kestrel.markets/src/session/harness/ai-sdk-client.ts

A typed, token-efficient language + runtime for agentic trading: agents author bounded plans, the runtime fires them at the tick. CLI + typed library + MCP server.

/**
 * # session/harness/ai-sdk-client — the BYOK Vercel-AI-SDK {@link LlmClient} (the ONLY provider file)
 *
 * The concrete, wire-touching `LlmClient` behind {@link liveAgent} (ADR-0013 (c): the Vercel AI SDK
 * is the default BYOK driver). This is the ONE module in the tree that imports the provider packages
 * (`ai`, `@ai-sdk/amazon-bedrock`, `@ai-sdk/anthropic`, `@ai-sdk/google`) — `../simulate.ts`,
 * `../agent.ts`, `../index.ts`, and `src/index.ts` never reach it, so the deterministic core /
 * node-light root stay provider-free (ADR-0013 Consequences).
 *
 * ## BYOK, scrubbed at capture (ADR-0013 (e))
 * The key is supplied through the ENVIRONMENT only (`AWS_BEARER_TOKEN_BEDROCK` for Bedrock,
 * `ANTHROPIC_API_KEY` for the direct Anthropic path, `GEMINI_API_KEY`/`GOOGLE_GENERATIVE_AI_API_KEY`/
 * `GOOGLE_API_KEY` for the Google/Gemini path) and injected into the provider at construction.
 * It NEVER enters {@link AgentConfig} (content-hashed into `ConfigId`), and the wire evidence this
 * client captures is run through {@link scrubSecrets} — with every credential env value AND the
 * provided secrets — BEFORE the bytes are attached to the completion, so no artifact can carry a
 * credential (an m9i acceptance criterion, met structurally rather than by convention).
 */

import { generateText, wrapLanguageModel, type LanguageModel, type LanguageModelMiddleware, type ModelMessage } from "ai";
import { createAmazonBedrock } from "@ai-sdk/amazon-bedrock";
import { createAnthropic } from "@ai-sdk/anthropic";
import { createGoogleGenerativeAI } from "@ai-sdk/google";
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
import { createOpenAI } from "@ai-sdk/openai";
import { createAzure } from "@ai-sdk/azure";
import { cacheTtlNeverReachesWire, type AgentConfig, type CacheTtl, type LlmClient, type LlmCompletion } from "../agent.ts";
import { CREDENTIAL_ENV_VARS, scrubSecrets } from "./prompt.ts";
import { FIREWORKS_PROVIDER_NAME, decodingMiddleware, defaultDecodingFor, type DecodingConstraint } from "./decoding.ts";
import {
  VERTEX_MAAS_PROVIDER_NAME,
  gcloudAdcTokenSource,
  requireVertexMaasProject,
  resolveVertexMaasLocation,
  vertexMaasBaseUrl,
  type AdcExecFn,
  type AdcTokenSource,
} from "./vertex-maas.ts";

/** Which BYOK provider the client drives — the ENV name each reads its key from, never persisted:
 * `bedrock` → `AWS_BEARER_TOKEN_BEDROCK`; `anthropic` → `ANTHROPIC_API_KEY`; `google` →
 * `GEMINI_API_KEY` / `GOOGLE_GENERATIVE_AI_API_KEY` / `GOOGLE_API_KEY` (first present wins);
 * `gateway` → `AI_GATEWAY_API_KEY` (the Vercel AI Gateway, an OpenAI-compatible surface — the CASH
 * lane, billed per token, so reserved for lots of small cheap runs or models with no credit lane);
 * `local` → `LOCAL_LLM_BASE_URL` (a local mlx_lm.server / vLLM / Ollama serving open-weight watchers —
 * the ZERO-CASH, CFG-capable lane for the large watcher fleet; `apiKey` optional);
 * `openai` → `OPENAI_API_KEY` (the direct OpenAI API — a CASH lane; constrained by strict JSON Schema,
 * NOT GBNF, which OpenAI cannot take);
 * `azure` → `AZURE_OPENAI_ENDPOINT` + `AZURE_OPENAI_API_KEY` + `AZURE_OPENAI_API_VERSION` (the same GPT
 * models on pre-committed CREDITS — so a LARGE fan-out of a GPT watcher rides Azure, and only a SMALL
 * spot-check rides cash OpenAI; same strict-JSON-Schema constraint);
 * `fireworks` → `FIREWORKS_API_KEY` (an OpenAI-compatible surface for the BIG OPEN weights — DeepSeek,
 * Llama-405B, Qwen — a CASH lane, but the only hosted one that takes **GBNF** grammar-constrained
 * decoding, so it is where an open model gets the same emission floor the local lane gets);
 * `vertex-maas` → NO env key: an **ADC bearer** refreshed via `gcloud auth application-default
 * print-access-token` (Vertex Model Garden MaaS — DeepSeek V3.2 / Qwen3-Next-80B / Kimi K2 Thinking on
 * GCP CREDITS, over the OpenAI-compatible `locations/global` endpoint; see `./vertex-maas.ts` for the
 * region trap and the dead-ADC doctrine, kestrel-k8he).
 *
 * `codex-cli` is the ZERO-CASH ChatGPT-SUBSCRIPTION lane (`gpt-5.6-sol` and the other GPT variants). It is
 * NOT an AI-SDK provider and is NOT built here: no OpenAI-compatible HTTP endpoint accepts a ChatGPT
 * subscription OAuth token, so it is served by SHELLING OUT to the Codex CLI (`codex exec`) in
 * `./codex-cli-client.ts`. It is a member of this union so the whole harness can speak ONE provider
 * vocabulary; {@link harnessLlmClient} in `./harness-client.ts` is the dispatcher that routes it there, and
 * {@link buildModel} FAILS CLOSED if it ever reaches the SDK path.
 *
 * There is deliberately NO `subagent` member. A Claude Code subagent is ONE-SHOT — spawned, it works, it
 * returns text — so it can author a plan DOC but can never serve as a per-turn client inside the harness's
 * multi-turn loop. The `subagent` LANE in {@link roster} is therefore an AUTHORING lane only (see
 * `scripts/bench/author-cells.ts`), not a provider. Multi-turn watching still requires a real API endpoint.
 *
 * See {@link roster} (cost-tier policy) and {@link resolveLane} for which lane a given model routes to at
 * each run size, and `./decoding.ts` for which constrained-decoding encoding each lane can honor. */
export type HarnessProvider =
  | "bedrock"
  | "anthropic"
  | "google"
  | "gateway"
  | "local"
  | "codex-cli"
  | "openai"
  | "azure"
  | "fireworks"
  | "vertex-maas";

/** Fireworks' OpenAI-compatible inference endpoint. The lane that serves the BIG open weights AND
 * accepts a GBNF grammar in `response_format` (see `./decoding.ts`). */
export const FIREWORKS_BASE_URL = "https://api.fireworks.ai/inference/v1";

/** The Azure OpenAI `api-version` used when `AZURE_OPENAI_API_VERSION` is unset. Azure pins its request
 * surface per api-version; an older one may silently DROP the `pattern` keyword from a strict schema
 * (see the L4 note in `artifacts/grammar/watcher.schema.json`), which is why this is overridable. */
export const AZURE_API_VERSION_DEFAULT = "2024-10-21";

/** The Vercel AI Gateway OpenAI-compatible base URL (the CASH lane). Model ids ride the
 * provider-prefixed form, e.g. `anthropic/claude-haiku-4.5`, `openai/gpt-5.6-sol`. */
export const GATEWAY_BASE_URL = "https://ai-gateway.vercel.sh/v1";

/** The gateway skip-cache header. MUST be sent `true` on EVERY gateway request — else the AI Gateway may
 * serve n replicates ONE cached completion, fabricating a σ=0 (a contamination-class bug: a cache artifact
 * that masquerades as measured determinism). Set at gateway-client construction so no call site can forget
 * it. Matches the hosted adapter's convention (scripts/bench/hosted/gateway-lane.ts). */
export const GATEWAY_SKIP_CACHE_HEADER = "cf-aig-skip-cache";

/** The default local OpenAI-compatible base URL when `LOCAL_LLM_BASE_URL` is unset — the port an
 * `mlx_lm.server` / vLLM / Ollama-compatible server listens on for the zero-cash watcher fleet. */
export const LOCAL_BASE_URL_DEFAULT = "http://localhost:8080/v1";

/** The Vercel AI Gateway rejects `max_tokens` below this with HTTP 400; the gateway lane floors its
 * output ceiling here so a tiny watcher budget can't turn a live call into a 400 (VERIFIED LIVE). */
export const GATEWAY_MIN_OUTPUT_TOKENS = 16;

/** A sane default reasoning budget (in output tokens) for a REASONING model on the gateway lane
 * (kestrel-a1f). Fable 5 is a thinking-always-on model: over the Vercel gateway's OpenAI-compatible
 * surface its reasoning/thinking tokens are drawn from the SAME `max_tokens` budget as the visible
 * reply. So a small authoring ceiling (the 1024 that shipped) is spent ENTIRELY on reasoning and the
 * visible plan text comes back EMPTY (finish_reason=length, no supersede). The gateway lane widens
 * the wire ceiling by the caller's `reasoningHeadroomTokens` so the requested VISIBLE-output budget
 * still survives after the model finishes thinking. This constant is a good default headroom for a
 * bounded authoring turn; callers pass it (or their own value) via {@link AiSdkClientOptions}. */
export const GATEWAY_REASONING_HEADROOM_TOKENS = 12000;

/** The slice of the AI-SDK `generateText` result this client reads. Declared so the temperature-capability
 * path can be exercised offline through {@link AiSdkClientOptions.generate} (no network, no provider SDK).
 * `inputTokenDetails.cacheRead/cacheWriteTokens` are the AI SDK's flattened provider-native prompt-cache
 * counters (Anthropic `cache_read/cache_creation_input_tokens`, Bedrock `cacheRead/cacheWriteInputTokens`) —
 * the on-record source of the m9i.5 cache economics (kestrel-rul). */
export interface GenerateResult {
  readonly text: string;
  readonly reasoningText?: string;
  readonly usage: {
    readonly inputTokens?: number;
    readonly outputTokens?: number;
    readonly inputTokenDetails?: {
      readonly cacheReadTokens?: number;
      readonly cacheWriteTokens?: number;
      readonly noCacheTokens?: number;
    };
    readonly outputTokenDetails?: { readonly reasoningTokens?: number };
  };
  readonly request?: unknown;
  readonly response?: { readonly id?: string; readonly modelId?: string; readonly headers?: unknown };
  readonly providerMetadata?: unknown;
  readonly finishReason?: unknown;
  /** The AI SDK's call warnings. Read for ONE purpose: a provider that SILENTLY DROPS a setting reports it
   * here rather than throwing (see {@link droppedTemperature}). Two shapes exist across SDK generations:
   * the older `{ type: "unsupported-setting", setting }` and ai v7's `{ type: "unsupported", feature }`
   * (`src/logger/log-warnings.ts` formats the latter — the "The feature X is not supported" console line). */
  readonly warnings?: readonly { readonly type?: string; readonly setting?: string; readonly feature?: string; readonly details?: string }[];
}

/** Did the provider DROP `temperature` rather than honor it? The OpenAI reasoning models (the whole gpt-5
 * family) do not accept `temperature`, and the AI SDK does NOT throw for them — it strips the field and
 * emits an `unsupported-setting` warning, so the call succeeds having silently used the model's default.
 *
 * That is a quiet lie in the evidence, and exactly the class of thing kestrel-gvx exists to stop: without
 * this check the captured wire record would report `temperature: 0` for a call whose temperature field
 * never went on the wire, and a reader would credit a greedy-decoding result to a run that never asked for
 * one. Detecting the warning lets `complete` record `"model-default"` — what ACTUALLY happened.
 *
 * (Distinct from {@link rejectsTemperature}, which handles the providers — the newest Bedrock Sonnet — that
 * ERROR on the field instead of dropping it. Same lie, two different provider behaviors.)
 *
 * BOTH warning shapes are matched: the older `unsupported-setting`/`setting` pair AND ai v7's
 * `unsupported`/`feature` pair. The installed ai v7 emits ONLY the latter — caught live on the azure lane
 * (azure-foundry-sweep v1, 2026-07-18): gpt-5.4 deployments dropped `temperature` with a v7-shape warning,
 * the old matcher missed it, and 24/24 rows' wire evidence claimed `temperature: 1` for a field that never
 * rode the wire — the exact kestrel-gvx lie this function exists to stop, inert for one shipped generation
 * because its fixture pinned the synthetic old shape rather than what the real SDK emits. */
export function droppedTemperature(r: GenerateResult): boolean {
  return (r.warnings ?? []).some(
    (w) =>
      (w.type === "unsupported-setting" && w.setting === "temperature") ||
      (w.type === "unsupported" && w.feature === "temperature"),
  );
}

/** One AI-SDK conversation `ModelMessage` this client sends. Only `user`/`assistant` ride `messages`;
 * the cached SYSTEM prefix does NOT (AI SDK v7 rejects a leading `system`-role message in `messages`:
 * *"System messages are not allowed in the prompt or messages fields. Use the instructions option
 * instead"*) — it rides the top-level `system`/`instructions` field as an {@link AiSdkSystemMessage}.
 * `providerOptions` carries the provider-native prompt-cache breakpoint (Anthropic `cacheControl` /
 * Bedrock `cachePoint`) on the conversation message the harness marked cacheable. */
export interface AiSdkMessage {
  readonly role: "user" | "assistant";
  readonly content: string;
  readonly providerOptions?: Record<string, unknown>;
}

/** The cached SYSTEM prefix as an AI-SDK `SystemModelMessage` — the shape the top-level
 * `system`/`instructions` option accepts (`Instructions = string | SystemModelMessage | SystemModelMessage[]`
 * in ai v7), so the byte-stable system prefix can carry the provider-native cache breakpoint WITHOUT being a
 * leading `system`-role message inside `messages` (which the v7 core rejects outright). Both the Bedrock and
 * Anthropic providers fold this into their top-level system field with the cache point attached. */
export interface AiSdkSystemMessage {
  readonly role: "system";
  readonly content: string;
  readonly providerOptions?: Record<string, unknown>;
}

/** The `generateText` args this client passes (a subset). `temperature` is OMITTED for a model that rejects
 * the field outright — see {@link aiSdkLlmClient}'s capability-aware `complete`. `system` is ALWAYS set: a
 * plain string for the uncached path (and for Google, whose caching is implicit — no breakpoint), or an
 * {@link AiSdkSystemMessage} carrying the cache breakpoint when the byte-stable system prefix is cached. */
export interface GenerateArgs {
  readonly model: LanguageModel;
  readonly system?: string | AiSdkSystemMessage;
  readonly messages: readonly AiSdkMessage[];
  readonly temperature?: number;
  readonly maxOutputTokens: number;
  readonly maxRetries: number;
  /** Top-level provider-native `providerOptions` for the whole request — currently the extended-thinking
   * config from {@link thinkingProviderOptions} (`{ anthropic: { thinking } }` / `{ bedrock: { reasoningConfig } }`
   * / `{ google: { thinkingConfig } }`). OMITTED (undefined) when `thinkingLevel:"none"`, so a no-thinking
   * request is byte-identical to before this knob was wired. Distinct from the per-MESSAGE `providerOptions`
   * on {@link AiSdkMessage}/{@link AiSdkSystemMessage}, which carry the prompt-cache breakpoint. */
  readonly providerOptions?: Record<string, unknown>;
}

/** The provider-native prompt-cache breakpoint for `provider`, as AI-SDK `providerOptions` — or `undefined`
 * when the provider has NO explicit breakpoint mechanism. Anthropic marks an `ephemeral` `cacheControl`
 * block; Bedrock inserts a `cachePoint`; **Google/Gemini returns `undefined`** — Gemini has no per-message
 * cache-control primitive in this SDK version (implicit context caching is automatic and its cache-read
 * tokens still flow through {@link GenerateResult} usage; explicit `cachedContent` needs an out-of-band
 * cache-create call, so it is NOT wired here — we do not fake a breakpoint the provider cannot honor).
 * Applied to the last content part of the message it rides (the AI SDK falls back to message-level
 * `providerOptions` for the final part).
 *
 * `ttl` (kestrel-wa0j.1) opts the breakpoint into a non-default cache TTL. `undefined` ⇒ the provider
 * default and BYTE-IDENTICAL breakpoint objects to before the axis existed (no `ttl` key at all). Both
 * native lanes take it (verified against the installed SDKs: `@ai-sdk/anthropic` 4.0.12
 * `cacheControl.ttl: "5m"|"1h"`; `@ai-sdk/amazon-bedrock` 5.0.17 `cachePoint.ttl: "5m"|"1h"`). A TTL on a
 * lane with NO breakpoint primitive THROWS — the ConfigId claims a TTL, and returning `undefined` would
 * silently drop a configured knob (fail-closed, never silently inert). */
export function cacheBreakpoint(provider: HarnessProvider, ttl?: CacheTtl): Record<string, unknown> | undefined {
  switch (provider) {
    case "bedrock":
      return { bedrock: { cachePoint: { type: "default", ...(ttl !== undefined ? { ttl } : {}) } } };
    case "anthropic":
      return { anthropic: { cacheControl: { type: "ephemeral", ...(ttl !== undefined ? { ttl } : {}) } } };
    case "google":
    case "gateway":
    case "local":
    case "openai":
    case "azure":
    case "fireworks":
    case "vertex-maas":
    case "codex-cli":
      // No explicit per-message cache-control primitive on these lanes: Gemini caches implicitly (no
      // breakpoint); the OpenAI-compatible surface caches automatically on a stable prefix; the Codex
      // CLI's caching is provider-internal. We do not fake a breakpoint the wire cannot honor — and a
      // configured TTL these lanes cannot express must FAIL, not silently ride a ConfigId column that
      // claims a cache TTL the provider never saw (kestrel-gvx: the evidence never lies about the wire).
      if (ttl !== undefined) {
        throw new Error(
          cacheTtlNeverReachesWire(
            ttl,
            `provider "${provider}" has no explicit prompt-cache breakpoint to carry a TTL`,
            `Use provider "anthropic" or "bedrock", or drop cacheTtl.`,
          ),
        );
      }
      return undefined;
  }
}

/** The strategist/watcher reasoning-effort dial, mirrored from {@link AgentConfig.thinkingLevel} (it rides
 * the ConfigId, so it is part of run identity). `none` is the default EVERYWHERE (author-cells.ts included). */
export type ThinkingLevel = AgentConfig["thinkingLevel"];

/** LEVEL → extended-thinking budget, in *thinking* tokens (SEPARATE from the visible-reply `maxOutputTokens`
 * ceiling — see {@link aiSdkLlmClient}, which raises the wire ceiling ABOVE this so `max_tokens > budget`).
 *
 * `none` → 0 ⇒ thinking is DISABLED: NO `providerOptions.*.thinking` key rides the wire at all, so a `none`
 * request is BYTE-IDENTICAL to before this knob was wired (the additive-opt-in guarantee). `low` starts at
 * 1024 — the Anthropic API's HARD MINIMUM `budget_tokens` (a value below 1024 is rejected), so it is the
 * smallest budget that can legally enable thinking. The rest step up geometrically to give the reasoning-level
 * sweep (task#19) real separation without crossing the 32k soft-ceiling where Anthropic recommends streaming
 * to avoid long-request timeouts: `medium` 4096, `high` 12000 (matches {@link GATEWAY_REASONING_HEADROOM_TOKENS},
 * the empirically-sufficient authoring headroom from kestrel-a1f), `max` 24000. */
export const THINKING_BUDGET: Record<ThinkingLevel, number> = {
  none: 0,
  low: 1024,
  medium: 4096,
  high: 12000,
  max: 24000,
};

/** LEVEL → the OpenAI-surface `reasoning_effort` enum for the gpt/open lanes (`openai`/`azure`/`fireworks`),
 * which express effort as an ENUM, not a token budget (kestrel-0drf — this mapping used to be a documented
 * no-op, so every "gpt-5 at high thinking" row compared the seat against itself).
 *
 * The 5→3 granularity decision, explicit: `reasoning_effort` is only guaranteed THREE values across every
 * OpenAI reasoning generation (o-series: low/medium/high; gpt-5 adds `minimal`; newer snapshots add more),
 * so `max` CLAMPS to `high` — the strongest value every generation accepts. A newer-only top value
 * (`xhigh`) would turn the dial's top stop into a wire error on a model that would have honored `high` —
 * a knob that fails exactly where it should bite hardest. `none` never reaches this table: no key rides
 * the wire at all (the additive-opt-in guarantee), so a `none` request stays byte-identical to before the
 * dial was wired. (The codex-cli lane has its own 5→5 mapping — `codexReasoningEffort` in
 * `./codex-cli-client.ts` — because the CLI's dial does reach `ultra`.) */
export const REASONING_EFFORT: Record<Exclude<ThinkingLevel, "none">, "low" | "medium" | "high"> = {
  low: "low",
  medium: "medium",
  high: "high",
  max: "high", // clamp — see the doc above
};

/** Map a {@link ThinkingLevel} to the provider-native extended-thinking/effort `providerOptions` for
 * `provider`, or `undefined` when NO thinking config should ride the wire — i.e. `none` (thinking disabled ⇒
 * byte-identical to today), OR a lane with no native per-request primitive (`gateway`/`local`: the
 * OpenAI-compatible surface shares one `max_tokens` budget and is handled by the separate gateway
 * `reasoningHeadroomTokens` mechanism; `codex-cli`: the subprocess lane maps the SAME dial onto the CLI's
 * `model_reasoning_effort` flag in `./codex-cli-client.ts`, never through providerOptions). The wired lanes
 * each take a different key, VERIFIED against the installed SDKs (`@ai-sdk/anthropic`+`@ai-sdk/google`
 * 4.0.12, `@ai-sdk/amazon-bedrock` 5.0.17, `@ai-sdk/openai`+`@ai-sdk/azure` 4.0.14,
 * `@ai-sdk/openai-compatible` 3.0.9):
 *   - anthropic → `{ anthropic: { thinking: { type: "enabled", budgetTokens } } }`  (AnthropicProviderOptions.thinking)
 *   - bedrock   → `{ bedrock: { reasoningConfig: { type: "enabled", budgetTokens } } }`  (BedrockProviderOptions.reasoningConfig)
 *   - google    → `{ google: { thinkingConfig: { thinkingBudget } } }`  (GoogleGenerativeAIProviderOptions.thinkingConfig)
 *   - openai/azure → `{ openai: { reasoningEffort } }`  (both serve @ai-sdk/openai's Chat Completions model,
 *     which parses providerOptions under the `openai` key — azure.chat included — and maps `reasoningEffort`
 *     onto the `reasoning_effort` request field; see {@link REASONING_EFFORT} for the 5→3 clamp)
 *   - fireworks → `{ fireworks: { reasoningEffort } }`  (the openai-compatible model reads providerOptions
 *     under its provider name and maps `reasoningEffort` → `reasoning_effort` on the body)
 */
export function thinkingProviderOptions(
  provider: HarnessProvider,
  level: ThinkingLevel,
): Record<string, unknown> | undefined {
  if (level === "none") return undefined; // thinking disabled — no key on the wire (byte-identical to today)
  const budgetTokens = THINKING_BUDGET[level];
  switch (provider) {
    case "anthropic":
      return { anthropic: { thinking: { type: "enabled", budgetTokens } } };
    case "bedrock":
      return { bedrock: { reasoningConfig: { type: "enabled", budgetTokens } } };
    case "google":
      return { google: { thinkingConfig: { thinkingBudget: budgetTokens } } };
    case "gateway":
    case "local":
    case "codex-cli":
    case "vertex-maas":
      // No native per-request thinking-budget primitive on these lanes — thinkingLevel is a no-op here (the
      // gateway lane uses `reasoningHeadroomTokens` instead; codex-cli maps effort in ./codex-cli-client.ts).
      // vertex-maas: the Model Garden OpenAI-compatible surface documents no `reasoning_effort` field —
      // UNVERIFIED against the live endpoint, so nothing is sent rather than a field the wire may drop
      // silently (kestrel-gvx: the evidence never claims a knob the provider never saw). Follow-up if the
      // MaaS surface is probed to honor it.
      return undefined;
    case "openai":
    case "azure":
      // The OpenAI reasoning family expresses effort as `reasoning_effort` (an enum, not a token budget).
      // Both lanes serve it through @ai-sdk/openai's Chat Completions model (the azure provider constructs
      // OpenAIChatLanguageModel with provider "azure.chat"), which parses providerOptions under the
      // `openai` key and puts `reasoningEffort` on the body as `reasoning_effort` (verified @ai-sdk/openai
      // 4.0.14). This was a documented no-op until kestrel-0drf — every "gpt at thinkingLevel X" row ran
      // at the provider-default effort, comparing the seat against itself.
      return { openai: { reasoningEffort: REASONING_EFFORT[level] } };
    case "fireworks":
      // Fireworks' OpenAI-compatible surface takes the same `reasoning_effort` request field; the
      // openai-compatible chat model reads providerOptions under its provider name (`fireworks`) and maps
      // `reasoningEffort` onto the body (verified @ai-sdk/openai-compatible 3.0.9). Same clamped mapping.
      return { [FIREWORKS_PROVIDER_NAME]: { reasoningEffort: REASONING_EFFORT[level] } };
  }
}

/** The injectable provider call — the real one wraps `generateText`; a test passes a stand-in. */
export type GenerateFn = (args: GenerateArgs) => Promise<GenerateResult>;

/** Does this provider error mean the model REJECTS the `temperature` field outright, rather than a transient
 * failure? The newest Bedrock Sonnet generation (`us.anthropic.claude-sonnet-5`) raises "`temperature` is
 * deprecated for this model" even at 0. Matched narrowly so ANY other error still fails the cell explicitly
 * (m9i.1 — never a blanket retry that masks a real provider failure). */
export function rejectsTemperature(e: unknown): boolean {
  const msg = (e instanceof Error ? e.message : String(e)).toLowerCase();
  return msg.includes("temperature") && /deprecat|not support|unsupported|not allowed|unexpected|invalid|unknown/.test(msg);
}

export interface AiSdkClientOptions {
  readonly provider: HarnessProvider;
  /** The provider-native model id (e.g. a Bedrock inference-profile id, or an Anthropic model id). */
  readonly modelId: string;
  /** AWS region (Bedrock: defaults to `AWS_REGION`/`AWS_DEFAULT_REGION`, else `us-east-1`). On
   * `vertex-maas` this is the LOCATION override and only `"global"` is legal — any other value THROWS
   * the clear region-trap error at construction (us-central1 rejects all MaaS slugs; never a silent
   * rewrite — see `./vertex-maas.ts`). Ignored on every other lane. */
  readonly region?: string;
  /** A hard output-token ceiling so a runaway reply can't blow the budget (default 2048). */
  readonly maxOutputTokens?: number;
  /** GATEWAY-ONLY reasoning headroom, in output tokens (default 0 — no change for any other lane, and
   * no change for gateway callers that don't set it). For a REASONING model reached via the Vercel
   * gateway (Fable 5: thinking always on), the OpenAI-compatible `max_tokens` budget is shared between
   * the reasoning/thinking trace and the visible reply — so a small `maxOutputTokens` is consumed by
   * reasoning alone and the visible text returns EMPTY (kestrel-a1f). Setting this WIDENS the wire
   * ceiling to `maxOutputTokens + reasoningHeadroomTokens`, reserving `maxOutputTokens` for the visible
   * output after the model finishes thinking. Ignored off the gateway lane (Bedrock/Anthropic/Google
   * separate their thinking budget natively; local carries no reasoning budget here). See
   * {@link GATEWAY_REASONING_HEADROOM_TOKENS} for a sane default a bounded authoring turn can pass. */
  readonly reasoningHeadroomTokens?: number;
  /** Extra secret strings to scrub from wire evidence, beyond the credential env vars (belt-and-suspenders). */
  readonly secrets?: readonly string[];
  /** Capture scrubbed wire evidence onto each completion (default true). */
  readonly captureWire?: boolean;
  /** Test seam ONLY: a stand-in for the provider `generateText` (defaults to the real one). Exists so the
   * temperature-capability logic can be exercised offline; NEVER set on a production client. */
  readonly generate?: GenerateFn;
  /**
   * CONSTRAINED DECODING. `"auto"` (the default) gives each lane the encoding it can honor — GBNF on
   * `fireworks`, strict JSON Schema on `openai`/`azure`, nothing on the lanes with no lever (see
   * `./decoding.ts`). `"off"` sends an unconstrained request. An explicit {@link DecodingConstraint}
   * overrides both (a specialized, frame-kernel-bound grammar, say).
   *
   * The default is `"auto"` and not `"off"` on purpose: a runner that adds a lane should get the emission
   * floor the benchmark's claims depend on WITHOUT having to know that it must ask for it. Forgetting to
   * ask would not fail loudly — it would just quietly produce a worse α and be misread as judgment.
   */
  readonly decoding?: DecodingConstraint | "auto" | "off";
  /** TEST SEAM ONLY: a stand-in `fetch` for the provider. The offline lane tests pass one that CAPTURES
   * the serialized request body (proving the grammar/schema really is attached to the outbound payload)
   * and returns a canned response — so the assertion is made against the real provider-package code path
   * with ZERO network calls. NEVER set on a production client. */
  readonly fetch?: typeof globalThis.fetch;
  /** `vertex-maas` ONLY test seam: a stand-in for the `gcloud auth application-default
   * print-access-token` subprocess the ADC token source spawns (same pattern as `codexExec`).
   * NEVER set on a production client. */
  readonly adcExec?: AdcExecFn;
  /** `vertex-maas` ONLY: a pre-built ADC token source (test seam / shared cache across clients).
   * Defaults to `gcloudAdcTokenSource({ exec: adcExec })`. */
  readonly adcTokenSource?: AdcTokenSource;
}

/** Gather the concrete credential values present in the environment (never logged, only fed to the
 * scrubber) plus any explicitly-provided secrets, so ANY occurrence in wire evidence is redacted by
 * identity regardless of framing. */
function secretsToScrub(extra: readonly string[] = []): string[] {
  const env: string[] = [];
  for (const name of CREDENTIAL_ENV_VARS) {
    const v = process.env[name];
    if (typeof v === "string" && v.trim().length >= 6) env.push(v);
  }
  return [...env, ...extra];
}

/** The Google/Gemini API key, read from the environment ONLY (never persisted, scrubbed at capture). The
 * SDK's own default env name is `GOOGLE_GENERATIVE_AI_API_KEY`; we also honor `GEMINI_API_KEY` and
 * `GOOGLE_API_KEY` (the overnight `.dev.vars` convention) so the same key material works whatever the shell
 * calls it — first present, non-trivial value wins. Returns `undefined` when none is set (the provider then
 * falls back to its own default lookup, still env-only). */
function googleApiKeyFromEnv(): string | undefined {
  for (const name of ["GEMINI_API_KEY", "GOOGLE_GENERATIVE_AI_API_KEY", "GOOGLE_API_KEY"] as const) {
    const v = process.env[name];
    if (typeof v === "string" && v.trim().length >= 6) return v;
  }
  return undefined;
}

/** The constraint this client will actually send: the caller's explicit one, or the lane's default under
 * `"auto"`, or none under `"off"`. Exported so a runner can LOG what it is about to constrain with (and
 * so the offline tests can assert the routing without constructing a provider). */
export function resolveDecoding(opts: AiSdkClientOptions): DecodingConstraint | undefined {
  const d = opts.decoding ?? "auto";
  if (d === "off") return undefined;
  if (d === "auto") return defaultDecodingFor(opts.provider);
  return d;
}

/** The AI SDK's `LanguageModel` is `string | LanguageModelV2|V3|V4` — the string form is the registry
 * shorthand (`"openai/gpt-5"`), which this harness never uses (it constructs providers explicitly). Narrow
 * it away so the model can be handed to `wrapLanguageModel`, which takes only the object form. */
type LanguageModelObject = Exclude<LanguageModel, string>;

/** The base (unconstrained) model for a lane. Split out from {@link buildModel} so the constraint
 * wrapper is applied in exactly ONE place and cannot be forgotten on a new lane. */
function buildBaseModel(opts: AiSdkClientOptions): LanguageModelObject {
  if (opts.provider === "codex-cli") {
    // FAIL-CLOSED. `codex-cli` is a SUBPROCESS lane (`./codex-cli-client.ts`), not an AI-SDK provider — there
    // is no OpenAI-compatible endpoint that accepts a ChatGPT-subscription OAuth token. Reaching here means a
    // caller bypassed the `harnessLlmClient` dispatcher; silently constructing an SDK model would send the
    // call to a CASH gateway instead of the FREE subscription — precisely the cost regression this lane
    // exists to prevent. Route through `harnessLlmClient` (`./harness-client.ts`) instead.
    throw new Error(
      `provider "codex-cli" is not an AI-SDK provider — it is the ChatGPT-subscription SUBPROCESS lane. ` +
        `Construct it via harnessLlmClient() from ./harness-client.ts (which routes to codexCliClient), ` +
        `never aiSdkLlmClient() (fail-closed: a silent SDK fallback would bill CASH on the gateway).`,
    );
  }
  if (opts.provider === "bedrock") {
    const region = opts.region ?? process.env.AWS_REGION ?? process.env.AWS_DEFAULT_REGION ?? "us-east-1";
    // apiKey defaults to AWS_BEARER_TOKEN_BEDROCK from the environment (never passed as a literal here).
    const bedrock = createAmazonBedrock({ region });
    return bedrock(opts.modelId);
  }
  if (opts.provider === "google") {
    // Google/Gemini path. The key is read from the environment (never a config literal) and scrubbed from
    // wire evidence like every other credential. We pass it explicitly ONLY because the harness accepts
    // several env aliases the SDK's default lookup (GOOGLE_GENERATIVE_AI_API_KEY) does not — the VALUE is
    // still env-sourced and never persisted. Absent ⇒ let the provider do its own (env-only) default lookup.
    const apiKey = googleApiKeyFromEnv();
    const google = createGoogleGenerativeAI(apiKey !== undefined ? { apiKey } : {});
    return google(opts.modelId);
  }
  if (opts.provider === "gateway") {
    // Vercel AI Gateway — an OpenAI-compatible surface (the CASH lane). The key is env-only
    // (`AI_GATEWAY_API_KEY`, scrubbed at capture like every credential); absent ⇒ the provider sends
    // no Authorization header and the gateway 401s honestly (never a config literal). Model ids ride
    // the provider-prefixed form (e.g. `anthropic/claude-haiku-4.5`). The min-16 output floor is
    // enforced by {@link aiSdkLlmClient}, not here (buildModel carries no per-call budget).
    const apiKey = process.env.AI_GATEWAY_API_KEY;
    const gateway = createOpenAICompatible({
      name: "gateway",
      baseURL: GATEWAY_BASE_URL,
      ...(apiKey !== undefined ? { apiKey } : {}),
      // cf-aig-skip-cache: true on EVERY gateway request (NON-OPTIONAL). The Vercel/Cloudflare AI Gateway
      // caches by request body; WITHOUT this header, n replicates of the same frame can silently collapse
      // into ONE cached completion — a contamination-class bug that FABRICATES a zero-variance (σ=0) result
      // that reads as measured determinism but is a cache artifact. Setting it here (at construction) means
      // it rides every call the gateway client ever makes and no call site can forget it. Mirrors the
      // hosted eval adapter's `gatewayHeaders` convention (scripts/bench/hosted/gateway-lane.ts, GATEWAY_
      // SKIP_CACHE_HEADER). Any other lane is unaffected (this header is gateway-only).
      headers: { [GATEWAY_SKIP_CACHE_HEADER]: "true" },
    });
    return gateway(opts.modelId);
  }
  if (opts.provider === "local") {
    // Local OpenAI-compatible server (mlx_lm.server / vLLM / Ollama) — the ZERO-CASH, CFG-capable lane
    // for the open-weight watcher fleet. Base URL from `LOCAL_LLM_BASE_URL` (default localhost:8080/v1);
    // apiKey is optional (a local server usually ignores it) — a dummy keeps the SDK happy if one leaks in.
    const baseURL = process.env.LOCAL_LLM_BASE_URL ?? LOCAL_BASE_URL_DEFAULT;
    const apiKey = process.env.LOCAL_LLM_API_KEY ?? "local";
    const local = createOpenAICompatible({ name: "local", baseURL, apiKey, ...(opts.fetch !== undefined ? { fetch: opts.fetch } : {}) });
    return local(opts.modelId);
  }
  if (opts.provider === "openai") {
    // Direct OpenAI — the CASH lane. `apiKey` defaults to OPENAI_API_KEY from the ENVIRONMENT (never a
    // config literal, never persisted, scrubbed from wire evidence like every other credential). Absent ⇒
    // the SDK raises its own "API key is missing" at call time, which is the honest failure.
    //
    // `.chat()` — CHAT COMPLETIONS, explicitly, not the (default) Responses API. Three reasons:
    //   1. The strict schema rides `response_format.json_schema` on Chat Completions but `text.format` on
    //      Responses. Pinning Chat Completions gives all three new lanes (openai / azure / fireworks) ONE
    //      wire contract, so `./decoding.ts` has one thing to attach and the tests one thing to assert.
    //   2. Azure serves Chat Completions on every api-version; its Responses surface is api-version-gated
    //      and 404s on an older one — and the SAME schema must reach both lanes or the comparison is moot.
    //   3. The watcher is text-in/text-out with a schema. Nothing in Responses (statefulness, built-in
    //      tools, reasoning-item plumbing) is used here, so its surface area is pure risk.
    const openai = createOpenAI({ ...(opts.fetch !== undefined ? { fetch: opts.fetch } : {}) });
    return openai.chat(opts.modelId);
  }
  if (opts.provider === "azure") {
    // Azure OpenAI — the same GPT models on pre-committed CREDITS (so a LARGE fan-out routes here and only
    // a SMALL spot-check pays cash on the direct OpenAI lane — see `./roster.ts`).
    //
    // `AZURE_OPENAI_ENDPOINT` is the resource endpoint (e.g. https://my-resource.openai.azure.com). We pass
    // it as `baseURL` rather than deriving a `resourceName`, because a private/proxied endpoint is common
    // and a resourceName cannot express one. `opts.modelId` is the **DEPLOYMENT NAME**, not the OpenAI model
    // id — on Azure the deployment is the addressable unit, and `useDeploymentBasedUrls` puts it in the path.
    const endpoint = process.env.AZURE_OPENAI_ENDPOINT;
    const apiVersion = process.env.AZURE_OPENAI_API_VERSION ?? AZURE_API_VERSION_DEFAULT;
    const apiKey = process.env.AZURE_OPENAI_API_KEY;
    const azure = createAzure({
      ...(endpoint !== undefined && endpoint.trim() !== "" ? { baseURL: `${endpoint.replace(/\/+$/, "")}/openai` } : {}),
      ...(apiKey !== undefined ? { apiKey } : {}),
      apiVersion,
      useDeploymentBasedUrls: true,
      ...(opts.fetch !== undefined ? { fetch: opts.fetch } : {}),
    });
    return azure.chat(opts.modelId); //  Chat Completions — see the `openai` branch for why, in full
  }
  if (opts.provider === "vertex-maas") {
    // Vertex Model Garden MaaS — DeepSeek V3.2 / Qwen3-Next-80B / Kimi K2 Thinking on GCP CREDITS, over
    // the OpenAI-compatible endpoint (kestrel-k8he; live probe wf_e58255fd-e01). Two fail-closed guards
    // fire HERE, at construction, before anything can reach the wire:
    //   - PROJECT: the endpoint URL embeds the GCP project id; absent env project ⇒ throw naming the
    //     env vars (a blank project could only 404 opaquely).
    //   - THE REGION TRAP: only `locations/global` serves MaaS — us-central1 rejects every slug. A
    //     configured region (opts.region or a GOOGLE_VERTEX_LOCATION/GOOGLE_CLOUD_LOCATION in the
    //     shell) that is not "global" throws the clear region error; NEVER a silent rewrite to global.
    const project = requireVertexMaasProject();
    const location = resolveVertexMaasLocation(opts.region);
    // AUTH IS AN ADC BEARER, NOT AN API KEY. Every outbound call gets `Authorization: Bearer <token>`
    // attached by this fetch wrapper, with the token refreshed via `gcloud auth application-default
    // print-access-token` (cached ~45 min; a DEAD ADC throws AdcAuthError naming the exact
    // `gcloud auth application-default login` reauth command — surface the login, never silently skip
    // the lane). The wrapper sets the header LAST so it rides the request whatever headers the SDK
    // built; the injected test fetch (if any) sits UNDERNEATH the wrapper, so the offline fixtures
    // exercise the real bearer-attachment path (delete this attachment ⇒ the fixture goes red).
    const tokenSource = opts.adcTokenSource ?? gcloudAdcTokenSource({ ...(opts.adcExec !== undefined ? { exec: opts.adcExec } : {}) });
    const baseFetch = opts.fetch ?? globalThis.fetch;
    const fetchWithAdcBearer = (async (input: Parameters<typeof globalThis.fetch>[0], init?: RequestInit) => {
      const token = await tokenSource.token();
      const headers = new Headers(init?.headers);
      headers.set("authorization", `Bearer ${token}`);
      return baseFetch(input, { ...init, headers });
    }) as typeof globalThis.fetch;
    const maas = createOpenAICompatible({
      name: VERTEX_MAAS_PROVIDER_NAME,
      baseURL: vertexMaasBaseUrl(project, location),
      fetch: fetchWithAdcBearer,
    });
    return maas(opts.modelId);
  }
  if (opts.provider === "fireworks") {
    // Fireworks — an OpenAI-compatible surface for the BIG OPEN weights (DeepSeek V3/R1, Llama-405B,
    // Qwen3-235B, Kimi K2). A CASH lane, but the ONLY hosted one that takes a **GBNF grammar** in
    // `response_format`, so an open model gets the same emission floor the local lane gets (bead
    // kestrel-y31). The grammar is attached by the decoding middleware in `buildModel`, not here.
    const apiKey = process.env.FIREWORKS_API_KEY;
    const fireworks = createOpenAICompatible({
      name: FIREWORKS_PROVIDER_NAME,
      baseURL: FIREWORKS_BASE_URL,
      ...(apiKey !== undefined ? { apiKey } : {}),
      ...(opts.fetch !== undefined ? { fetch: opts.fetch } : {}),
    });
    return fireworks(opts.modelId);
  }
  // Direct Anthropic path — apiKey defaults to ANTHROPIC_API_KEY from the environment.
  const anthropic = createAnthropic({});
  return anthropic(opts.modelId);
}

/**
 * The lane's model, with its constrained-decoding constraint WRAPPED AROUND IT. The wrap (rather than a
 * per-call argument) is deliberate: the constraint then rides every call the model makes, so no call site
 * — present or future — can forget it and quietly hand back an unconstrained run that looks like a
 * judgment result (bead kestrel-y31).
 */
function buildModel(opts: AiSdkClientOptions): LanguageModel {
  // Fail-closed: never construct a client with no model — an empty/whitespace id is an unresolved
  // model, and silently proceeding would send a request that either errors on the wire or, worse,
  // gets a provider default, corrupting the run's model identity (kestrel-rul.1). The friendly-name →
  // inference-profile MAPPING lives in `./roster.ts` (resolveBedrockModelId); this is the last-line
  // structural guard so a mis-resolution can never reach the wire as a blank id.
  if (opts.modelId.trim() === "") {
    throw new Error(`empty model id for provider "${opts.provider}" — refusing to construct a client with no model (fail-closed, kestrel-rul.1).`);
  }
  const base = buildBaseModel(opts);
  const constraint = resolveDecoding(opts);
  if (constraint === undefined) return base;
  // The middleware is typed structurally in `./decoding.ts` (which imports NO provider package); bridge it
  // to the SDK's `LanguageModelMiddleware` here, in the provider-quarantine file (the ONE impure module).
  const middleware = decodingMiddleware(constraint) as unknown as LanguageModelMiddleware;
  return wrapLanguageModel({ model: base, middleware });
}

/**
 * Construct the BYOK AI-SDK {@link LlmClient}. Retries, silent fallbacks, and middleware repair are
 * OFF by default (ADR-0013 (c) / m9i.1: the harness must never quietly turn one config into another).
 * Each `complete` returns the reply text, the raw reasoning trace (free evidence), provider-native
 * usage, and — unless disabled — the SCRUBBED wire evidence.
 */
export function aiSdkLlmClient(opts: AiSdkClientOptions): LlmClient {
  // vertex-maas: the ADC token source is hoisted OUT of buildModel so every token it ever issues can be
  // fed to scrubSecrets at capture time. An ADC bearer lives in NO env var, so the CREDENTIAL_ENV_VARS
  // scrub alone would never catch it — this is the same belt-and-suspenders that keeps every other
  // credential out of wire evidence (ADR-0013 (e)), extended to a credential that is minted per call.
  const adcSource =
    opts.provider === "vertex-maas"
      ? (opts.adcTokenSource ?? gcloudAdcTokenSource({ ...(opts.adcExec !== undefined ? { exec: opts.adcExec } : {}) }))
      : undefined;
  const model = buildModel(adcSource !== undefined ? { ...opts, adcTokenSource: adcSource } : opts);
  // The gateway lane floors its output ceiling at 16 — the Vercel AI Gateway 400s on `max_tokens<16`
  // (VERIFIED LIVE). Other lanes keep the caller's exact budget (or the 2048 default).
  const requestedMaxOutputTokens = opts.maxOutputTokens ?? 2048;
  // The gateway lane floors at 16 (the gateway 400s below that) AND, for a reasoning model, ADDS the
  // caller's reasoning headroom on top: over the OpenAI-compatible surface the thinking trace and the
  // visible reply share one `max_tokens` budget, so without headroom a small ceiling is burned entirely
  // on reasoning and the plan text returns empty (kestrel-a1f). Headroom is gateway-only and defaults to
  // 0, so every other lane — and any gateway caller that doesn't opt in — keeps its exact prior budget.
  const reasoningHeadroom = opts.provider === "gateway" ? (opts.reasoningHeadroomTokens ?? 0) : 0;
  const maxOutputTokens =
    opts.provider === "gateway"
      ? Math.max(GATEWAY_MIN_OUTPUT_TOKENS, requestedMaxOutputTokens) + reasoningHeadroom
      : requestedMaxOutputTokens;
  const captureWire = opts.captureWire ?? true;
  const secrets = secretsToScrub(opts.secrets);
  const generate: GenerateFn =
    opts.generate ??
    (async ({ system, messages, temperature, maxOutputTokens: callMaxOutputTokens, maxRetries, providerOptions }) => {
      const r = await generateText({
        model,
        // `system` is either a plain string or an {@link AiSdkSystemMessage} carrying the cache breakpoint;
        // the ai v7 top-level `system`/`instructions` option accepts both (`Instructions = string |
        // SystemModelMessage | SystemModelMessage[]`). The `providerOptions` record is looser than the SDK's
        // `ProviderOptions`, so it is bridged here in the provider-quarantine file (the ONE impure module).
        ...(system !== undefined ? { system: system as unknown as string } : {}),
        // AI-SDK `ModelMessage` is a role-discriminated union whose content type depends on the role; the
        // client's uniform `{ role, content: string, providerOptions? }` shape is a structural member of it,
        // bridged here in the provider-quarantine file (the ONE impure module) so the rest of the client
        // depends only on the narrow `AiSdkMessage` shape.
        messages: messages as unknown as ModelMessage[],
        ...(temperature !== undefined ? { temperature } : {}),
        // The per-CALL ceiling (raised above the thinking budget when thinking is enabled — see `complete`);
        // for a no-thinking request it equals the closure `maxOutputTokens`, so the wire is unchanged.
        maxOutputTokens: callMaxOutputTokens,
        maxRetries,
        // Top-level extended-thinking config, or ABSENT for `thinkingLevel:"none"` (byte-identical to today).
        // The record is looser than the SDK's `ProviderOptions`; bridged here in the provider-quarantine file.
        ...(providerOptions !== undefined ? { providerOptions: providerOptions as never } : {}),
      });
      // The SDK's rich generic result is a structural superset of the slice this client reads; bridge it to
      // {@link GenerateResult} here in the provider-quarantine file (the ONE impure module) so the rest of the
      // client — and the injectable test seam — depend only on the narrow shape.
      return r as unknown as GenerateResult;
    });

  return {
    async complete(req): Promise<LlmCompletion> {
      // Translate the harness's cache signals to provider-native controls (kestrel-rul). The growing
      // conversation rides `messages` (user/assistant only); a marked message carries the cache breakpoint.
      // The cached SYSTEM prefix rides the TOP-LEVEL `system`/`instructions` field as a SystemModelMessage
      // carrying the breakpoint — NOT a leading system-role message in `messages`, which the ai v7 core
      // rejects outright ("System messages are not allowed in the prompt or messages fields. Use the
      // instructions option instead"); that rejection is exactly what fell `conversation-cached` back to a
      // 100% provider-error run on live Bedrock (dry-run-1). Both the Bedrock (cachePoint) and Anthropic
      // (cacheControl) providers fold a system message's providerOptions into their top-level system field
      // with the cache point attached. When a provider has no breakpoint (Google: `bp === undefined`), the
      // system stays a plain string and no per-message providerOptions is added — Gemini caches implicitly,
      // and its cache-read tokens still surface through usage. Uncached policies also keep the plain string.
      // `req.cacheTtl` (kestrel-wa0j.1) rides every breakpoint this request marks; absent ⇒ the provider
      // default (byte-identical), and a TTL on a lane with no breakpoint primitive throws (fail-closed).
      const bp = cacheBreakpoint(opts.provider, req.cacheTtl);
      const convo: AiSdkMessage[] = req.messages.map((m) => ({
        role: m.role,
        content: m.content,
        ...(m.cache === true && bp !== undefined ? { providerOptions: bp } : {}),
      }));
      const cacheSystem = req.cacheSystem === true;
      const system: string | AiSdkSystemMessage =
        cacheSystem && bp !== undefined
          ? { role: "system", content: req.system, providerOptions: bp }
          : req.system;
      // Wire the strategist/watcher reasoning-effort dial (`thinkingLevel`) to the provider-native
      // extended-thinking/effort config (the dead-knob fix). `none` ⇒ `thinking === undefined` ⇒ NO
      // `providerOptions.*` key and the ceiling is untouched, so the request is byte-identical to before
      // this knob was wired; only an explicit `low`+ opts in. The gpt/open lanes (openai/azure/fireworks)
      // ride the same seam as `reasoning_effort` (kestrel-0drf); only gateway/local/codex-cli get
      // `undefined` here (gateway uses `reasoningHeadroomTokens`; codex-cli maps the dial onto its CLI
      // flag in ./codex-cli-client.ts — see {@link thinkingProviderOptions}).
      const thinking = thinkingProviderOptions(opts.provider, req.sampling.t