@warmdrift/kgauto-compiler 2.0.0-alpha.15 → 2.0.0-alpha.17

package/dist/index.js

@@ -1475,6 +1475,215 @@ var PROFILES_RAW = [
       // sequential tools — same as V4-Flash
     }
   },
+  // ── OpenAI ──
+  // alpha.16 (2026-05-17): close the half-supported provider gap. env.ts
+  // already registered OPENAI_API_KEY + executeOpenAI + normalizeOpenAILike
+  // + lowerOpenAI all existed; profile entries were missing, so the
+  // alpha.10 auto-filter would mark openai-keyed models reachable but
+  // there were no profiles to filter IN. Half-supported is now fully
+  // supported. PB request `openai-provider-profiles` (2026-05-16).
+  //
+  // Profile data verified against developers.openai.com/api/docs/pricing
+  // + per-model pages 2026-05-17. L-049/L-081 step-zero: no AI-trained
+  // numbers — fetched live from OpenAI's docs. As of 2026-05, OpenAI's
+  // current flagship is gpt-5.5 (2025-12 cutoff); gpt-5.4-{base,mini,nano}
+  // are the workhorse family. gpt-4.1 + gpt-4o are legacy.
+  //
+  // Both 5.5 and 5.4 carry a 272K input-token pricing cliff (2x input,
+  // 1.5x output beyond that). Modeled as a `downgrade_quality_warning`
+  // cliff because it ranks the model down at large-context shapes — the
+  // semantics of "this model is now 2x more expensive" map onto the
+  // existing penalty mechanism. Cost-watcher will catch high-context
+  // spikes empirically; the cliff prevents naive routing into the doubled
+  // pricing zone.
+  {
+    id: "gpt-5.5",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 105e4,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 272e3,
+        action: "downgrade_quality_warning",
+        reason: "OpenAI pricing tier shift: >272K input tokens billed at 2x input + 1.5x output rates"
+      }
+    ],
+    costInputPer1m: 5,
+    costOutputPer1m: 30,
+    lowering: {
+      system: { mode: "inline" },
+      // OpenAI caching is implicit (auto-applied to repeated prefixes
+      // ≥1024 tokens for prompt_tokens_details.cached_tokens). No
+      // wire-format marker. Discount: 10x for cached input ($0.50/$5.00).
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["reasoning", "agentic_coding", "long_context", "structured_output", "reliable_tool_use", "reasoning_effort_knob"],
+    weaknesses: ["cost", "pricing_cliff_at_272k"],
+    notes: "OpenAI frontier (2026-05). 1M context (1.05M total), 128K max output, 2025-12 cutoff. Reasoning effort knob (none/low/medium/high/xhigh). Pricing cliff at 272K input.",
+    // Frontier-tier perf hypothesis. Anchored to Opus 4.7 row (similar
+    // price/positioning). Brain evidence will refine; no telemetry yet.
+    archetypePerf: {
+      critique: 9,
+      plan: 9,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      // parallel tool support good but cliff at 272K hurts deep multi-step
+      summarize: 7,
+      // overkill for tolerant archetype
+      classify: 7
+      // overkill; cheaper models cover this
+    }
+  },
+  {
+    id: "gpt-5.4",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 105e4,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 272e3,
+        action: "downgrade_quality_warning",
+        reason: "OpenAI pricing tier shift: >272K input tokens billed at 2x input + 1.5x output rates"
+      }
+    ],
+    costInputPer1m: 2.5,
+    costOutputPer1m: 15,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["reasoning", "long_context", "structured_output", "reliable_tool_use"],
+    weaknesses: ["pricing_cliff_at_272k"],
+    notes: "OpenAI workhorse (2026-05). 1M context (1.05M total), 128K max output, 2025-08 cutoff. Pricing cliff at 272K input. Pairs cleanly with Sonnet 4.6 on cost ($2.50/$15.00 vs $3.00/$15.00).",
+    // Anchored to Sonnet 4.6 row (similar price/positioning). Slight
+    // anthropic-side edge on agentic coding per master plan vibe.
+    archetypePerf: {
+      critique: 8,
+      plan: 8,
+      generate: 8,
+      ask: 8,
+      extract: 8,
+      transform: 8,
+      hunt: 7,
+      summarize: 7,
+      classify: 7
+    }
+  },
+  {
+    id: "gpt-5.4-mini",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 4e5,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 0.75,
+    costOutputPer1m: 4.5,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "speed", "agentic_coding", "structured_output", "reliable_tool_use"],
+    weaknesses: ["reasoning_depth"],
+    notes: "OpenAI mini-tier (2026-05). 400K context, 128K max output, 2025-08 cutoff. OpenAI describes as 'strongest mini model for coding, computer use, subagents.' Cache discount 10x ($0.075 input).",
+    // Mini-tier hypothesis. Anchored to Haiku 4.5 + Flash row pricing.
+    // Cost is slightly higher than Haiku ($0.75 vs $0.50 input) but
+    // OpenAI claims strong coding/subagent perf.
+    archetypePerf: {
+      ask: 7,
+      generate: 7,
+      extract: 7,
+      transform: 7,
+      classify: 7,
+      summarize: 7,
+      hunt: 7,
+      plan: 6,
+      critique: 5
+      // reasoning depth gap — frontier models handle this
+    }
+  },
+  {
+    id: "gpt-5.4-nano",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 4e5,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 0.2,
+    costOutputPer1m: 1.25,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "speed", "volume", "structured_output"],
+    weaknesses: ["reasoning_depth", "no_computer_use"],
+    notes: "OpenAI nano-tier (2026-05). 400K context, 128K max output, 2025-08 cutoff. 'Cheapest GPT-5.4-class for simple high-volume tasks.' No fine-tuning, no computer-use tools. Cache discount 10x.",
+    // Nano-tier. Anchored to Flash-Lite row ($0.10/$0.40 vs nano's
+    // $0.20/$1.25). Slightly more expensive than Flash-Lite but with
+    // OpenAI brand reliability. Good fit for classify/summarize floor.
+    archetypePerf: {
+      classify: 7,
+      summarize: 6,
+      ask: 6,
+      transform: 6,
+      extract: 6,
+      generate: 5,
+      hunt: 5,
+      plan: 4,
+      critique: 3
+      // not for reasoning archetypes
+    }
+  },
   // ── Auto-onboarded (UNVERIFIED) ──
   // Cloned by scripts/auto-onboard-models.mjs from a same-family template.
   // Each entry's pricing/context/cliffs/lowering reflects the template, NOT
@@ -2563,10 +2772,14 @@ var loadChainsFromBrain = createBrainQueryCache({
 // src/fallback.ts
 var STARTER_CHAINS = {
   // Reasoning floor — never degrade. Walk UP on 429 to Opus → cross-provider.
+  // alpha.16: gpt-5.5 appended as third-provider critique floor (frontier-tier,
+  // archetypePerf=9). Cross-provider-tail invariant has somewhere to land when
+  // both Anthropic + Google are unreachable (consumer adds only OpenAI key).
   critique: [
     "claude-opus-4-7",
     "claude-sonnet-4-6",
-    "gemini-2.5-pro"
+    "gemini-2.5-pro",
+    "gpt-5.5"
   ],
   // Reasoning matters — Sonnet primary; walk UP to Opus on 429 (rare exception
   // to "always cheaper"); cross-provider via Pro; DeepSeek Pro as tier 3 floor.
@@ -2577,25 +2790,29 @@ var STARTER_CHAINS = {
     "deepseek-v4-pro"
   ],
   // Quality + cost match. Walk Sonnet → Haiku same-provider, Pro cross,
-  // Flash floor for the open-posture chain.
+  // gpt-5.4-mini as third-provider tail (alpha.16 — closes the mono-Anthropic
+  // gap when consumer has only ANTHROPIC + OPENAI keys; archetypePerf=7).
   generate: [
     "claude-sonnet-4-6",
     "claude-haiku-4-5",
     "gemini-2.5-pro",
-    "gemini-2.5-flash"
+    "gpt-5.4-mini"
   ],
   ask: [
     "claude-sonnet-4-6",
     "claude-haiku-4-5",
     "gemini-2.5-pro",
-    "gemini-2.5-flash"
+    "gpt-5.4-mini"
   ],
   // Structured-output archetype — Flash skipped (alpha.8 MAX_TOKENS cliff),
-  // DeepSeek skipped (no brain evidence). Floor at Haiku.
+  // DeepSeek skipped (no brain evidence). Floor at Haiku. alpha.16: gpt-5.4
+  // appended as third-provider extract floor (archetypePerf=8, native
+  // structured-output support).
   extract: [
     "claude-sonnet-4-6",
     "claude-haiku-4-5",
-    "gemini-2.5-pro"
+    "gemini-2.5-pro",
+    "gpt-5.4"
   ],
   // Forgiving archetype — Sonnet primary but Flash safely floors it.
   transform: [
@@ -2708,9 +2925,321 @@ function ensureCrossProviderTail(opts) {
   return { chain };
 }
 
+// src/glassbox/types.ts
+var GLASSBOX_STREAM_TTL_MS = 6e4;
+
+// src/glassbox/pubsub-memory.ts
+var MemoryPubSub = class {
+  subscribers = /* @__PURE__ */ new Map();
+  async publish(traceId, event) {
+    const subs = this.subscribers.get(traceId);
+    if (!subs || subs.size === 0) return;
+    for (const sub of subs) {
+      if (sub.closed) continue;
+      try {
+        sub.controller.enqueue(event);
+      } catch {
+        sub.closed = true;
+        continue;
+      }
+      this.refreshTtl(traceId, sub);
+    }
+  }
+  subscribe(traceId) {
+    const self = this;
+    let sub;
+    return new ReadableStream({
+      start(controller) {
+        sub = {
+          controller,
+          ttlTimer: setTimeout(() => {
+            self.closeSubscriber(traceId, sub);
+          }, GLASSBOX_STREAM_TTL_MS),
+          closed: false
+        };
+        let set = self.subscribers.get(traceId);
+        if (!set) {
+          set = /* @__PURE__ */ new Set();
+          self.subscribers.set(traceId, set);
+        }
+        set.add(sub);
+      },
+      cancel() {
+        if (sub) self.removeSubscriber(traceId, sub);
+      }
+    });
+  }
+  /**
+   * Refresh the rolling TTL for a subscriber after an event lands. Replaces
+   * the existing timer with a fresh 60s one.
+   */
+  refreshTtl(traceId, sub) {
+    clearTimeout(sub.ttlTimer);
+    sub.ttlTimer = setTimeout(() => {
+      this.closeSubscriber(traceId, sub);
+    }, GLASSBOX_STREAM_TTL_MS);
+  }
+  /**
+   * Close the subscriber's stream cleanly and remove from the fan-out set.
+   * Idempotent — safe to call multiple times.
+   */
+  closeSubscriber(traceId, sub) {
+    if (sub.closed) return;
+    sub.closed = true;
+    clearTimeout(sub.ttlTimer);
+    try {
+      sub.controller.close();
+    } catch {
+    }
+    this.removeSubscriber(traceId, sub);
+  }
+  removeSubscriber(traceId, sub) {
+    clearTimeout(sub.ttlTimer);
+    const set = this.subscribers.get(traceId);
+    if (!set) return;
+    set.delete(sub);
+    if (set.size === 0) this.subscribers.delete(traceId);
+  }
+  /**
+   * Test-only reset. Tears down all subscribers, clears all state. Calling
+   * outside of tests is harmless but cancels every active stream.
+   */
+  _reset() {
+    for (const [, set] of this.subscribers) {
+      for (const sub of set) {
+        this.closeSubscriber("", sub);
+      }
+    }
+    this.subscribers.clear();
+  }
+};
+
+// src/glassbox/pubsub-upstash.ts
+var UpstashPubSub = class {
+  url;
+  token;
+  fetchImpl;
+  blockMs;
+  maxLen;
+  constructor(cfg) {
+    this.url = cfg.url.replace(/\/$/, "");
+    this.token = cfg.token;
+    this.fetchImpl = cfg.fetchImpl ?? globalThis.fetch.bind(globalThis);
+    this.blockMs = cfg.blockMs ?? 100;
+    this.maxLen = cfg.maxLen ?? 100;
+  }
+  async publish(traceId, event) {
+    const key = streamKey(traceId);
+    const payload = JSON.stringify(event);
+    await this.cmd([
+      "XADD",
+      key,
+      "MAXLEN",
+      "~",
+      String(this.maxLen),
+      "*",
+      "event",
+      payload
+    ]);
+    await this.cmd(["EXPIRE", key, String(Math.ceil(GLASSBOX_STREAM_TTL_MS / 1e3))]);
+  }
+  subscribe(traceId) {
+    const key = streamKey(traceId);
+    const self = this;
+    let cursor = "$";
+    let cancelled = false;
+    let ttlDeadline = Date.now() + GLASSBOX_STREAM_TTL_MS;
+    return new ReadableStream({
+      async start(controller) {
+        try {
+          while (!cancelled && Date.now() < ttlDeadline) {
+            const resp = await self.cmd([
+              "XREAD",
+              "BLOCK",
+              String(self.blockMs),
+              "STREAMS",
+              key,
+              cursor
+            ]);
+            if (cancelled) break;
+            const parsed = parseXReadResult(resp.result);
+            if (parsed.entries.length === 0) {
+              continue;
+            }
+            for (const entry of parsed.entries) {
+              const evt = decodeEvent(entry.fields);
+              if (evt) {
+                try {
+                  controller.enqueue(evt);
+                } catch {
+                  cancelled = true;
+                  break;
+                }
+              }
+              cursor = entry.id;
+            }
+            ttlDeadline = Date.now() + GLASSBOX_STREAM_TTL_MS;
+          }
+        } catch (err) {
+          if (!cancelled) {
+            try {
+              controller.error(err);
+            } catch {
+            }
+            return;
+          }
+        }
+        try {
+          controller.close();
+        } catch {
+        }
+      },
+      cancel() {
+        cancelled = true;
+      }
+    });
+  }
+  async cmd(args) {
+    const res = await this.fetchImpl(this.url, {
+      method: "POST",
+      headers: {
+        Authorization: `Bearer ${this.token}`,
+        "Content-Type": "application/json"
+      },
+      body: JSON.stringify(args)
+    });
+    if (!res.ok) {
+      throw new Error(`Upstash ${args[0]} failed: HTTP ${res.status}`);
+    }
+    const json = await res.json();
+    if (json.error) {
+      throw new Error(`Upstash ${args[0]} failed: ${json.error}`);
+    }
+    return json;
+  }
+};
+function streamKey(traceId) {
+  return `glassbox:trace:${traceId}`;
+}
+function decodeEvent(fields) {
+  const raw = fields["event"];
+  if (!raw) return void 0;
+  try {
+    const parsed = JSON.parse(raw);
+    if (typeof parsed.kind === "string" && typeof parsed.at === "number") {
+      return parsed;
+    }
+    return void 0;
+  } catch {
+    return void 0;
+  }
+}
+function parseXReadResult(raw) {
+  if (!Array.isArray(raw)) return { entries: [] };
+  const entries = [];
+  for (const stream of raw) {
+    if (!Array.isArray(stream) || stream.length < 2) continue;
+    const streamEntries = stream[1];
+    if (!Array.isArray(streamEntries)) continue;
+    for (const entry of streamEntries) {
+      if (!Array.isArray(entry) || entry.length < 2) continue;
+      const id = String(entry[0]);
+      const flat = entry[1];
+      if (!Array.isArray(flat)) continue;
+      const fields = {};
+      for (let i = 0; i < flat.length; i += 2) {
+        const k = flat[i];
+        const v = flat[i + 1];
+        if (typeof k === "string") fields[k] = String(v ?? "");
+      }
+      entries.push({ id, fields });
+    }
+  }
+  return { entries };
+}
+
+// src/glassbox/emit.ts
+var activePubSub;
+function getPubSub() {
+  if (activePubSub) return activePubSub;
+  const url = readEnv("UPSTASH_REDIS_URL");
+  const token = readEnv("UPSTASH_REDIS_TOKEN");
+  if (url && token) {
+    activePubSub = new UpstashPubSub({ url, token });
+  } else {
+    activePubSub = new MemoryPubSub();
+  }
+  return activePubSub;
+}
+function readEnv(key) {
+  try {
+    if (typeof process !== "undefined" && process.env) {
+      const v = process.env[key];
+      return v && v.trim() !== "" ? v : void 0;
+    }
+  } catch {
+  }
+  return void 0;
+}
+function emitGlassboxEvent(traceId, kind, data) {
+  if (!traceId) return;
+  const event = { kind, at: Date.now(), data };
+  const ps = getPubSub();
+  try {
+    const p = ps.publish(traceId, event);
+    if (p && typeof p.then === "function") {
+      p.catch(() => {
+      });
+    }
+  } catch {
+  }
+}
+function emitCompileStart(traceId, data) {
+  emitGlassboxEvent(traceId, "compile.start", data);
+}
+function emitCompileDone(traceId, data) {
+  emitGlassboxEvent(traceId, "compile.done", data);
+}
+function emitExecuteAttempt(traceId, data) {
+  emitGlassboxEvent(traceId, "execute.attempt", data);
+}
+function emitExecuteSuccess(traceId, data) {
+  emitGlassboxEvent(traceId, "execute.success", data);
+}
+function emitAdvisoryFired(traceId, data) {
+  emitGlassboxEvent(traceId, "advisory.fired", data);
+}
+function emitFallbackWalked(traceId, data) {
+  emitGlassboxEvent(traceId, "fallback.walked", data);
+}
+
 // src/call.ts
 async function call(ir, opts = {}) {
+  const traceId = generateTraceId();
+  safeEmit(
+    () => emitCompileStart(traceId, {
+      appId: ir.appId,
+      archetype: ir.intent.archetype,
+      models: ir.models
+    })
+  );
   const initial = compileAndRegister(ir, opts);
+  safeEmit(
+    () => emitCompileDone(traceId, {
+      target: initial.target,
+      provider: initial.provider,
+      fallbackChain: initial.fallbackChain,
+      tokensIn: initial.tokensIn,
+      estimatedCostUsd: initial.estimatedCostUsd,
+      mutationsApplied: initial.mutationsApplied,
+      advisories: initial.advisories
+    })
+  );
+  for (const adv of initial.advisories) {
+    safeEmit(
+      () => emitAdvisoryFired(traceId, { code: adv.code, message: adv.message })
+    );
+  }
   const start = Date.now();
   const attempts = [];
   const rawTargets = [initial.target, ...initial.fallbackChain];
@@ -2765,6 +3294,47 @@ async function call(ir, opts = {}) {
       }
     }
   }
+  let policyBlockedFiltered;
+  if (opts.policy?.blockedModels && opts.policy.blockedModels.length > 0) {
+    const blocked = new Set(opts.policy.blockedModels);
+    const filtered = [];
+    const dropped = [];
+    for (const t of targetsToTry) {
+      if (blocked.has(t)) {
+        dropped.push(t);
+      } else {
+        filtered.push(t);
+      }
+    }
+    if (dropped.length > 0) {
+      policyBlockedFiltered = dropped;
+      targetsToTry = filtered;
+    }
+    if (targetsToTry.length === 0) {
+      const latencyMs2 = Date.now() - start;
+      await record({
+        handle: initial.handle,
+        tokensIn: 0,
+        tokensOut: 0,
+        latencyMs: latencyMs2,
+        success: false,
+        errorType: "all_blocked_by_policy",
+        promptPreview: extractPromptPreview(ir)
+      });
+      const blockedAttempts = dropped.map((m) => ({
+        model: m,
+        status: "terminal",
+        errorCode: "blocked_by_policy",
+        message: `Skipped \u2014 model ${m} is in CompilePolicy.blockedModels`
+      }));
+      throw new CallError(
+        `call(): all chain targets blocked by CompilePolicy.blockedModels: [${dropped.join(", ")}]`,
+        blockedAttempts,
+        void 0,
+        "all_blocked_by_policy"
+      );
+    }
+  }
   let activeCompile = initial;
   let lastErr;
   const failedProviders = /* @__PURE__ */ new Set();
@@ -2800,6 +3370,9 @@ async function call(ir, opts = {}) {
         continue;
       }
     }
+    safeEmit(
+      () => emitExecuteAttempt(traceId, { model: targetModel, attemptIndex: i })
+    );
     const exec = await execute(activeCompile.request, {
       apiKeys: opts.apiKeys,
       fetchImpl: opts.fetchImpl,
@@ -2809,6 +3382,14 @@ async function call(ir, opts = {}) {
     if (validated.ok) {
       attempts.push({ model: targetModel, status: "success" });
       const latencyMs2 = Date.now() - start;
+      safeEmit(
+        () => emitExecuteSuccess(traceId, {
+          model: targetModel,
+          tokensIn: validated.response.tokens.input,
+          tokensOut: validated.response.tokens.output,
+          latencyMs: latencyMs2
+        })
+      );
       await record({
         handle: initial.handle,
         tokensIn: validated.response.tokens.input,
@@ -2825,6 +3406,20 @@ async function call(ir, opts = {}) {
         cacheCreationInputTokens: validated.response.tokens.cacheCreated
       });
       const fellOver = targetModel !== initial.target;
+      const fallbackReason = fellOver ? normalizeFallbackReason(attempts) : void 0;
+      if (fellOver) {
+        const firstFailed = attempts.find((a) => a.status !== "success");
+        if (firstFailed) {
+          safeEmit(
+            () => emitFallbackWalked(traceId, {
+              from: initial.target,
+              to: targetModel,
+              reason: fallbackReason ?? "unknown",
+              attempt: firstFailed
+            })
+          );
+        }
+      }
       return {
         handle: initial.handle,
         actualModel: targetModel,
@@ -2836,8 +3431,10 @@ async function call(ir, opts = {}) {
         attempts,
         servedBy: targetModel,
         fellOverFrom: fellOver ? initial.target : void 0,
-        fallbackReason: fellOver ? normalizeFallbackReason(attempts) : void 0,
-        unreachableFiltered
+        fallbackReason,
+        unreachableFiltered,
+        policyBlockedFiltered,
+        traceId
       };
     }
     attempts.push({
@@ -2866,8 +3463,9 @@ async function call(ir, opts = {}) {
     promptPreview: extractPromptPreview(ir)
   });
   const filteredNote = unreachableFiltered && unreachableFiltered.length > 0 ? ` (also auto-filtered: [${unreachableFiltered.join(", ")}] \u2014 no API key)` : "";
+  const blockedNote = policyBlockedFiltered && policyBlockedFiltered.length > 0 ? ` (also policy-blocked: [${policyBlockedFiltered.join(", ")}])` : "";
   throw new CallError(
-    `call(): all attempts failed${lastErr ? ` \u2014 ${lastErr.errorCode}: ${lastErr.message}` : ""}${filteredNote}`,
+    `call(): all attempts failed${lastErr ? ` \u2014 ${lastErr.errorCode}: ${lastErr.message}` : ""}${filteredNote}${blockedNote}`,
     attempts,
     lastErr?.status,
     lastErr?.errorCode
@@ -2933,6 +3531,23 @@ function normalizeFallbackReason(attempts) {
   if (code === "auth" || code === "auth_inferred") return "provider_auth_failed";
   return "provider_error";
 }
+function generateTraceId() {
+  try {
+    const g = globalThis;
+    if (g.crypto && typeof g.crypto.randomUUID === "function") {
+      return g.crypto.randomUUID();
+    }
+  } catch {
+  }
+  const hex = (n) => Math.floor(Math.random() * Math.pow(16, n)).toString(16).padStart(n, "0");
+  return `${hex(8)}-${hex(4)}-${hex(4)}-${hex(4)}-${hex(12)}`;
+}
+function safeEmit(fn) {
+  try {
+    fn();
+  } catch {
+  }
+}
 
 // src/oracle.ts
 var DEFAULT_DIMENSIONS = ["correctness", "completeness", "conciseness", "format"];