { input_tokens?: number; output_tokens?: number } }): { usage?: TokenUsage } {\n  const { input_tokens, output_tokens } = response.usage ?? {};\n  if (input_tokens === undefined || output_tokens === undefined) return {};\n  return { usage: { inputTokens: input_tokens, outputTokens: output_tokens } };\n}\n","// Inference client interface - all implementations must follow this contract\n\n/**\n * What the call actually cost, as the PROVIDER counted it — never estimated\n * here. Optional because a provider may not report it (and a call that fails\n * before generating has nothing to report); absent means unknown, and a\n * consumer must treat it as unknown rather than substituting a guess.\n */\nexport interface TokenUsage {\n  inputTokens: number;\n  outputTokens: number;\n}\n\nexport interface InferenceResponse {\n  text: string;\n  stopReason: 'end_turn' | 'max_tokens' | 'stop_sequence' | string;\n  usage?: TokenUsage;\n}\n\n/**\n * Raw JSON Schema for ONE array element of a structured generation — a plain\n * object, not a TS type and not a validator instance. Both providers consume\n * JSON Schema directly (Anthropic as the array-root schema under\n * `output_config.format`; Ollama as `format: { type: 'array', items: … }`),\n * so anything richer would abstract one shape with two consumers.\n *\n * Constrain it to what both providers enforce: objects,\n * `string`/`number`/`boolean`/`null`, `enum`, `const`, `required`, and\n * `additionalProperties: false`. Numeric and string constraints (`minimum`,\n * `maxLength`) are NOT enforced by Anthropic strict mode — declaring them\n * buys nothing and misleads the reader.\n */\nexport type ElementSchema = Record<string, unknown>;\n\n/**\n * A structured generation's result: the elements the model produced, plus the\n * provider's stop reason (consumers gate on 'max_tokens' — truncation is data\n * loss, not \"fewer items\").\n *\n * `items` is `T[]`, never a string: there is no representable value meaning\n * \"here is some text I could not read.\" An implementation that cannot deliver\n * the array THROWS — failure is distinct from empty by construction.\n *\n * `T` is a caller assertion, not a runtime guarantee: nothing verifies the\n * element schema and `T` agree, and the type parameter is erased. Declare the\n * schema and `T` adjacently at the call site so drift is visible in one\n * place, and keep per-element structural guards on the consuming side.\n */\nexport interface StructuredResponse<T> {\n  items: T[];\n  stopReason: 'end_turn' | 'max_tokens' | 'stop_sequence' | string;\n  usage?: TokenUsage;\n}\n\n/**\n * A provider's actual ceilings for the configured model, discovered from the\n * provider itself (Anthropic Models API; Ollama `/api/show`) — never\n * hand-maintained constants. Detection budget arithmetic derives from these.\n */\nexport interface InferenceLimits {\n  /**\n   * The context window in tokens. Semantics differ by provider shape:\n   * Anthropic reports maximum *input* tokens (output has its own ceiling);\n   * Ollama reports the *shared* input+output window and mirrors it in\n   * `maxOutputTokens` (there is no separate output ceiling), so\n   * `maxOutputTokens === contextTokens` signals a shared window.\n   */\n  contextTokens: number;\n  /** Maximum output tokens per generation. */\n  maxOutputTokens: number;\n  /**\n   * The provider's own worst-case output-rate model, in output tokens per\n   * hour, when it publishes one. Anthropic's SDK projects a call's maximum\n   * duration as `max_tokens / rate` (client.js `calculateNonstreamingTimeout`,\n   * 128_000/hour) and refuses non-streaming calls projected past 10 minutes —\n   * the one duration statement that provider surface makes. Consumers with\n   * their own call deadline derive a duration-safe output budget from it\n   * (ABANDONED-INFERENCE P4). Absent for providers whose rates are\n   * unknowable a priori (Ollama — local hardware). Absence does NOT mean no\n   * duration bound: the detection consumer applies its own conservative\n   * assumed floor rate instead (OLLAMA-DETECTION-TESTING P3b) — an unbounded\n   * budget turned model repetition loops into hour-long transient burns.\n   */\n  outputTokensPerHour?: number;\n}\n\n/**\n * Thrown when a structured generation's response cannot be read as the\n * requested array — never coerced to `[]` (empty is a legitimate, distinct\n * outcome). One class for every implementation, because the message shape\n * and the classification contract must not diverge between providers.\n *\n * Carries the provider's stop reason because the cause classifies\n * differently downstream: `max_tokens` means the JSON was cut off by the\n * output budget — the same input truncates the same way, so a retry is\n * guaranteed waste — while any other reason is model misbehavior a retry\n * may legitimately fix.\n *\n * Also thrown — on either generation path — when a response arrives EMPTY:\n * a thinking model can exhaust the whole output budget on hidden reasoning\n * before its first response character (measured live, gpt-oss:120b-cloud\n * 2026-09-05). Truncated-to-nothing is still truncation, and it needs the\n * same stop-reason ride to classify correctly.\n */\nexport class StructuredReadError extends Error {\n  override readonly name = 'StructuredReadError';\n  constructor(detail: string, readonly stopReason: string, options?: ErrorOptions) {\n    super(`Structured response could not be read: ${detail} (stop_reason: ${stopReason})`, options);\n  }\n}\n\nexport interface InferenceClient {\n  /** Provider type identifier (e.g. 'anthropic', 'ollama') */\n  readonly type: string;\n\n  /** Model identifier used for generation (e.g. 'claude-opus-4-6', 'llama3') */\n  readonly modelId: string;\n\n  /**\n   * How many INDEPENDENT inference calls a caller should run concurrently\n   * against this provider for a throughput gain (DETECTION-QUALITY-THROUGHPUT\n   * P6 — detection's per-type fan-out reads this).\n   *\n   * This is a property of the provider's economics, which is why it lives on\n   * the provider and not in the caller. A HOSTED API whose per-account rate\n   * limit sits far above one job's usage has real spare capacity, so >1\n   * genuinely parallelizes. A LOCAL single-model server (Ollama) is 1: its\n   * throughput is hardware-bound, so concurrent requests only queue or split\n   * one GPU — no aggregate speedup, and N live KV-cache contexts is memory\n   * pressure that can OOM. There is no honest default across those two worlds,\n   * so this is required, not optional.\n   *\n   * Hard-coded per implementation for now; the natural seam for future\n   * per-provider or admin tuning (a value that later comes from config changes\n   * only where this is SET, not the callers).\n   */\n  readonly maxConcurrency: number;\n\n  /**\n   * Whether detection should run the count-verifier against this provider's\n   * extractions (OLLAMA-DETECTION-TESTING P3c; universalized to every real\n   * provider by user ruling 2026-09-05 — unverified completeness is not a\n   * savings). Declared HERE, per implementation, because @semiont/jobs does no\n   * provider-specific switching (architecture ruling, same date): whatever\n   * varies by provider is a capability on this contract, like\n   * `maxConcurrency`. The mock alone defaults false, so deterministic tests\n   * opt in explicitly rather than paying a queue-popping count call by\n   * surprise.\n   */\n  readonly verifyDetectionYield: boolean;\n\n  /**\n   * The provider's actual context/output ceilings for `modelId`. Discovered\n   * lazily on first call and cached for the client's lifetime; a failed\n   * discovery is NOT cached — the next call retries. Throws when the ceilings\n   * cannot be determined (unknown model, discovery endpoint unreachable):\n   * fail-loud, never a guessed floor.\n   */\n  limits(): Promise<InferenceLimits>;\n\n  /**\n   * Generate text from a prompt (simple interface).\n   *\n   * `signal` (here and on every generation method — a trailing optional\n   * parameter, deliberately not an options bag; STRUCTURED-INFERENCE removed\n   * that shape on purpose): true cancellation, ABANDONED-INFERENCE P1.\n   * Implementations MUST thread it to their transport so an abort tears down\n   * the underlying request — and, for SDKs with internal retry loops, ends\n   * those too — rejecting promptly. Accepting the parameter and ignoring it\n   * is a defect worse than not having it: cancellation tests pass against\n   * such an adapter while zombie requests keep running (and billing) in\n   * production.\n   */\n  generateText(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<string>;\n\n  /**\n   * Generate text with detailed response information\n   */\n  generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<InferenceResponse>;\n\n  /**\n   * Generate a JSON array whose elements satisfy `elementSchema`, as parsed\n   * values — the structured counterpart of `generateTextWithMetadata`, and\n   * the ONLY generation surface detection may use.\n   *\n   * The return type carries the guarantee the old `format: 'json'` option\n   * left in a comment: callers receive `T[]` or an exception. When the\n   * provider's answer cannot be read as an array (the SDK hands tool input\n   * over as an unparsed string, the response is missing the array, the\n   * grammar was not honoured), implementations THROW a\n   * \"Structured response could not be read\" error — they never coerce to\n   * `[]`, because empty is a legitimate, distinct outcome.\n   */\n  generateStructured<T>(\n    prompt: string,\n    maxTokens: number,\n    temperature: number,\n    elementSchema: ElementSchema,\n    signal?: AbortSignal,\n  ): Promise<StructuredResponse<T>>;\n}\n","// Ollama implementation of InferenceClient interface\n// Uses native Ollama HTTP API (no SDK dependency)\n\nimport { Agent } from 'undici';\nimport { estimateTokens, isNumber, isObject } from '@semiont/core';\nimport type { Logger } from '@semiont/core';\nimport { recordInferenceUsage } from '@semiont/observability';\nimport { ElementSchema, InferenceClient, InferenceLimits, InferenceResponse, StructuredReadError, StructuredResponse } from '../interface.js';\n\n// With `stream: false` Ollama sends nothing — not even response headers —\n// until the whole generation finishes, so any transport-level header timeout\n// is a hidden generation ceiling: undici's 300s default killed every longer\n// generation as a retryable-looking `TypeError: fetch failed` before the\n// caller's own bound could fire. Both transport timeouts are disabled here so\n// the caller's AbortSignal is the single bound on a generate call.\n//\n// The assertion bridges a declaration skew only: @types/node types this slot\n// via undici-types@8, while the Agent must come from undici@7 to match Node\n// 24's bundled copy — an undici@8 Agent is rejected at runtime with\n// UND_ERR_INVALID_ARG (measured 2026-09-04; transport.test.ts gates both the\n// runtime compatibility and the failure shape).\ntype FetchDispatcher = NonNullable<RequestInit['dispatcher']>;\nexport const unboundedTransport = new Agent({ headersTimeout: 0, bodyTimeout: 0 }) as unknown as FetchDispatcher;\n\n// Slack added to the chars/4 prompt estimate when sizing `num_ctx`:\n// proportional to the estimate (the heuristic's error grows with prompt size)\n// plus a small fixed allowance for the model's chat template. The risk profile\n// is asymmetric — an undersized window silently clips input (the exact hole\n// managed num_ctx exists to close) while an oversized one only costs memory —\n// so the slack leans generous. Always capped at the model's real window.\nconst NUM_CTX_ESTIMATE_SLACK = 0.2;\nconst NUM_CTX_TEMPLATE_ALLOWANCE = 64;\n\ninterface OllamaGenerateResponse {\n  response: string;\n  done: boolean;\n  done_reason?: string;\n  /**\n   * Hidden reasoning from a thinking model. Requests always send\n   * `think: false`, but cloud-hosted reasoning models ignore it (measured\n   * live, gpt-oss:120b-cloud 2026-09-05): the thinking happens anyway,\n   * spends the caller's `num_predict` budget, and its tokens are folded\n   * invisibly into `eval_count`.\n   */\n  thinking?: string;\n  /** Number of prompt tokens evaluated. Available on most Ollama versions. */\n  prompt_eval_count?: number;\n  /** Number of tokens generated — INCLUDING any hidden thinking tokens. */\n  eval_count?: number;\n}\n\nexport class OllamaInferenceClient implements InferenceClient {\n  readonly type = 'ollama' as const;\n  // Local single model: generation throughput is hardware-bound, so concurrent\n  // requests queue or split one GPU for no aggregate speedup — and each live\n  // context costs KV-cache memory. Detection runs its types sequentially here\n  // (DETECTION-QUALITY-THROUGHPUT P6).\n  readonly maxConcurrency = 1;\n  // Where the collapse risk was MEASURED (F7) — the verifier's original home.\n  readonly verifyDetectionYield = true;\n  readonly modelId: string;\n  private baseURL: string;\n  private logger?: Logger;\n\n  private limitsPromise?: Promise<InferenceLimits>;\n\n  constructor(model: string, baseURL?: string, logger?: Logger) {\n    this.baseURL = (baseURL || 'http://localhost:11434').replace(/\\/+$/, '');\n    this.modelId = model;\n    this.logger = logger;\n  }\n\n  limits(): Promise<InferenceLimits> {\n    if (!this.limitsPromise) {\n      this.limitsPromise = this.discoverLimits().catch((err: unknown) => {\n        // Never cache a failed discovery — a transient outage would otherwise\n        // pin every future call to the same rejection.\n        this.limitsPromise = undefined;\n        throw err;\n      });\n    }\n    return this.limitsPromise;\n  }\n\n  private async discoverLimits(): Promise<InferenceLimits> {\n    const res = await fetch(`${this.baseURL}/api/show`, {\n      method: 'POST',\n      headers: { 'Content-Type': 'application/json' },\n      body: JSON.stringify({ model: this.modelId }),\n    });\n    if (!res.ok) {\n      throw new Error(\n        `Failed to discover model limits: /api/show returned ${res.status} for '${this.modelId}'`,\n      );\n    }\n    const data: unknown = await res.json();\n    const modelInfo = isObject(data) && isObject(data['model_info']) ? data['model_info'] : undefined;\n    const contextTokens = readContextLength(modelInfo);\n    if (contextTokens === undefined) {\n      throw new Error(`/api/show reports no context length for '${this.modelId}'`);\n    }\n    // Shared window: input and output draw from the same context — there is\n    // no separate output ceiling, so the window is published as both (the\n    // `maxOutputTokens === contextTokens` shape consumers key the split on).\n    return { contextTokens, maxOutputTokens: contextTokens };\n  }\n\n  async generateText(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<string> {\n    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature, signal);\n    return response.text;\n  }\n\n  async generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<InferenceResponse> {\n    return this.generate(prompt, maxTokens, temperature, undefined, signal);\n  }\n\n  async generateStructured<T>(\n    prompt: string,\n    maxTokens: number,\n    temperature: number,\n    elementSchema: ElementSchema,\n    signal?: AbortSignal,\n  ): Promise<StructuredResponse<T>> {\n    // Grammar-constrained sampling: the schema goes to Ollama's `format`\n    // parameter, which constrains generation itself — same mechanism as the\n    // old bare array schema, now element-typed. The response text is then\n    // parsed here, and anything that does not read as an array is a THROW,\n    // never a coerced [] — \"could not read the model\" must stay distinct\n    // from \"the model found nothing.\"\n    const response = await this.generate(prompt, maxTokens, temperature, elementSchema, signal);\n\n    let parsed: unknown;\n    try {\n      parsed = JSON.parse(response.text);\n    } catch (err) {\n      this.logger?.error('Structured response could not be read', {\n        model: this.modelId,\n        textLength: response.text.length,\n        stopReason: response.stopReason,\n      });\n      throw new StructuredReadError('response is not valid JSON', response.stopReason, { cause: err });\n    }\n    if (!Array.isArray(parsed)) {\n      this.logger?.error('Structured response could not be read', {\n        model: this.modelId,\n        parsedType: typeof parsed,\n        stopReason: response.stopReason,\n      });\n      throw new StructuredReadError(`parsed to ${typeof parsed}, not an array`, response.stopReason);\n    }\n\n    return { items: parsed as T[], stopReason: response.stopReason, ...(response.usage ? { usage: response.usage } : {}) };\n  }\n\n  private async generate(\n    prompt: string,\n    maxTokens: number,\n    temperature: number,\n    elementSchema: ElementSchema | undefined,\n    signal?: AbortSignal,\n  ): Promise<InferenceResponse> {\n    this.logger?.debug('Generating text with Ollama', {\n      model: this.modelId,\n      promptLength: prompt.length,\n      maxTokens,\n      temperature,\n      structured: elementSchema !== undefined,\n    });\n\n    // Managed context window: size num_ctx to cover this request, capped at\n    // the model's discovered window. Without an explicit num_ctx Ollama uses\n    // the model's *default* window and SILENTLY CLIPS any prompt beyond it —\n    // input loss with no error (found 2026-07-30).\n    const limits = await this.limits();\n    const promptTokens = estimateTokens(prompt);\n    if (promptTokens + maxTokens > limits.contextTokens) {\n      throw new Error(\n        `Prompt (~${promptTokens} tokens) + output budget (${maxTokens}) exceed the ` +\n        `'${this.modelId}' context window (${limits.contextTokens} tokens)`,\n      );\n    }\n    const numCtx = Math.min(\n      limits.contextTokens,\n      promptTokens + maxTokens\n        + Math.ceil(promptTokens * NUM_CTX_ESTIMATE_SLACK) + NUM_CTX_TEMPLATE_ALLOWANCE,\n    );\n\n    const url = `${this.baseURL}/api/generate`;\n    const start = performance.now();\n\n    // Ollama's `format` parameter accepts either the literal string\n    // `\"json\"` (any valid JSON, including objects, numbers, etc.) or a\n    // JSON schema (constrains the top-level shape). The structured contract\n    // is \"an array of elements matching the caller's schema,\" so we pass an\n    // array schema wrapping it — the bare `\"json\"` string would let the\n    // model satisfy \"valid JSON\" with `{\"entities\": [...]}` and break every\n    // consumer that maps over the top-level value.\n    const body: Record<string, unknown> = {\n      model: this.modelId,\n      prompt,\n      stream: false,\n      think: false,\n      options: {\n        num_predict: maxTokens,\n        num_ctx: numCtx,\n        temperature,\n      },\n    };\n    if (elementSchema !== undefined) {\n      // Grammar-enforced for locally served models. Cloud-routed models\n      // (`*-cloud`) treat this as ADVISORY only — a schema-violating response\n      // is possible there (measured live 2026-09-05: bare strings where the\n      // schema required objects) and surfaces as a StructuredReadError.\n      body['format'] = { type: 'array', items: elementSchema };\n    }\n\n    let res: Response;\n    try {\n      // True cancellation (ABANDONED-INFERENCE P1): the signal tears down the\n      // socket, so an aborted call cannot keep generating on the server's\n      // dime after its job is gone.\n      res = await fetch(url, {\n        method: 'POST',\n        headers: { 'Content-Type': 'application/json' },\n        body: JSON.stringify(body),\n        signal,\n        dispatcher: unboundedTransport,\n      });\n    } catch (err) {\n      recordInferenceUsage({\n        provider: this.type,\n        model: this.modelId,\n        durationMs: performance.now() - start,\n        outcome: 'error',\n      });\n      throw err;\n    }\n\n    if (!res.ok) {\n      recordInferenceUsage({\n        provider: this.type,\n        model: this.modelId,\n        durationMs: performance.now() - start,\n        outcome: 'error',\n      });\n      const body = await res.text();\n      this.logger?.error('Ollama API error', {\n        model: this.modelId,\n        status: res.status,\n        body,\n      });\n      throw new Error(`Ollama API error (${res.status}): ${body}`);\n    }\n\n    const data = await res.json() as OllamaGenerateResponse;\n    const stopReason = mapStopReason(data.done_reason);\n\n    if (data.thinking) {\n      // The adapter cannot prevent ignored `think: false` — it can only make\n      // the cost visible: the thinking billed, and `eval_count` (and any rate\n      // derived from it) is inflated by tokens that never reached the\n      // response.\n      this.logger?.warn('Model produced hidden thinking despite think:false', {\n        model: this.modelId,\n        thinkingChars: data.thinking.length,\n      });\n    }\n\n    if (!data.response) {\n      recordInferenceUsage({\n        provider: this.type,\n        model: this.modelId,\n        durationMs: performance.now() - start,\n        outcome: 'error',\n        inputTokens: data.prompt_eval_count,\n        outputTokens: data.eval_count,\n      });\n      this.logger?.error('Empty response from Ollama', {\n        model: this.modelId,\n        stopReason,\n        thinkingChars: data.thinking?.length,\n      });\n      // Truncated-to-nothing is still truncation: a thinking model can burn\n      // the entire output budget before its first response character. The\n      // stop reason must ride the error so `max_tokens` classifies\n      // deterministic (and subdivides) instead of masquerading as an\n      // unrecognized retryable mystery; any other stop reason stays\n      // retryable, exactly as an untyped error did.\n      throw new StructuredReadError('response is empty', stopReason);\n    }\n\n    recordInferenceUsage({\n      provider: this.type,\n      model: this.modelId,\n      durationMs: performance.now() - start,\n      outcome: 'success',\n      inputTokens: data.prompt_eval_count,\n      outputTokens: data.eval_count,\n    });\n\n    this.logger?.info('Text generation completed', {\n      model: this.modelId,\n      textLength: data.response.length,\n      stopReason,\n    });\n\n    return {\n      text: data.response,\n      stopReason,\n      ...(data.prompt_eval_count !== undefined && data.eval_count !== undefined\n        ? { usage: { inputTokens: data.prompt_eval_count, outputTokens: data.eval_count } }\n        : {}),\n    };\n  }\n}\n\n/**\n * The context length lives in `model_info` under an architecture-prefixed key\n * (e.g. `llama.context_length`); `general.architecture` names the prefix.\n * Falls back to any `*.context_length` key for models whose metadata omits\n * the architecture field.\n */\nfunction readContextLength(modelInfo: Record<string, unknown> | undefined): number | undefined {\n  if (!modelInfo) return undefined;\n  const arch = modelInfo['general.architecture'];\n  if (typeof arch === 'string') {\n    const direct = modelInfo[`${arch}.context_length`];\n    if (isNumber(direct) && direct > 0) return direct;\n  }\n  const fallbackKey = Object.keys(modelInfo).find(k => k.endsWith('.context_length'));\n  if (fallbackKey !== undefined) {\n    const fallback = modelInfo[fallbackKey];\n    if (isNumber(fallback) && fallback > 0) return fallback;\n  }\n  return undefined;\n}\n\nfunction mapStopReason(doneReason: string | undefined): string {\n  switch (doneReason) {\n    case 'stop': return 'end_turn';\n    case 'length': return 'max_tokens';\n    default: return doneReason || 'unknown';\n  }\n}\n","// Factory for creating inference client instances based on configuration\n\nimport type { Logger } from '@semiont/core';\nimport { InferenceClient } from './interface.js';\nimport { AnthropicInferenceClient } from './implementations/anthropic.js';\nimport { OllamaInferenceClient } from './implementations/ollama.js';\n\nexport type InferenceClientType = 'anthropic' | 'ollama';\n\nexport interface InferenceClientConfig {\n  type: InferenceClientType;\n  apiKey?: string;\n  model: string;\n  endpoint?: string;\n  baseURL?: string;\n}\n\nexport function createInferenceClient(config: InferenceClientConfig, logger?: Logger): InferenceClient {\n  switch (config.type) {\n    case 'anthropic': {\n      if (!config.apiKey || config.apiKey.trim() === '') {\n        throw new Error('apiKey is required for Anthropic inference client');\n      }\n      return new AnthropicInferenceClient(\n        config.apiKey,\n        config.model,\n        config.endpoint || config.baseURL,\n        logger\n      );\n    }\n\n    case 'ollama': {\n      return new OllamaInferenceClient(\n        config.model,\n        config.endpoint || config.baseURL,\n        logger\n      );\n    }\n\n    default:\n      throw new Error(`Unsupported inference client type: ${config.type}`);\n  }\n}\n","// Mock implementation of InferenceClient for testing\n\nimport { ElementSchema, InferenceClient, InferenceLimits, InferenceResponse, StructuredReadError, StructuredResponse } from '../interface.js';\n\n// Generous defaults so existing consumers never trip chunking or window\n// guards unless a test injects tighter limits deliberately.\nconst GENEROUS_LIMITS: InferenceLimits = {\n  contextTokens: 1_000_000,\n  maxOutputTokens: 1_000_000,\n  // Generous RATE too: without one the mock reads as rate-silent, which since\n  // OLLAMA-DETECTION-TESTING P3b/P3c opts consumers into the assumed duration\n  // floor and the count-verifier. Tests exercise those by injecting\n  // rate-silent limits deliberately, never by the default.\n  outputTokensPerHour: 3_600_000_000,\n};\n\nexport class MockInferenceClient implements InferenceClient {\n  readonly type = 'mock' as const;\n  readonly modelId = 'mock-model' as const;\n  // Deterministic default for tests; a test exercising concurrency sets its own.\n  readonly maxConcurrency = 1;\n  // Deterministic default: a count call pops the shared response queue, so\n  // tests exercising the verifier declare their own true rather than every\n  // consumer paying it by surprise.\n  readonly verifyDetectionYield = false;\n  private responses: string[] = [];\n  private responseIndex: number = 0;\n  private stopReasons: string[] = [];\n  private injectedLimits: InferenceLimits;\n  public calls: Array<{ prompt: string; maxTokens: number; temperature: number; elementSchema?: ElementSchema }> = [];\n\n  constructor(responses: string[] = ['Mock response'], stopReasons?: string[], limits?: InferenceLimits) {\n    this.responses = responses;\n    this.stopReasons = stopReasons || responses.map(() => 'end_turn');\n    this.injectedLimits = limits ?? GENEROUS_LIMITS;\n  }\n\n  async limits(): Promise<InferenceLimits> {\n    return this.injectedLimits;\n  }\n\n  async generateText(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<string> {\n    const response = await this.generateTextWithMetadata(prompt, maxTokens, temperature, signal);\n    return response.text;\n  }\n\n  async generateTextWithMetadata(prompt: string, maxTokens: number, temperature: number, signal?: AbortSignal): Promise<InferenceResponse> {\n    throwIfAborted(signal);\n    this.calls.push({ prompt, maxTokens, temperature });\n    return this.nextResponse();\n  }\n\n  /**\n   * Structured surface: pops the same responses queue and PARSES the entry,\n   * mirroring the real contract — a queued string that is not a JSON array\n   * throws \"could not be read\", so tests inject the malformed shape simply by\n   * queuing it (`setResponses(['not json'])`). The element schema is recorded\n   * on `calls` so tests can assert what the caller declared.\n   */\n  async generateStructured<T>(\n    prompt: string,\n    maxTokens: number,\n    temperature: number,\n    elementSchema: ElementSchema,\n    signal?: AbortSignal,\n  ): Promise<StructuredResponse<T>> {\n    throwIfAborted(signal);\n    this.calls.push({ prompt, maxTokens, temperature, elementSchema });\n    const { text, stopReason } = this.nextResponse();\n\n    let parsed: unknown;\n    try {\n      parsed = JSON.parse(text);\n    } catch (err) {\n      throw new StructuredReadError('response is not valid JSON', stopReason, { cause: err });\n    }\n    if (!Array.isArray(parsed)) {\n      throw new StructuredReadError(`parsed to ${typeof parsed}, not an array`, stopReason);\n    }\n    return { items: parsed as T[], stopReason };\n  }\n\n  private nextResponse(): InferenceResponse {\n    const text = this.responses[this.responseIndex];\n    const stopReason = this.stopReasons[this.responseIndex] || 'end_turn';\n\n    if (this.responseIndex < this.responses.length - 1) {\n      this.responseIndex++;\n    }\n\n    return { text, stopReason };\n  }\n\n  // Test helper methods\n  reset(): void {\n    this.calls = [];\n    this.responseIndex = 0;\n  }\n\n  setResponses(responses: string[], stopReasons?: string[]): void {\n    this.responses = responses;\n    this.stopReasons = stopReasons || responses.map(() => 'end_turn');\n    this.responseIndex = 0;\n  }\n}\n\n/**\n * The mock honors the signal like a real adapter (ABANDONED-INFERENCE P1's\n * accept-and-drop trap: an adapter that takes the parameter and ignores it\n * lets cancellation tests pass while proving nothing). The mock resolves\n * synchronously, so an entry check is the whole contract — rejecting the way\n * an aborted `fetch` does, with an `AbortError`-named DOMException.\n */\nfunction throwIfAborted(signal?: AbortSignal): void {\n  if (signal?.aborted) {\n    throw new DOMException('This operation was aborted', 'AbortError');\n  }\n}\n"],"mappings":";AAEA,OAAO,eAAe;AACtB,SAAS,gBAA6B;AACtC,SAAS,4BAA4B;;;ACoG9B,IAAM,sBAAN,cAAkC,MAAM;AAAA,EAE7C,YAAY,QAAyB,YAAoB,SAAwB;AAC/E,UAAM,0CAA0C,MAAM,kBAAkB,UAAU,KAAK,OAAO;AAD3D;AAAA,EAErC;AAAA,EAFqC;AAAA,EADnB,OAAO;AAI3B;;;AD7FA,IAAM,yBAAyB;AAK/B,IAAM,iCAAiC,KAAK,MAAM,yBAAyB,CAAC;AAyBrE,IAAM,2BAAN,MAA0D;AAAA,EACtD,OAAO;AAAA;AAAA;AAAA;AAAA;AAAA,EAKP,iBAAiB;AAAA;AAAA;AAAA;AAAA;AAAA,EAKjB,uBAAuB;AAAA,EACvB;AAAA,EACD;AAAA,EACA;AAAA,EACA;AAAA,EAER,YAAY,QAAgB,OAAe,SAAkB,QAAiB;AAC5E,SAAK,SAAS,IAAI,UAAU;AAAA,MAC1B;AAAA,MACA,SAAS,WAAW;AAAA,IACtB,CAAC;AACD,SAAK,UAAU;AACf,SAAK,SAAS;AAAA,EAChB;AAAA,EAEA,SAAmC;AACjC,WAAO,KAAK,SAAS,EAAE,KAAK,OAAK,EAAE,MAAM;AAAA,EAC3C;AAAA,EAEQ,WAAoC;AAC1C,QAAI,CAAC,KAAK,kBAAkB;AAC1B,WAAK,mBAAmB,KAAK,cAAc,EAAE,MAAM,CAAC,QAAiB;AAGnE,aAAK,mBAAmB;AACxB,cAAM;AAAA,MACR,CAAC;AAAA,IACH;AACA,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,gBAAyC;AAMrD,UAAM,OAAO,MAAM,KAAK,OAAO,OAAO,SAAS,KAAK,OAAO,EAAE,MAAM,CAAC,QAAiB;AACnF,YAAM,IAAI;AAAA,QACR,wCAAwC,KAAK,OAAO;AAAA,QACpD,EAAE,OAAO,IAAI;AAAA,MACf;AAAA,IACF,CAAC;AACD,QAAI,KAAK,oBAAoB,QAAQ,KAAK,cAAc,MAAM;AAC5D,YAAM,IAAI,MAAM,sDAAsD,KAAK,OAAO,GAAG;AAAA,IACvF;AAIA,UAAM,MAAe;AACrB,UAAM,6BACJ,SAAS,GAAG,KACZ,SAAS,IAAI,cAAc,CAAC,KAC5B,SAAS,IAAI,cAAc,EAAE,oBAAoB,CAAC,KAClD,IAAI,cAAc,EAAE,oBAAoB,EAAE,WAAW,MAAM;AAC7D,WAAO;AAAA,MACL,QAAQ;AAAA,QACN,eAAe,KAAK;AAAA,QACpB,iBAAiB,KAAK;AAAA,QACtB,qBAAqB;AAAA,MACvB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,eAAe,QAAmD,QAAkD;AAQ1H,QAAI,OAAO,aAAa,gCAAgC;AACtD,aAAO,KAAK,OAAO,SAAS,OAAO,QAAQ,EAAE,OAAO,CAAC,EAAE,aAAa;AAAA,IACtE;AACA,WAAO,KAAK,OAAO,SAAS,OAAO,QAAQ,EAAE,OAAO,CAAC;AAAA,EACvD;AAAA,EAEA,MAAM,aAAa,QAAgB,WAAmB,aAAqB,QAAuC;AAChH,UAAM,WAAW,MAAM,KAAK,yBAAyB,QAAQ,WAAW,aAAa,MAAM;AAC3F,WAAO,SAAS;AAAA,EAClB;AAAA,EAEA,MAAM,yBAAyB,QAAgB,WAAmB,aAAqB,QAAkD;AACvI,SAAK,QAAQ,MAAM,yCAAyC;AAAA,MAC1D,OAAO,KAAK;AAAA,MACZ,cAAc,OAAO;AAAA,MACrB;AAAA,MACA;AAAA,IACF,CAAC;AAED,UAAM,SAAoD;AAAA,MACxD,OAAO,KAAK;AAAA,MACZ,YAAY;AAAA,MACZ;AAAA,MACA,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,OAAO,CAAC;AAAA,IAC9C;AAEA,UAAM,QAAQ,YAAY,IAAI;AAC9B,UAAM,WAAW,MAAM,KAAK,gBAAgB,QAAQ,OAAO,MAAM;AAEjE,UAAM,cAAc,SAAS,QAAQ,KAAK,OAAK,EAAE,SAAS,MAAM;AAChE,QAAI,CAAC,eAAe,YAAY,SAAS,QAAQ;AAC/C,WAAK,YAAY,OAAO,QAAQ;AAChC,WAAK,QAAQ,MAAM,yCAAyC;AAAA,QAC1D,OAAO,KAAK;AAAA,QACZ,cAAc,SAAS,QAAQ,IAAI,OAAK,EAAE,IAAI;AAAA,MAChD,CAAC;AACD,YAAM,IAAI,MAAM,uCAAuC;AAAA,IACzD;AACA,UAAM,OAAO,YAAY;AAEzB,yBAAqB;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,OAAO,KAAK;AAAA,MACZ,YAAY,YAAY,IAAI,IAAI;AAAA,MAChC,SAAS;AAAA,MACT,aAAa,SAAS,OAAO;AAAA,MAC7B,cAAc,SAAS,OAAO;AAAA,IAChC,CAAC;AAED,SAAK,QAAQ,KAAK,6BAA6B;AAAA,MAC7C,OAAO,KAAK;AAAA,MACZ,YAAY,KAAK;AAAA,MACjB,YAAY,SAAS;AAAA,MACrB,WAAW,YAAY,QAAQ;AAAA,IACjC,CAAC;AAED,WAAO;AAAA,MACL;AAAA,MACA,YAAY,SAAS,eAAe;AAAA,IACtC;AAAA,EACF;AAAA,EAEA,MAAM,mBACJ,QACA,WACA,aACA,eACA,QACgC;AAOhC,UAAM,YAAY,MAAM,KAAK,SAAS;AACtC,QAAI,CAAC,UAAU,4BAA4B;AACzC,YAAM,IAAI;AAAA,QACR,UAAU,KAAK,OAAO;AAAA,MAMxB;AAAA,IACF;AAEA,SAAK,QAAQ,MAAM,sDAAsD;AAAA,MACvE,OAAO,KAAK;AAAA,MACZ,cAAc,OAAO;AAAA,MACrB;AAAA,MACA;AAAA,IACF,CAAC;AAED,UAAM,SAAoD;AAAA,MACxD,OAAO,KAAK;AAAA,MACZ,YAAY;AAAA,MACZ;AAAA,MACA,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,OAAO,CAAC;AAAA;AAAA;AAAA,MAG5C,eAAe;AAAA,QACb,QAAQ;AAAA,UACN,MAAM;AAAA,UACN,QAAQ,EAAE,MAAM,SAAS,OAAO,cAAc;AAAA,QAChD;AAAA,MACF;AAAA,IACF;AAEA,UAAM,QAAQ,YAAY,IAAI;AAC9B,UAAM,WAAW,MAAM,KAAK,gBAAgB,QAAQ,OAAO,MAAM;AAEjE,UAAM,cAAc,SAAS,QAAQ,KAAK,OAAK,EAAE,SAAS,MAAM;AAChE,QAAI,CAAC,eAAe,YAAY,SAAS,QAAQ;AAC/C,WAAK,YAAY,OAAO,QAAQ;AAChC,WAAK,QAAQ,MAAM,oDAAoD;AAAA,QACrE,OAAO,KAAK;AAAA,QACZ,cAAc,SAAS,QAAQ,IAAI,OAAK,EAAE,IAAI;AAAA,MAChD,CAAC;AACD,YAAM,IAAI,MAAM,kDAAkD;AAAA,IACpE;AAOA,QAAI;AACJ,QAAI;AACF,eAAS,KAAK,MAAM,YAAY,IAAI;AAAA,IACtC,SAAS,KAAK;AACZ,WAAK,YAAY,OAAO,QAAQ;AAChC,WAAK,QAAQ,MAAM,yCAAyC;AAAA,QAC1D,OAAO,KAAK;AAAA,QACZ,YAAY,YAAY,KAAK;AAAA,QAC7B,YAAY,SAAS;AAAA,MACvB,CAAC;AACD,YAAM,IAAI,oBAAoB,8BAA8B,SAAS,eAAe,WAAW,EAAE,OAAO,IAAI,CAAC;AAAA,IAC/G;AACA,QAAI,CAAC,MAAM,QAAQ,MAAM,GAAG;AAC1B,WAAK,YAAY,OAAO,QAAQ;AAChC,WAAK,QAAQ,MAAM,yCAAyC;AAAA,QAC1D,OAAO,KAAK;AAAA,QACZ,YAAY,OAAO;AAAA,QACnB,YAAY,SAAS;AAAA,MACvB,CAAC;AACD,YAAM,IAAI,oBAAoB,aAAa,OAAO,MAAM,kBAAkB,SAAS,eAAe,SAAS;AAAA,IAC7G;AAEA,yBAAqB;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,OAAO,KAAK;AAAA,MACZ,YAAY,YAAY,IAAI,IAAI;AAAA,MAChC,SAAS;AAAA,MACT,aAAa,SAAS,OAAO;AAAA,MAC7B,cAAc,SAAS,OAAO;AAAA,IAChC,CAAC;AAED,SAAK,QAAQ,KAAK,mCAAmC;AAAA,MACnD,OAAO,KAAK;AAAA,MACZ,OAAO,OAAO;AAAA,MACd,YAAY,SAAS;AAAA,MACrB,WAAW,YAAY,QAAQ;AAAA,IACjC,CAAC;AAED,WAAO;AAAA,MACL,OAAO;AAAA,MACP,YAAY,SAAS,eAAe;AAAA,MACpC,GAAG,QAAQ,QAAQ;AAAA,IACrB;AAAA,EACF;AAAA;AAAA,EAGA,MAAc,gBAAgB,QAAmD,OAAe,QAAkD;AAChJ,QAAI;AACF,aAAO,MAAM,KAAK,eAAe,QAAQ,MAAM;AAAA,IACjD,SAAS,KAAK;AACZ,2BAAqB;AAAA,QACnB,UAAU,KAAK;AAAA,QACf,OAAO,KAAK;AAAA,QACZ,YAAY,YAAY,IAAI,IAAI;AAAA,QAChC,SAAS;AAAA,MACX,CAAC;AACD,YAAM;AAAA,IACR;AAAA,EACF;AAAA,EAEQ,YAAY,OAAe,UAAmC;AACpE,yBAAqB;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,OAAO,KAAK;AAAA,MACZ,YAAY,YAAY,IAAI,IAAI;AAAA,MAChC,SAAS;AAAA,MACT,aAAa,SAAS,OAAO;AAAA,MAC7B,cAAc,SAAS,OAAO;AAAA,IAChC,CAAC;AAAA,EACH;AACF;AAQA,SAAS,YAAY,UAAuC;AAC1D,MAAI,SAAS,QAAQ,KAAK,OAAO,SAAS,aAAa,MAAM,UAAU;AACrE,WAAO,SAAS,aAAa;AAAA,EAC/B;AACA,SAAO;AACT;AAOA,SAAS,QAAQ,UAAiG;AAChH,QAAM,EAAE,cAAc,cAAc,IAAI,SAAS,SAAS,CAAC;AAC3D,MAAI,iBAAiB,UAAa,kBAAkB,OAAW,QAAO,CAAC;AACvE,SAAO,EAAE,OAAO,EAAE,aAAa,cAAc,cAAc,cAAc,EAAE;AAC7E;;;AE5VA,SAAS,aAAa;AACtB,SAAS,gBAAgB,UAAU,YAAAA,iBAAgB;AAEnD,SAAS,wBAAAC,6BAA4B;AAgB9B,IAAM,qBAAqB,IAAI,MAAM,EAAE,gBAAgB,GAAG,aAAa,EAAE,CAAC;AAQjF,IAAM,yBAAyB;AAC/B,IAAM,6BAA6B;AAoB5B,IAAM,wBAAN,MAAuD;AAAA,EACnD,OAAO;AAAA;AAAA;AAAA;AAAA;AAAA,EAKP,iBAAiB;AAAA;AAAA,EAEjB,uBAAuB;AAAA,EACvB;AAAA,EACD;AAAA,EACA;AAAA,EAEA;AAAA,EAER,YAAY,OAAe,SAAkB,QAAiB;AAC5D,SAAK,WAAW,WAAW,0BAA0B,QAAQ,QAAQ,EAAE;AACvE,SAAK,UAAU;AACf,SAAK,SAAS;AAAA,EAChB;AAAA,EAEA,SAAmC;AACjC,QAAI,CAAC,KAAK,eAAe;AACvB,WAAK,gBAAgB,KAAK,eAAe,EAAE,MAAM,CAAC,QAAiB;AAGjE,aAAK,gBAAgB;AACrB,cAAM;AAAA,MACR,CAAC;AAAA,IACH;AACA,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,iBAA2C;AACvD,UAAM,MAAM,MAAM,MAAM,GAAG,KAAK,OAAO,aAAa;AAAA,MAClD,QAAQ;AAAA,MACR,SAAS,EAAE,gBAAgB,mBAAmB;AAAA,MAC9C,MAAM,KAAK,UAAU,EAAE,OAAO,KAAK,QAAQ,CAAC;AAAA,IAC9C,CAAC;AACD,QAAI,CAAC,IAAI,IAAI;AACX,YAAM,IAAI;AAAA,QACR,uDAAuD,IAAI,MAAM,SAAS,KAAK,OAAO;AAAA,MACxF;AAAA,IACF;AACA,UAAM,OAAgB,MAAM,IAAI,KAAK;AACrC,UAAM,YAAYC,UAAS,IAAI,KAAKA,UAAS,KAAK,YAAY,CAAC,IAAI,KAAK,YAAY,IAAI;AACxF,UAAM,gBAAgB,kBAAkB,SAAS;AACjD,QAAI,kBAAkB,QAAW;AAC/B,YAAM,IAAI,MAAM,4CAA4C,KAAK,OAAO,GAAG;AAAA,IAC7E;AAIA,WAAO,EAAE,eAAe,iBAAiB,cAAc;AAAA,EACzD;AAAA,EAEA,MAAM,aAAa,QAAgB,WAAmB,aAAqB,QAAuC;AAChH,UAAM,WAAW,MAAM,KAAK,yBAAyB,QAAQ,WAAW,aAAa,MAAM;AAC3F,WAAO,SAAS;AAAA,EAClB;AAAA,EAEA,MAAM,yBAAyB,QAAgB,WAAmB,aAAqB,QAAkD;AACvI,WAAO,KAAK,SAAS,QAAQ,WAAW,aAAa,QAAW,MAAM;AAAA,EACxE;AAAA,EAEA,MAAM,mBACJ,QACA,WACA,aACA,eACA,QACgC;AAOhC,UAAM,WAAW,MAAM,KAAK,SAAS,QAAQ,WAAW,aAAa,eAAe,MAAM;AAE1F,QAAI;AACJ,QAAI;AACF,eAAS,KAAK,MAAM,SAAS,IAAI;AAAA,IACnC,SAAS,KAAK;AACZ,WAAK,QAAQ,MAAM,yCAAyC;AAAA,QAC1D,OAAO,KAAK;AAAA,QACZ,YAAY,SAAS,KAAK;AAAA,QAC1B,YAAY,SAAS;AAAA,MACvB,CAAC;AACD,YAAM,IAAI,oBAAoB,8BAA8B,SAAS,YAAY,EAAE,OAAO,IAAI,CAAC;AAAA,IACjG;AACA,QAAI,CAAC,MAAM,QAAQ,MAAM,GAAG;AAC1B,WAAK,QAAQ,MAAM,yCAAyC;AAAA,QAC1D,OAAO,KAAK;AAAA,QACZ,YAAY,OAAO;AAAA,QACnB,YAAY,SAAS;AAAA,MACvB,CAAC;AACD,YAAM,IAAI,oBAAoB,aAAa,OAAO,MAAM,kBAAkB,SAAS,UAAU;AAAA,IAC/F;AAEA,WAAO,EAAE,OAAO,QAAe,YAAY,SAAS,YAAY,GAAI,SAAS,QAAQ,EAAE,OAAO,SAAS,MAAM,IAAI,CAAC,EAAG;AAAA,EACvH;AAAA,EAEA,MAAc,SACZ,QACA,WACA,aACA,eACA,QAC4B;AAC5B,SAAK,QAAQ,MAAM,+BAA+B;AAAA,MAChD,OAAO,KAAK;AAAA,MACZ,cAAc,OAAO;AAAA,MACrB;AAAA,MACA;AAAA,MACA,YAAY,kBAAkB;AAAA,IAChC,CAAC;AAMD,UAAM,SAAS,MAAM,KAAK,OAAO;AACjC,UAAM,eAAe,eAAe,MAAM;AAC1C,QAAI,eAAe,YAAY,OAAO,eAAe;AACnD,YAAM,IAAI;AAAA,QACR,YAAY,YAAY,6BAA6B,SAAS,iBAC1D,KAAK,OAAO,qBAAqB,OAAO,aAAa;AAAA,MAC3D;AAAA,IACF;AACA,UAAM,SAAS,KAAK;AAAA,MAClB,OAAO;AAAA,MACP,eAAe,YACX,KAAK,KAAK,eAAe,sBAAsB,IAAI;AAAA,IACzD;AAEA,UAAM,MAAM,GAAG,KAAK,OAAO;AAC3B,UAAM,QAAQ,YAAY,IAAI;AAS9B,UAAM,OAAgC;AAAA,MACpC,OAAO,KAAK;AAAA,MACZ;AAAA,MACA,QAAQ;AAAA,MACR,OAAO;AAAA,MACP,SAAS;AAAA,QACP,aAAa;AAAA,QACb,SAAS;AAAA,QACT;AAAA,MACF;AAAA,IACF;AACA,QAAI,kBAAkB,QAAW;AAK/B,WAAK,QAAQ,IAAI,EAAE,MAAM,SAAS,OAAO,cAAc;AAAA,IACzD;AAEA,QAAI;AACJ,QAAI;AAIF,YAAM,MAAM,MAAM,KAAK;AAAA,QACrB,QAAQ;AAAA,QACR,SAAS,EAAE,gBAAgB,mBAAmB;AAAA,QAC9C,MAAM,KAAK,UAAU,IAAI;AAAA,QACzB;AAAA,QACA,YAAY;AAAA,MACd,CAAC;AAAA,IACH,SAAS,KAAK;AACZ,MAAAC,sBAAqB;AAAA,QACnB,UAAU,KAAK;AAAA,QACf,OAAO,KAAK;AAAA,QACZ,YAAY,YAAY,IAAI,IAAI;AAAA,QAChC,SAAS;AAAA,MACX,CAAC;AACD,YAAM;AAAA,IACR;AAEA,QAAI,CAAC,IAAI,IAAI;AACX,MAAAA,sBAAqB;AAAA,QACnB,UAAU,KAAK;AAAA,QACf,OAAO,KAAK;AAAA,QACZ,YAAY,YAAY,IAAI,IAAI;AAAA,QAChC,SAAS;AAAA,MACX,CAAC;AACD,YAAMC,QAAO,MAAM,IAAI,KAAK;AAC5B,WAAK,QAAQ,MAAM,oBAAoB;AAAA,QACrC,OAAO,KAAK;AAAA,QACZ,QAAQ,IAAI;AAAA,QACZ,MAAAA;AAAA,MACF,CAAC;AACD,YAAM,IAAI,MAAM,qBAAqB,IAAI,MAAM,MAAMA,KAAI,EAAE;AAAA,IAC7D;AAEA,UAAM,OAAO,MAAM,IAAI,KAAK;AAC5B,UAAM,aAAa,cAAc,KAAK,WAAW;AAEjD,QAAI,KAAK,UAAU;AAKjB,WAAK,QAAQ,KAAK,sDAAsD;AAAA,QACtE,OAAO,KAAK;AAAA,QACZ,eAAe,KAAK,SAAS;AAAA,MAC/B,CAAC;AAAA,IACH;AAEA,QAAI,CAAC,KAAK,UAAU;AAClB,MAAAD,sBAAqB;AAAA,QACnB,UAAU,KAAK;AAAA,QACf,OAAO,KAAK;AAAA,QACZ,YAAY,YAAY,IAAI,IAAI;AAAA,QAChC,SAAS;AAAA,QACT,aAAa,KAAK;AAAA,QAClB,cAAc,KAAK;AAAA,MACrB,CAAC;AACD,WAAK,QAAQ,MAAM,8BAA8B;AAAA,QAC/C,OAAO,KAAK;AAAA,QACZ;AAAA,QACA,eAAe,KAAK,UAAU;AAAA,MAChC,CAAC;AAOD,YAAM,IAAI,oBAAoB,qBAAqB,UAAU;AAAA,IAC/D;AAEA,IAAAA,sBAAqB;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,OAAO,KAAK;AAAA,MACZ,YAAY,YAAY,IAAI,IAAI;AAAA,MAChC,SAAS;AAAA,MACT,aAAa,KAAK;AAAA,MAClB,cAAc,KAAK;AAAA,IACrB,CAAC;AAED,SAAK,QAAQ,KAAK,6BAA6B;AAAA,MAC7C,OAAO,KAAK;AAAA,MACZ,YAAY,KAAK,SAAS;AAAA,MAC1B;AAAA,IACF,CAAC;AAED,WAAO;AAAA,MACL,MAAM,KAAK;AAAA,MACX;AAAA,MACA,GAAI,KAAK,sBAAsB,UAAa,KAAK,eAAe,SAC5D,EAAE,OAAO,EAAE,aAAa,KAAK,mBAAmB,cAAc,KAAK,WAAW,EAAE,IAChF,CAAC;AAAA,IACP;AAAA,EACF;AACF;AAQA,SAAS,kBAAkB,WAAoE;AAC7F,MAAI,CAAC,UAAW,QAAO;AACvB,QAAM,OAAO,UAAU,sBAAsB;AAC7C,MAAI,OAAO,SAAS,UAAU;AAC5B,UAAM,SAAS,UAAU,GAAG,IAAI,iBAAiB;AACjD,QAAI,SAAS,MAAM,KAAK,SAAS,EAAG,QAAO;AAAA,EAC7C;AACA,QAAM,cAAc,OAAO,KAAK,SAAS,EAAE,KAAK,OAAK,EAAE,SAAS,iBAAiB,CAAC;AAClF,MAAI,gBAAgB,QAAW;AAC7B,UAAM,WAAW,UAAU,WAAW;AACtC,QAAI,SAAS,QAAQ,KAAK,WAAW,EAAG,QAAO;AAAA,EACjD;AACA,SAAO;AACT;AAEA,SAAS,cAAc,YAAwC;AAC7D,UAAQ,YAAY;AAAA,IAClB,KAAK;AAAQ,aAAO;AAAA,IACpB,KAAK;AAAU,aAAO;AAAA,IACtB;AAAS,aAAO,cAAc;AAAA,EAChC;AACF;;;ACtUO,SAAS,sBAAsB,QAA+B,QAAkC;AACrG,UAAQ,OAAO,MAAM;AAAA,IACnB,KAAK,aAAa;AAChB,UAAI,CAAC,OAAO,UAAU,OAAO,OAAO,KAAK,MAAM,IAAI;AACjD,cAAM,IAAI,MAAM,mDAAmD;AAAA,MACrE;AACA,aAAO,IAAI;AAAA,QACT,OAAO;AAAA,QACP,OAAO;AAAA,QACP,OAAO,YAAY,OAAO;AAAA,QAC1B;AAAA,MACF;AAAA,IACF;AAAA,IAEA,KAAK,UAAU;AACb,aAAO,IAAI;AAAA,QACT,OAAO;AAAA,QACP,OAAO,YAAY,OAAO;AAAA,QAC1B;AAAA,MACF;AAAA,IACF;AAAA,IAEA;AACE,YAAM,IAAI,MAAM,sCAAsC,OAAO,IAAI,EAAE;AAAA,EACvE;AACF;;;ACpCA,IAAM,kBAAmC;AAAA,EACvC,eAAe;AAAA,EACf,iBAAiB;AAAA;AAAA;AAAA;AAAA;AAAA,EAKjB,qBAAqB;AACvB;AAEO,IAAM,sBAAN,MAAqD;AAAA,EACjD,OAAO;AAAA,EACP,UAAU;AAAA;AAAA,EAEV,iBAAiB;AAAA;AAAA;AAAA;AAAA,EAIjB,uBAAuB;AAAA,EACxB,YAAsB,CAAC;AAAA,EACvB,gBAAwB;AAAA,EACxB,cAAwB,CAAC;AAAA,EACzB;AAAA,EACD,QAA0G,CAAC;AAAA,EAElH,YAAY,YAAsB,CAAC,eAAe,GAAG,aAAwB,QAA0B;AACrG,SAAK,YAAY;AACjB,SAAK,cAAc,eAAe,UAAU,IAAI,MAAM,UAAU;AAChE,SAAK,iBAAiB,UAAU;AAAA,EAClC;AAAA,EAEA,MAAM,SAAmC;AACvC,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,aAAa,QAAgB,WAAmB,aAAqB,QAAuC;AAChH,UAAM,WAAW,MAAM,KAAK,yBAAyB,QAAQ,WAAW,aAAa,MAAM;AAC3F,WAAO,SAAS;AAAA,EAClB;AAAA,EAEA,MAAM,yBAAyB,QAAgB,WAAmB,aAAqB,QAAkD;AACvI,mBAAe,MAAM;AACrB,SAAK,MAAM,KAAK,EAAE,QAAQ,WAAW,YAAY,CAAC;AAClD,WAAO,KAAK,aAAa;AAAA,EAC3B;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,mBACJ,QACA,WACA,aACA,eACA,QACgC;AAChC,mBAAe,MAAM;AACrB,SAAK,MAAM,KAAK,EAAE,QAAQ,WAAW,aAAa,cAAc,CAAC;AACjE,UAAM,EAAE,MAAM,WAAW,IAAI,KAAK,aAAa;AAE/C,QAAI;AACJ,QAAI;AACF,eAAS,KAAK,MAAM,IAAI;AAAA,IAC1B,SAAS,KAAK;AACZ,YAAM,IAAI,oBAAoB,8BAA8B,YAAY,EAAE,OAAO,IAAI,CAAC;AAAA,IACxF;AACA,QAAI,CAAC,MAAM,QAAQ,MAAM,GAAG;AAC1B,YAAM,IAAI,oBAAoB,aAAa,OAAO,MAAM,kBAAkB,UAAU;AAAA,IACtF;AACA,WAAO,EAAE,OAAO,QAAe,WAAW;AAAA,EAC5C;AAAA,EAEQ,eAAkC;AACxC,UAAM,OAAO,KAAK,UAAU,KAAK,aAAa;AAC9C,UAAM,aAAa,KAAK,YAAY,KAAK,aAAa,KAAK;AAE3D,QAAI,KAAK,gBAAgB,KAAK,UAAU,SAAS,GAAG;AAClD,WAAK;AAAA,IACP;AAEA,WAAO,EAAE,MAAM,WAAW;AAAA,EAC5B;AAAA;AAAA,EAGA,QAAc;AACZ,SAAK,QAAQ,CAAC;AACd,SAAK,gBAAgB;AAAA,EACvB;AAAA,EAEA,aAAa,WAAqB,aAA8B;AAC9D,SAAK,YAAY;AACjB,SAAK,cAAc,eAAe,UAAU,IAAI,MAAM,UAAU;AAChE,SAAK,gBAAgB;AAAA,EACvB;AACF;AASA,SAAS,eAAe,QAA4B;AAClD,MAAI,QAAQ,SAAS;AACnB,UAAM,IAAI,aAAa,8BAA8B,YAAY;AAAA,EACnE;AACF;","names":["isObject","recordInferenceUsage","isObject","recordInferenceUsage","body"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@semiont/inference",
-  "version": "0.5.29",
+  "version": "0.5.31",
   "engines": {
     "node": ">=24.0.0"
   },
@@ -27,9 +27,10 @@
     "test:coverage": "vitest run --coverage"
   },
   "dependencies": {
-    "@anthropic-ai/sdk": "^0.120.0",
-    "@semiont/core": "0.5.29",
-    "@semiont/observability": "0.5.29"
+    "@anthropic-ai/sdk": "^0.123.0",
+    "@semiont/core": "0.5.31",
+    "@semiont/observability": "0.5.31",
+    "undici": "^7.29.0"
   },
   "devDependencies": {
     "@vitest/coverage-v8": "^4.1.11",