agent-accelerator 0.3.0 → 0.4.5

package/src/providers/shared.ts

@@ -0,0 +1,190 @@
+/**
+ * Shared provider helpers (Q-45-partial).
+ *
+ * Single home for the `redactedHeaders` / `readErrorPayload` /
+ * `parseArguments` / `normalizeThinkingParts` helpers previously duplicated
+ * across the four REST transports, plus timeout/signal and id helpers.
+ */
+import { toConciseProviderError } from "../utils/errors.ts";
+
+/** Parses a JSON tool-argument string; non-objects degrade to `{ raw }`. */
+export function parseArguments(raw: string | undefined): Record<string, unknown> {
+  if (!raw) return {};
+  try {
+    const parsed: unknown = JSON.parse(raw);
+    if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
+      return parsed as Record<string, unknown>;
+    }
+    return { raw };
+  } catch {
+    return { raw };
+  }
+}
+
+/** Trims/collapses assembled thinking parts (no edge-tripling). */
+export function normalizeThinkingParts(parts: string[]): string | undefined {
+  const cleaned = parts
+    .map((p) => p.replace(/\n{3,}/g, "\n\n").trim())
+    .filter((p) => p.length > 0);
+  return cleaned.length > 0 ? cleaned.join("\n") : undefined;
+}
+
+/**
+ * Redacts secret-bearing headers for the raw audit trail. Superset of the
+ * four adapter versions: `authorization` (any case) and `x-goog-api-key` are
+ * redacted; all other headers pass through untouched.
+ */
+export function redactedHeaders(headers: Record<string, string>): Record<string, string> {
+  const out: Record<string, string> = {};
+  for (const [k, v] of Object.entries(headers)) {
+    const lower = k.toLowerCase();
+    out[k] = lower === "authorization" || lower === "x-goog-api-key" ? "[REDACTED]" : v;
+  }
+  return out;
+}
+
+/**
+ * Extracts `{ message, code, errorType, raw }` from a non-2xx JSON body. Handles
+ * OpenAI `type`, OpenRouter `metadata.error_type` + `metadata.raw`, and Google
+ * `status` code shapes; unknown bodies degrade to a 300-char slice.
+ */
+export function readErrorPayload(
+  bodyText: string
+): { message: string; code?: string | number; errorType?: string; raw?: string } {
+  try {
+    const parsed: unknown = JSON.parse(bodyText);
+    const first = Array.isArray(parsed) ? parsed[0] : parsed;
+    const err = (first as { error?: { message?: string; code?: string | number; type?: string; status?: string | number; metadata?: { error_type?: string; raw?: unknown } } })?.error;
+    if (err && typeof err.message === "string") {
+      const code = err.code ?? err.status;
+      const errorType =
+        err.type ?? err.metadata?.error_type ?? (first as { error_type?: string })?.error_type;
+      const rawCandidate =
+        (err as { metadata?: { raw?: unknown } }).metadata?.raw ??
+        (first as { metadata?: { raw?: unknown } })?.metadata?.raw;
+      const raw = typeof rawCandidate === "string" && rawCandidate.trim() ? rawCandidate.trim().slice(0, 500) : undefined;
+      return {
+        message: err.message,
+        ...(code !== undefined ? { code } : {}),
+        ...(typeof errorType === "string" ? { errorType } : {}),
+        ...(raw ? { raw } : {}),
+      };
+    }
+    return { message: bodyText.slice(0, 300) };
+  } catch {
+    return { message: bodyText.slice(0, 300) };
+  }
+}
+
+/** Retryable incomplete-stream error for truncated SSE (Q-23). */
+export function incompleteStreamError(provider: string, modelId: string): Error {
+  const err = toConciseProviderError(
+    Object.assign(new Error("Stream ended without terminal event"), {
+      statusCode: 502,
+      transient: true,
+      retryable: true,
+      url: provider,
+    }),
+    provider,
+    modelId
+  );
+  (err as any).code = "incomplete_stream";
+  return err;
+}
+
+/** Per-request timeout ms from provider options (Q-22). */
+export function timeoutFor(options: { timeoutMs?: number } | undefined): number | undefined {
+  const t = (options as { timeoutMs?: unknown } | undefined)?.timeoutMs;
+  return typeof t === "number" && Number.isFinite(t) && t > 0 ? Math.floor(t) : undefined;
+}
+
+/** Merges a parent signal with a timeout into one effective signal. Caller must call cleanup when the request settles so the timer never holds the loop open. */
+export function combinedSignal(signal?: AbortSignal, timeoutMs?: number): { signal: AbortSignal | undefined; cleanup: () => void } {
+  if (timeoutMs === undefined) return { signal, cleanup: () => {} };
+  const ctrl = new AbortController();
+  const timer = setTimeout(() => {
+    try { ctrl.abort(); } catch {}
+  }, timeoutMs);
+  try { (timer as unknown as { unref?: () => void }).unref?.(); } catch {}
+  const cleanup = () => {
+    try { clearTimeout(timer); } catch {}
+  };
+  if (signal) {
+    if (signal.aborted) {
+      cleanup();
+      ctrl.abort();
+    } else {
+      signal.addEventListener("abort", () => {
+        cleanup();
+        try { ctrl.abort(); } catch {}
+      }, { once: true });
+    }
+  }
+  return { signal: ctrl.signal, cleanup };
+}
+
+/** Q-34: crypto-backed tool-call id, no Math.random primary. */
+export function newToolCallId(prefix = "call"): string {
+  try {
+    const c = (globalThis as any)?.crypto;
+    if (c?.randomUUID) return `${prefix}_${c.randomUUID().replace(/-/g, "").slice(0, 12)}`;
+  } catch {}
+  try {
+    const nc: any = (globalThis as any)?.process?.getBuiltinModule?.("node:crypto") ?? null;
+    if (nc?.randomUUID) return `${prefix}_${nc.randomUUID().replace(/-/g, "").slice(0, 12)}`;
+  } catch {}
+  return `${prefix}_${Math.random().toString(36).slice(2, 10)}`;
+}
+
+/**
+ * True when custom headers already carry auth (proxy/Copilot flows).
+ * Case-insensitive, non-empty `Authorization` or `cf-aig-authorization`.
+ * Callers return `"unused"` instead of throwing missing-key.
+ */
+export function hasAuthHeader(headers: Record<string, string> | undefined): boolean {
+  if (!headers) return false;
+  for (const [k, v] of Object.entries(headers)) {
+    const lower = k.toLowerCase();
+    if ((lower === "authorization" || lower === "cf-aig-authorization") && typeof v === "string" && v.trim().length > 0) {
+      return true;
+    }
+  }
+  return false;
+}
+
+/** True for truncated streamed tool args (`{ raw }` marker from `parseStreamedToolArguments`). */
+export function hasTruncatedToolArguments(args: unknown): boolean {
+  if (!args || typeof args !== "object" || Array.isArray(args)) return false;
+  const keys = Object.keys(args as Record<string, unknown>);
+  return keys.length === 1 && keys[0] === "raw" && typeof (args as Record<string, unknown>).raw === "string";
+}
+
+/** Short hash for tool-ID clamping (FNV-1a, 8 hex chars, no deps). */
+export function shortHashId(value: string): string {
+  let h = 0x811c9dc5;
+  for (let i = 0; i < value.length; i++) {
+    h ^= value.charCodeAt(i);
+    h = Math.imul(h, 0x01000193);
+  }
+  return (h >>> 0).toString(16).padStart(8, "0");
+}
+
+/**
+ * Clamps a tool-call ID to the 40-char OpenAI limit (pi `api/openai-completions.ts:1194`).
+ * Pipe IDs (`callId|itemId`) become `callId_itemHash`; overlong IDs truncate
+ * with hash suffix. Plain IDs pass through (truncated if needed).
+ */
+export function clampToolCallId(id: string): string {
+  const clean = String(id ?? "").replace(/[^a-zA-Z0-9_-]/g, "_");
+  if (!clean.includes("|")) {
+    return clean.length > 40 ? clean.slice(0, 40) : clean || "call_unknown";
+  }
+  const sep = clean.indexOf("|");
+  const callId = clean.slice(0, sep).replace(/_+$/, "") || "call";
+  const itemId = clean.slice(sep + 1).replace(/_+$/, "");
+  const combined = itemId ? `${callId}_${itemId}` : callId;
+  if (combined.length <= 40) return combined;
+  const hash = shortHashId(id).slice(0, 8);
+  const prefix = callId.slice(0, Math.max(1, 40 - hash.length - 1));
+  return `${prefix}_${hash}`;
+}

package/src/providers.ts

@@ -10,6 +10,9 @@
  * redesigning this abstraction — only new per-provider mappers/adapters.
  */
 import type { CacheConfig, ServiceTier, ThinkingLevel } from "./types/core.ts";
+import type { StructuredOutputSpec } from "./types/model.ts";
+import { zodToJsonSchema } from "./tools/schema.ts";
+import { ValidationError } from "./types/errors.ts";
 
 /** Classification of a canonical capability on a given provider. */
 export type ProviderCapabilityStatus =
@@ -50,9 +53,8 @@ export function emitProviderWarning(options: {
   // Leading newline: warnings fire at the start of a turn's request build,
   // when the previous turn's streamed text may have left the terminal cursor
   // mid-line. Without it the warning glues onto streamed output.
-  console.warn(
-    `\n[Agent Accelerator] WARNING [${options.provider}] ${options.capability}${requested}: ${options.reason} Using ${options.fallback} instead.`
-  );
+  // Q-41 dual: telemetry bus (new) + console.warn (compat for existing captureWarnings tests; /cli logger replaces in 0.5).
+  { const msg = `[Agent Accelerator] WARNING [${options.provider}] ${options.capability}${requested}: ${options.reason} Using ${options.fallback} instead.`; try { (globalThis as any).__agentAccelTelemetry?.emitWarning?.({ code: "provider_warning", provider: options.provider, capability: options.capability, message: msg }); } catch {} try { (globalThis as any).__agentAccelWarnSink?.(msg, options); } catch {} try { console.warn(msg); } catch {} }
 }
 
 /** Clears deduped-warning state (mainly for tests). */
@@ -184,7 +186,7 @@ export function mapThinkingLevelToOpenRouterChat(
  * echoed but are NOT documented levels, so they clamp to `high` with a warning.
  *
  * @deprecated The Responses skin is discontinued for OpenRouter
- * (`src/providers/openrouter-responses.ts`, archived). Use
+ * (removed from core). Use
  * {@link mapThinkingLevelToOpenRouterChat} with the stable Chat Completions
  * transport instead.
  */
@@ -358,17 +360,17 @@ export function mapServiceTierToOpenAI(tier: ServiceTier | undefined): "flex" |
 
 /**
  * Applies canonical cache config for OpenAI. Session affinity flows via
- * `prompt_cache_key` (handled by the adapter); explicit retention control
- * lives in `prompt_cache_options` (gpt-5.6+ explicit breakpoints) which is
- * out of scope for the most-important subset, so retention/cachedContentId
- * are UNSUPPORTED: warn and drop. `prompt_cache_retention` is deprecated.
+ * `prompt_cache_key` (handled by the adapter); `long` retention maps to
+ * `prompt_cache_retention: "24h"` when the catalog supports it (handled by
+ * the adapter, pi `api/openai-responses.ts:97` rule). Other explicit
+ * retention modes and `cachedContentId` are UNSUPPORTED: warn and drop.
  */
 export function applyCacheForOpenAI(
   cache: CacheConfig | undefined,
   modelRef: string
 ): void {
   if (!cache) return;
-  if (cache.retention && cache.retention !== "implicit") {
+  if (cache.retention && cache.retention !== "implicit" && cache.retention !== "long") {
     emitProviderWarning({
       provider: "openai",
       capability: "cache retention",
@@ -490,3 +492,251 @@ export function parseStreamedToolArguments(startText: string, deltaText: string)
   }
   return { raw: startText + deltaText };
 }
+
+// ---------------------------------------------------------------------------
+// Structured outputs (canonical, pure, no I/O)
+// ---------------------------------------------------------------------------
+
+function isZodLike(value: unknown): boolean {
+  if (!value || typeof value !== "object") return false;
+  const v = value as Record<string, unknown>;
+  return (
+    typeof v["safeParse"] === "function" ||
+    typeof v["parse"] === "function" ||
+    "_def" in v ||
+    "~standard" in v
+  );
+}
+
+function looksLikeJsonSchema(value: unknown): boolean {
+  if (!isPlainObject(value)) return false;
+  const v = value as Record<string, unknown>;
+  return (
+    typeof v["type"] === "string" ||
+    isPlainObject(v["properties"]) ||
+    Array.isArray(v["anyOf"]) ||
+    Array.isArray(v["oneOf"]) ||
+    Array.isArray(v["allOf"]) ||
+    typeof v["$schema"] === "string" ||
+    isPlainObject(v["$defs"]) ||
+    isPlainObject(v["definitions"])
+  );
+}
+
+function sanitizeOutputName(raw: unknown): string {
+  const base = typeof raw === "string" && raw.trim() ? raw.trim() : "structured_output";
+  const cleaned = base.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 64) || "structured_output";
+  return cleaned;
+}
+
+function schemaToJsonSchema(schema: unknown, what: string): Record<string, unknown> {
+  if (isZodLike(schema)) {
+    try {
+      const converted = zodToJsonSchema(schema);
+      if (isPlainObject(converted)) return converted;
+    } catch {
+      // fall through to error below
+    }
+  }
+  if (isPlainObject(schema)) return schema as Record<string, unknown>;
+  throw new ValidationError(
+    `[Agent Accelerator] Invalid structured output ${what}: expected a Zod schema, a JSON Schema object, or { name, schema, strict }.`
+  );
+}
+
+/**
+ * Normalizes `AgentConfig.output` into a provider-agnostic spec.
+ *
+ * Accepts a Zod schema, a JSON Schema object, or an explicit
+ * `{ name?, description?, schema, strict? }` wrapper. Returns `undefined`
+ * when no output was requested. Throws `ValidationError` for invalid shapes.
+ *
+ * @example `normalizeStructuredOutput(z.object({ city: z.string() }))`
+ */
+export function normalizeStructuredOutput(output: unknown): StructuredOutputSpec | undefined {
+  if (output === undefined || output === null) return undefined;
+  if (isPlainObject(output) && "schema" in (output as Record<string, unknown>)) {
+    const rec = output as Record<string, unknown>;
+    const schema = schemaToJsonSchema(rec["schema"], "`output.schema`");
+    return {
+      name: sanitizeOutputName(rec["name"]),
+      schema,
+      strict: typeof rec["strict"] === "boolean" ? (rec["strict"] as boolean) : true,
+      ...(typeof rec["description"] === "string" && (rec["description"] as string).trim()
+        ? { description: (rec["description"] as string).trim() }
+        : {}),
+    };
+  }
+  if (isZodLike(output) || looksLikeJsonSchema(output)) {
+    return {
+      name: "structured_output",
+      schema: schemaToJsonSchema(output, "`output`"),
+      strict: true,
+    };
+  }
+  throw new ValidationError(
+    "[Agent Accelerator] Invalid structured output: expected a Zod schema, a JSON Schema object, or { name, schema, strict }."
+  );
+}
+
+/**
+ * Extracts the original Zod validator from a canonical output option, when
+ * present (bare Zod schema or `{ schema: Zod }`). Used by the loop for
+ * post-parse validation; `undefined` means JSON-parse only.
+ */
+export function extractOutputValidator(output: unknown): { safeParse: (v: unknown) => { success: boolean; data?: unknown; error?: { message: string } } } | undefined {
+  if (!output) return undefined;
+  if (isZodLike(output) && typeof (output as { safeParse?: unknown }).safeParse === "function") {
+    return output as { safeParse: (v: unknown) => { success: boolean; data?: unknown; error?: { message: string } } };
+  }
+  if (isPlainObject(output) && isZodLike((output as Record<string, unknown>)["schema"])) {
+    const inner = (output as Record<string, unknown>)["schema"] as { safeParse?: unknown };
+    if (typeof inner.safeParse === "function") {
+      return inner as { safeParse: (v: unknown) => { success: boolean; data?: unknown; error?: { message: string } } };
+    }
+  }
+  return undefined;
+}
+
+function stripCodeFences(text: string): string {
+  const trimmed = text.trim();
+  const fenced = trimmed.match(/^```(?:json)?\s*\n([\s\S]*?)\n```\s*$/i);
+  if (fenced && typeof fenced[1] === "string") return fenced[1].trim();
+  return trimmed;
+}
+
+/**
+ * Parses and validates a final answer against a canonical output option.
+ * Strips markdown fences, requires valid JSON, then applies the Zod validator
+ * when the original option carried one. Throws `ValidationError` on mismatch.
+ *
+ * @example `parseStructuredOutput('{"city":"Paris"}', z.object({ city: z.string() }))`
+ */
+export function parseStructuredOutput(text: string, output: unknown): unknown {
+  const cleaned = stripCodeFences(typeof text === "string" ? text : String(text ?? ""));
+  if (!cleaned) {
+    throw new ValidationError(
+      "[Agent Accelerator] Structured output validation failed: model returned empty text, expected JSON matching the output schema."
+    );
+  }
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(cleaned);
+  } catch {
+    throw new ValidationError(
+      `[Agent Accelerator] Structured output validation failed: model text is not valid JSON (${cleaned.slice(0, 120)}).`
+    );
+  }
+  const validator = extractOutputValidator(output);
+  if (validator) {
+    const res = validator.safeParse(parsed);
+    if (!res.success) {
+      const detail = res.error?.message ? ` ${res.error.message}`.slice(0, 500) : "";
+      throw new ValidationError(
+        `[Agent Accelerator] Structured output validation failed: JSON does not match the output schema.${detail}`
+      );
+    }
+    return res.data ?? parsed;
+  }
+  return parsed;
+}
+
+/**
+ * Maps a normalized spec onto OpenAI Responses `text.format`.
+ * `{ type: json_schema, name, schema, strict }` (docs: responses/create).
+ */
+export function mapOutputToOpenAI(
+  spec: StructuredOutputSpec | undefined
+): { format: { type: "json_schema"; name: string; schema: Record<string, unknown>; strict: boolean; description?: string } } | undefined {
+  if (!spec) return undefined;
+  return {
+    format: {
+      type: "json_schema",
+      name: spec.name,
+      schema: spec.schema,
+      strict: spec.strict ?? true,
+      ...(spec.description ? { description: spec.description } : {}),
+    },
+  };
+}
+
+/**
+ * Maps a normalized spec onto Google Interactions `response_format`.
+ * `{ type: text, mime_type: application/json, schema }` (docs: structured-output).
+ */
+export function mapOutputToGoogle(
+  spec: StructuredOutputSpec | undefined
+): { responseFormat: { type: "text"; mime_type: "application/json"; schema: Record<string, unknown> } } | undefined {
+  if (!spec) return undefined;
+  return {
+    responseFormat: {
+      type: "text",
+      mime_type: "application/json",
+      schema: spec.schema,
+    },
+  };
+}
+
+/** Chat Completions `response_format` wire shape (OpenRouter + compat). */
+export type ChatCompletionsResponseFormat =
+  | { type: "json_schema"; json_schema: { name: string; schema: Record<string, unknown>; strict: boolean; description?: string } };
+
+/**
+ * Maps a normalized spec onto Chat Completions `response_format`.
+ * `{ type: json_schema, json_schema: { name, schema, strict } }`
+ * (docs: openai completions/create, openrouter parameters).
+ */
+export function mapOutputToOpenRouterChat(
+  spec: StructuredOutputSpec | undefined
+): { response_format: ChatCompletionsResponseFormat } | undefined {
+  if (!spec) return undefined;
+  return {
+    response_format: {
+      type: "json_schema",
+      json_schema: {
+        name: spec.name,
+        schema: spec.schema,
+        strict: spec.strict ?? true,
+        ...(spec.description ? { description: spec.description } : {}),
+      },
+    },
+  };
+}
+
+/**
+ * Maps a normalized spec onto generic compat Chat Completions
+ * `response_format` (same standard shape as OpenRouter Chat).
+ */
+export function mapOutputToCompatChat(
+  spec: StructuredOutputSpec | undefined
+): { response_format: ChatCompletionsResponseFormat } | undefined {
+  return mapOutputToOpenRouterChat(spec);
+}
+
+/**
+ * Builds the emulated structured-output instruction for models without
+ * native `response_format` support (catalog `structured_output: false`).
+ * Appended to the system prompt so the model still targets JSON; the loop
+ * still validates into `AgentResponse.parsed`.
+ *
+ * @example `augmentSystemPromptWithStructuredOutput(sys, spec)`
+ */
+export function buildStructuredOutputInstruction(spec: StructuredOutputSpec): string {
+  return `Return ONLY valid JSON matching this JSON Schema, with no markdown fences, no commentary, no extra text.\nSchema: ${JSON.stringify(spec.schema)}`;
+}
+
+/**
+ * Returns a system prompt with the emulated structured-output instruction
+ * appended (stable per Agent since `output` is agent-level, so prefix-cache
+ * affinity is preserved). No-op when spec is undefined.
+ */
+export function augmentSystemPromptWithStructuredOutput(
+  systemPrompt: string | undefined,
+  spec: StructuredOutputSpec | undefined
+): string | undefined {
+  if (!spec) return systemPrompt;
+  const instruction = `[Structured Output]\n${buildStructuredOutputInstruction(spec)}`;
+  if (!systemPrompt) return instruction;
+  if (systemPrompt.includes("[Structured Output]")) return systemPrompt;
+  return `${systemPrompt}\n\n${instruction}`;
+}