eddie-jekyll 0.4.0 → 0.4.2

data/assets/eddie-agent-worker.js

@@ -1,12 +1,14 @@
 // SPDX-License-Identifier: GPL-3.0-only
 // Generated by widget/build.sh from widget/src/eddie-agent-worker.js and widget/src/lib/*.js; edit those instead.
 "use strict";
+const EDDIE_ASSET_VERSION = "c0fdc503e01a";
 const EddieLib = {};
 // SPDX-License-Identifier: GPL-3.0-only
 
-// Agent helpers shared by the widget and the agent worker: model selection,
-// prompts, plan parsing, evidence assembly and answer post-processing.
-// Pure functions; no WebLLM, no DOM.
+// Agent helpers the widget shows in its UI: model selection, evidence
+// assembly, stream display and the FAQ gate. Pure functions; no WebLLM, no
+// DOM. The prompts and answer post-processing that run beside the model
+// live in agent-llm.js (bundled only into the agent engine's hosts).
 
 (function (factory) {
   const api = factory();
@@ -21,7 +23,6 @@ const EddieLib = {};
   "use strict";
 
   const NOHIT = "The site doesn't cover that.";
-  const NOHIT_RE = /\bthe site (?:doesn['’]t|does not|didn['’]t|did not) cover (?:that|this|it)\.?/gi;
 
   const AGENT_MODEL_SIZES = {
     "Qwen3.5-0.8B": 0.4e9,
@@ -30,22 +31,6 @@ const EddieLib = {};
   };
   const TWO_GIB = 2 * 1024 * 1024 * 1024;
 
-  const PLAN_SCHEMA = {
-    type: "object",
-    properties: {
-      queries: { type: "array", items: { type: "string" }, minItems: 1, maxItems: 3 },
-    },
-    required: ["queries"],
-  };
-
-  function planPrompt(site) {
-    return `You write search queries for a site search engine. The site is ${site}. Reply with JSON only: {"queries": ["..."]}. Give 1 to 3 different short keyword queries (2 to 5 words each, no punctuation) that a site search engine would match against page text. Each query must be different. Do not answer the question.`;
-  }
-
-  function answerPrompt(site) {
-    return `You answer visitor questions about ${site} using only the numbered sources below the question. Answer the question directly in the first sentence. Write 1 to 3 sentences in your own words; never repeat a source's wording. End each sentence with the numbers of the sources it comes from, like [2] or [1][3]. Never cite a number that is not in the list. Do not add calculations or inferences that are not in the sources. If no source answers the question, your entire reply is: ${NOHIT} Never mix that sentence with an answer. Never use outside knowledge.`;
-  }
-
   /** Strip the WebLLM variant suffix to get the family name shown to visitors. */
   function baseModelId(id) {
     return String(id).replace(/-q\d+f(16|32)_\d+-MLC$/i, "");
@@ -58,7 +43,10 @@ const EddieLib = {};
 
   /**
    * Choose the WebLLM model id.
-   * opts: { mode: "auto"|"quality"|<id>, maxBufferSize, isMobile, hasF16 }
+   * opts: { mode: "auto"|"light"|"quality"|<id>, maxBufferSize, isMobile, hasF16 }
+   *
+   * "light" and "quality" are the two sizes the settings panel offers by
+   * name; "auto" picks between them from the adapter's buffer limit.
    */
   function selectAgentModel(opts) {
     const o = opts || {};
@@ -70,6 +58,8 @@ const EddieLib = {};
       base = big ? "Qwen3.5-2B" : "Qwen3.5-0.8B";
     } else if (mode === "quality") {
       base = "Qwen3.5-2B";
+    } else if (mode === "light") {
+      base = "Qwen3.5-0.8B";
     } else {
       return { id: mode, base: baseModelId(mode), sizeBytes: agentModelBytes(mode), explicit: true };
     }
@@ -85,14 +75,6 @@ const EddieLib = {};
     return /Mobi|Android|iPhone|iPad|iPod|Windows Phone/i.test(n.userAgent || "");
   }
 
-  /** Remove <think>…</think> blocks; a dangling <think> loses only the tag. */
-  function stripThink(text) {
-    if (!text) return "";
-    let out = String(text).replace(/<think>[\s\S]*?<\/think>/g, "");
-    out = out.replace(/<think>/g, "").replace(/<\/think>/g, "");
-    return out.trim();
-  }
-
   /**
    * Display text for a partial stream: complete think blocks removed, and
    * anything after an unclosed <think> hidden until it closes.
@@ -105,6 +87,123 @@ const EddieLib = {};
     return out.replace(/^\s+/, "");
   }
 
+  function urlKey(url) {
+    return String(url || "").replace(/#.*$/, "").replace(/\/+$/, "").toLowerCase();
+  }
+
+  /**
+   * Round-robin merge of several result lists, deduplicated by URL, at most
+   * `max` items. Each result keeps its own fields.
+   */
+  function mergeEvidence(lists, max) {
+    const limit = max == null ? 6 : max;
+    const seen = new Set();
+    const out = [];
+    const arrays = (lists || []).map((l) => (Array.isArray(l) ? l : []));
+    const longest = arrays.reduce((n, l) => Math.max(n, l.length), 0);
+    for (let i = 0; i < longest && out.length < limit; i++) {
+      for (const list of arrays) {
+        if (out.length >= limit) break;
+        const r = list[i];
+        if (!r || !r.url) continue;
+        const key = urlKey(r.url);
+        if (seen.has(key)) continue;
+        seen.add(key);
+        out.push(r);
+      }
+    }
+    return out;
+  }
+
+  // FAQ card gate: prefer the WASM's fused `confident` flag (qa_lookup v0.4.1+);
+  // older indexes only carry a dense score, so fall back to a plain cutoff.
+  function faqPasses(hit, qaMode) {
+    if (!hit || typeof hit !== "object") return false;
+    if (qaMode === "off") return false;
+    if (qaMode === "always") return true;
+    if (typeof hit.confident === "boolean") return hit.confident;
+    return typeof hit.score === "number" && hit.score >= 0.5;
+  }
+
+  // Turn confident QA hits into agent evidence items ("Q: … A: …") so the
+  // answer model sees the FAQ lane, not only chunk text.
+  function qaEvidence(hits, max) {
+    const limit = max == null ? 2 : max;
+    const out = [];
+    for (const h of Array.isArray(hits) ? hits : []) {
+      if (out.length >= limit) break;
+      if (!faqPasses(h, "auto")) continue;
+      const q = String(h.question || "").trim();
+      const a = String(h.answer || "").trim();
+      if (!q || !a) continue;
+      out.push({ title: "FAQ: " + q, url: h.source_url || "", text: "Q: " + q + "\nA: " + a, faq: true });
+    }
+    return out;
+  }
+
+  return {
+    faqPasses,
+    qaEvidence,
+    NOHIT,
+    AGENT_MODEL_SIZES,
+    baseModelId,
+    agentModelBytes,
+    selectAgentModel,
+    isMobileDevice,
+    visibleStreamText,
+    mergeEvidence,
+  };
+});
+
+// SPDX-License-Identifier: GPL-3.0-only
+
+// Agent LLM helpers used only where the model runs (agent-engine.js hosts:
+// the agent service worker and the page-side agent worker): prompts, plan
+// parsing and answer post-processing. Split out of agent.js so the widget
+// bundle carries only the model-selection and evidence helpers it shows in
+// the UI. Bundles that include this file always include agent.js first.
+
+(function (factory) {
+  const api = factory();
+  if (typeof module === "object" && module && module.exports) {
+    module.exports = api;
+  } else if (typeof EddieLib === "object" && EddieLib) {
+    Object.assign(EddieLib, api);
+  } else {
+    globalThis.EddieLib = Object.assign(globalThis.EddieLib || {}, api);
+  }
+})(function () {
+  "use strict";
+
+  // The fallback sentence lives in agent.js (the widget shows it too); the
+  // bundles concatenate agent.js ahead of this file, node requires it.
+  const NOHIT = (typeof module === "object" && module && module.exports ? require("./agent.js") : EddieLib).NOHIT;
+  const NOHIT_RE = /\bthe site (?:doesn['’]t|does not|didn['’]t|did not) cover (?:that|this|it)\.?/gi;
+
+  const PLAN_SCHEMA = {
+    type: "object",
+    properties: {
+      queries: { type: "array", items: { type: "string" }, minItems: 1, maxItems: 3 },
+    },
+    required: ["queries"],
+  };
+
+  function planPrompt(site) {
+    return `You write search queries for a site search engine. The site is ${site}. Reply with JSON only: {"queries": ["..."]}. Give 1 to 3 different short keyword queries (2 to 5 words each, no punctuation) that a site search engine would match against page text. Each query must be different. Do not answer the question.`;
+  }
+
+  function answerPrompt(site) {
+    return `You answer visitor questions about ${site} using only the numbered sources below the question. Answer the question directly in the first sentence. Write 1 to 3 sentences in your own words; never repeat a source's wording. End each sentence with the numbers of the sources it comes from, like [2] or [1][3]. Never cite a number that is not in the list. Do not add calculations or inferences that are not in the sources. If no source answers the question, your entire reply is: ${NOHIT} Never mix that sentence with an answer. Never use outside knowledge.`;
+  }
+
+  /** Remove <think>…</think> blocks; a dangling <think> loses only the tag. */
+  function stripThink(text) {
+    if (!text) return "";
+    let out = String(text).replace(/<think>[\s\S]*?<\/think>/g, "");
+    out = out.replace(/<think>/g, "").replace(/<\/think>/g, "");
+    return out.trim();
+  }
+
   function extractJsonObject(text) {
     const s = String(text);
     const start = s.indexOf("{");
@@ -149,34 +248,6 @@ const EddieLib = {};
     return out;
   }
 
-  function urlKey(url) {
-    return String(url || "").replace(/#.*$/, "").replace(/\/+$/, "").toLowerCase();
-  }
-
-  /**
-   * Round-robin merge of several result lists, deduplicated by URL, at most
-   * `max` items. Each result keeps its own fields.
-   */
-  function mergeEvidence(lists, max) {
-    const limit = max == null ? 6 : max;
-    const seen = new Set();
-    const out = [];
-    const arrays = (lists || []).map((l) => (Array.isArray(l) ? l : []));
-    const longest = arrays.reduce((n, l) => Math.max(n, l.length), 0);
-    for (let i = 0; i < longest && out.length < limit; i++) {
-      for (const list of arrays) {
-        if (out.length >= limit) break;
-        const r = list[i];
-        if (!r || !r.url) continue;
-        const key = urlKey(r.url);
-        if (seen.has(key)) continue;
-        seen.add(key);
-        out.push(r);
-      }
-    }
-    return out;
-  }
-
   /** Cut to `max` characters at a word boundary, with an ellipsis. */
   function truncateText(text, max) {
     const limit = max == null ? 700 : max;
@@ -266,20 +337,12 @@ const EddieLib = {};
   }
 
   return {
-    NOHIT,
-    AGENT_MODEL_SIZES,
     PLAN_SCHEMA,
     planPrompt,
     answerPrompt,
     sourcesPrompt,
-    baseModelId,
-    agentModelBytes,
-    selectAgentModel,
-    isMobileDevice,
     stripThink,
-    visibleStreamText,
     parsePlan,
-    mergeEvidence,
     truncateText,
     formatEvidence,
     postProcessAnswer,
@@ -288,249 +351,320 @@ const EddieLib = {};
 
 // SPDX-License-Identifier: GPL-3.0-only
 
-// Eddie agent worker (module worker, created on the first "Ask").
+// Eddie agent engine, host-independent.
 //
-// Runs WebLLM in the worker; retrieval stays in the widget, which owns the
-// search worker and passes evidence in. widget/build.sh concatenates
-// widget/src/lib/agent.js ahead of this file (EddieLib).
+// Everything the agent worker does (WebLLM load, plan, ask/stream, abort with
+// the drained-stream lock rule) behind an `env` the host supplies:
 //
-// Protocol (main thread -> worker):
-//   load  {model}
-//   plan  {requestId, question, site}
-//   ask   {requestId, question, site, evidence: [{title, url, text}]}
-//   abort {requestId?}
-// (worker -> main thread):
-//   progress    {text, progress}
-//   loaded      {model, loadMs}
-//   plan_result {requestId, queries, ms}
-//   token       {requestId, text}
-//   done        {requestId, answer, citations: [{n, url, title}], nohit, usage}
-//   aborted     {requestId}
-//   error       {requestId?, message}
-
-"use strict";
+//   createAgentEngine({
+//     post(message)     broadcast sink (unused today; progress goes to the loaders)
+//     loadWebLLM()      -> Promise of the WebLLM module
+//     now()             optional clock (tests)
+//   })
+//
+// `engine.handle(msg, reply)` dispatches one protocol message; `reply` is
+// the sink for that message's answers (progress/loaded for `load`,
+// plan_result, token/done/aborted for `ask`, error). Message shapes are in
+// widget/README.md ("Worker protocol", agent section).
 
-const WEBLLM_URL = "https://esm.run/@mlc-ai/web-llm@0.2.84";
-const EVIDENCE_CHARS = 700;
+(function (factory) {
+  const api = factory();
+  if (typeof module === "object" && module && module.exports) {
+    module.exports = api;
+  } else if (typeof EddieLib === "object" && EddieLib) {
+    Object.assign(EddieLib, api);
+  } else {
+    globalThis.EddieLib = Object.assign(globalThis.EddieLib || {}, api);
+  }
+})(function () {
+  "use strict";
 
-const lib = EddieLib;
+  const EVIDENCE_CHARS = 700;
+  const NOT_LOADED = "model not loaded";
+
+  function createAgentEngine(env) {
+    const lib = typeof EddieLib === "object" && EddieLib ? EddieLib : env.lib;
+    const now = env.now || (() => (typeof performance === "object" && performance.now ? performance.now() : Date.now()));
+
+    let webllm = null;
+    let engine = null;
+    let modelId = null;
+    let loading = null; // { model, promise, waiters: [reply] }
+    let active = null; // { requestId, aborted, reply }
+    let queue = Promise.resolve();
+
+    function handle(msg, reply) {
+      const m = msg || {};
+      const out = reply || env.post;
+      switch (m.type) {
+        case "load":
+          return load(m, out);
+        case "plan":
+          return enqueue(() => plan(m, out), m.requestId, out);
+        case "ask":
+          return enqueue(() => ask(m, out), m.requestId, out);
+        case "abort":
+          abort(m);
+          return Promise.resolve();
+        default:
+          postError(out, m.requestId, `unknown message type ${String(m.type)}`);
+          return Promise.resolve();
+      }
+    }
 
-let webllm = null;
-let engine = null;
-let modelId = null;
-let loading = null;
-let active = null; // { requestId, aborted }
-let queue = Promise.resolve();
+    /** Snapshot for the service worker's `state` reply. */
+    function snapshot() {
+      return {
+        model: modelId,
+        loaded: !!engine,
+        loading: loading ? loading.model : null,
+        active: active ? active.requestId : null,
+      };
+    }
 
-self.onmessage = function (e) {
-  const msg = e.data || {};
-  switch (msg.type) {
-    case "load":
-      load(msg);
-      break;
-    case "plan":
-      enqueue(() => plan(msg), msg.requestId);
-      break;
-    case "ask":
-      enqueue(() => ask(msg), msg.requestId);
-      break;
-    case "abort":
-      abort(msg);
-      break;
-    default:
-      postError(msg.requestId, `unknown message type ${String(msg.type)}`);
-  }
-};
+    function enqueue(fn, requestId, reply) {
+      queue = queue
+        .then(() => {
+          console.debug("eddie agent engine: start", requestId);
+          return fn();
+        })
+        .catch((err) => postError(reply, requestId, describe(err)))
+        .then(() => console.debug("eddie agent engine: end", requestId));
+      return queue;
+    }
 
-function enqueue(fn, requestId) {
-  queue = queue
-    .then(() => {
-      console.debug("eddie agent worker: start", requestId);
-      return fn();
-    })
-    .catch((err) => postError(requestId, describe(err)))
-    .then(() => console.debug("eddie agent worker: end", requestId));
-}
-
-async function load(msg) {
-  const model = String(msg.model || "");
-  if (!model) {
-    postError(undefined, "load: model is required");
-    return;
-  }
-  if (engine && modelId === model) {
-    self.postMessage({ type: "loaded", model, loadMs: 0, cached: true });
-    return;
-  }
-  if (loading) {
-    try {
-      await loading;
-    } catch (_) {
-      // fall through and try again
+    async function load(msg, reply) {
+      const model = String(msg.model || "");
+      if (!model) {
+        postError(reply, undefined, "load: model is required");
+        return;
+      }
+      if (engine && modelId === model) {
+        reply({ type: "loaded", model, loadMs: 0, cached: true });
+        return;
+      }
+      if (loading) {
+        // Another page (or an earlier message) is loading. Same model: join
+        // it, the fan-out delivers progress and loaded/error to us too.
+        // Different model: wait for it to settle, then load ours.
+        const join = loading.model === model;
+        if (join) loading.waiters.push(reply);
+        try {
+          await loading.promise;
+        } catch (_) {
+          // the fan-out already reported the error to the joined waiters
+        }
+        if (join) return;
+        if (engine && modelId === model) {
+          reply({ type: "loaded", model, loadMs: 0, cached: true });
+          return;
+        }
+      }
+      const job = { model, waiters: [reply], promise: null };
+      const fanout = (message) => {
+        for (const w of job.waiters) w(message);
+      };
+      job.promise = (async () => {
+        const t0 = now();
+        if (!webllm) {
+          fanout({ type: "progress", text: "Loading the WebLLM runtime…", progress: 0 });
+          webllm = await env.loadWebLLM();
+        }
+        if (engine) {
+          try {
+            await engine.unload();
+          } catch (_) {
+            // ignore
+          }
+          engine = null;
+          modelId = null;
+        }
+        const created = await webllm.CreateMLCEngine(model, {
+          initProgressCallback: (p) => {
+            fanout({
+              type: "progress",
+              text: p && p.text ? p.text : "Loading model…",
+              progress: p && typeof p.progress === "number" ? p.progress : null,
+            });
+          },
+        });
+        engine = created;
+        modelId = model;
+        fanout({ type: "loaded", model, loadMs: Math.round(now() - t0) });
+      })();
+      loading = job;
+      try {
+        await job.promise;
+      } catch (err) {
+        engine = null;
+        modelId = null;
+        fanout({ type: "error", requestId: undefined, message: describe(err) });
+      } finally {
+        if (loading === job) loading = null;
+      }
     }
-    if (engine && modelId === model) {
-      self.postMessage({ type: "loaded", model, loadMs: 0, cached: true });
-      return;
+
+    function requireEngine() {
+      if (!engine) throw new Error(NOT_LOADED);
     }
-  }
-  loading = (async () => {
-    const t0 = performance.now();
-    if (!webllm) {
-      self.postMessage({ type: "progress", text: "Loading the WebLLM runtime…", progress: 0 });
-      webllm = await import(WEBLLM_URL);
+
+    async function plan(msg, reply) {
+      requireEngine();
+      const question = String(msg.question || "").trim();
+      const site = String(msg.site || "this website");
+      const t0 = now();
+      const replyMsg = await engine.chat.completions.create({
+        messages: [
+          { role: "system", content: lib.planPrompt(site) },
+          { role: "user", content: question },
+        ],
+        temperature: 0,
+        max_tokens: 100,
+        response_format: { type: "json_object", schema: JSON.stringify(lib.PLAN_SCHEMA) },
+        extra_body: { enable_thinking: false },
+      });
+      const content = replyMsg && replyMsg.choices && replyMsg.choices[0] && replyMsg.choices[0].message ? replyMsg.choices[0].message.content : "";
+      const queries = lib.parsePlan(content, question);
+      reply({ type: "plan_result", requestId: msg.requestId, queries, ms: Math.round(now() - t0) });
     }
-    if (engine) {
+
+    async function ask(msg, reply) {
+      requireEngine();
+      const requestId = msg.requestId;
+      const question = String(msg.question || "").trim();
+      const site = String(msg.site || "this website");
+      const evidence = Array.isArray(msg.evidence) ? msg.evidence.filter((e) => e && e.url) : [];
+      if (evidence.length === 0) {
+        reply({
+          type: "done",
+          requestId,
+          answer: lib.NOHIT,
+          citations: [],
+          nohit: true,
+          raw: "",
+          usage: { ttftMs: 0, totalMs: 0, tps: null, completionTokens: 0 },
+        });
+        return;
+      }
+      active = { requestId, aborted: false, reply };
+      const t0 = now();
+      let first = 0;
+      let text = "";
+      let usage = null;
       try {
-        await engine.unload();
-      } catch (_) {
-        // ignore
+        const stream = await engine.chat.completions.create({
+          messages: [
+            { role: "system", content: lib.answerPrompt(site) },
+            { role: "user", content: lib.sourcesPrompt(evidence, question, EVIDENCE_CHARS) },
+          ],
+          stream: true,
+          stream_options: { include_usage: true },
+          temperature: 0,
+          frequency_penalty: 0.5,
+          presence_penalty: 0,
+          max_tokens: 220,
+          extra_body: { enable_thinking: false },
+        });
+        // Never break out of this loop: WebLLM releases its generation lock at
+        // the end of the async generator, and an early exit skips that release,
+        // hanging every later completion. After interruptGenerate() the stream
+        // ends by itself within one decode step; drop the tokens until then.
+        for await (const chunk of stream) {
+          if (active.aborted) continue;
+          const delta = chunk && chunk.choices && chunk.choices[0] && chunk.choices[0].delta ? chunk.choices[0].delta.content : null;
+          if (delta) {
+            if (!first) first = now();
+            text += delta;
+            reply({ type: "token", requestId, text: delta });
+          }
+          if (chunk && chunk.usage) usage = chunk.usage;
+        }
+      } finally {
+        const wasAborted = active && active.aborted;
+        active = null;
+        if (wasAborted) {
+          reply({ type: "aborted", requestId });
+          return;
+        }
       }
-      engine = null;
-      modelId = null;
+      const processed = lib.postProcessAnswer(text, evidence);
+      const totalMs = Math.round(now() - t0);
+      reply({
+        type: "done",
+        requestId,
+        answer: processed.answer,
+        citations: processed.citations,
+        nohit: processed.nohit,
+        raw: text,
+        usage: {
+          ttftMs: first ? Math.round(first - t0) : totalMs,
+          totalMs,
+          tps: usage && usage.extra && typeof usage.extra.decode_tokens_per_s === "number" ? Math.round(usage.extra.decode_tokens_per_s) : null,
+          completionTokens: usage ? usage.completion_tokens : null,
+        },
+      });
     }
-    const created = await webllm.CreateMLCEngine(model, {
-      initProgressCallback: (p) => {
-        self.postMessage({
-          type: "progress",
-          text: p && p.text ? p.text : "Loading model…",
-          progress: p && typeof p.progress === "number" ? p.progress : null,
-        });
-      },
-    });
-    engine = created;
-    modelId = model;
-    self.postMessage({ type: "loaded", model, loadMs: Math.round(performance.now() - t0) });
-  })();
-  try {
-    await loading;
-  } catch (err) {
-    engine = null;
-    modelId = null;
-    postError(undefined, describe(err));
-  } finally {
-    loading = null;
-  }
-}
-
-function requireEngine() {
-  if (!engine) throw new Error("model not loaded");
-}
-
-async function plan(msg) {
-  requireEngine();
-  const question = String(msg.question || "").trim();
-  const site = String(msg.site || "this website");
-  const t0 = performance.now();
-  const reply = await engine.chat.completions.create({
-    messages: [
-      { role: "system", content: lib.planPrompt(site) },
-      { role: "user", content: question },
-    ],
-    temperature: 0,
-    max_tokens: 100,
-    response_format: { type: "json_object", schema: JSON.stringify(lib.PLAN_SCHEMA) },
-    extra_body: { enable_thinking: false },
-  });
-  const content = reply && reply.choices && reply.choices[0] && reply.choices[0].message ? reply.choices[0].message.content : "";
-  const queries = lib.parsePlan(content, question);
-  self.postMessage({ type: "plan_result", requestId: msg.requestId, queries, ms: Math.round(performance.now() - t0) });
-}
-
-async function ask(msg) {
-  requireEngine();
-  const requestId = msg.requestId;
-  const question = String(msg.question || "").trim();
-  const site = String(msg.site || "this website");
-  const evidence = Array.isArray(msg.evidence) ? msg.evidence.filter((e) => e && e.url) : [];
-  if (evidence.length === 0) {
-    self.postMessage({
-      type: "done",
-      requestId,
-      answer: lib.NOHIT,
-      citations: [],
-      nohit: true,
-      raw: "",
-      usage: { ttftMs: 0, totalMs: 0, tps: null, completionTokens: 0 },
-    });
-    return;
-  }
-  active = { requestId, aborted: false };
-  const t0 = performance.now();
-  let first = 0;
-  let text = "";
-  let usage = null;
-  try {
-    const stream = await engine.chat.completions.create({
-      messages: [
-        { role: "system", content: lib.answerPrompt(site) },
-        { role: "user", content: lib.sourcesPrompt(evidence, question, EVIDENCE_CHARS) },
-      ],
-      stream: true,
-      stream_options: { include_usage: true },
-      temperature: 0,
-      frequency_penalty: 0.5,
-      presence_penalty: 0,
-      max_tokens: 220,
-      extra_body: { enable_thinking: false },
-    });
-    // Never break out of this loop: WebLLM releases its generation lock at
-    // the end of the async generator, and an early exit skips that release,
-    // hanging every later completion. After interruptGenerate() the stream
-    // ends by itself within one decode step; drop the tokens until then.
-    for await (const chunk of stream) {
-      if (active.aborted) continue;
-      const delta = chunk && chunk.choices && chunk.choices[0] && chunk.choices[0].delta ? chunk.choices[0].delta.content : null;
-      if (delta) {
-        if (!first) first = performance.now();
-        text += delta;
-        self.postMessage({ type: "token", requestId, text: delta });
+
+    function abort(msg) {
+      console.debug("eddie agent engine: abort", msg.requestId, active ? active.requestId : null);
+      if (!active) return;
+      if (msg.requestId != null && msg.requestId !== active.requestId) return;
+      active.aborted = true;
+      try {
+        if (engine) engine.interruptGenerate();
+      } catch (err) {
+        console.warn("eddie agent: interrupt failed", err);
       }
-      if (chunk && chunk.usage) usage = chunk.usage;
     }
-  } finally {
-    const wasAborted = active && active.aborted;
-    active = null;
-    if (wasAborted) {
-      self.postMessage({ type: "aborted", requestId });
-      return;
+
+    /** Abort the active run if it belongs to a page that went away. */
+    function abortIfOwner(reply) {
+      if (active && active.reply === reply) abort({ requestId: active.requestId });
+    }
+
+    function postError(reply, requestId, message) {
+      reply({ type: "error", requestId: requestId == null ? undefined : requestId, message });
     }
+
+    return { handle, state: snapshot, abortIfOwner };
   }
-  const processed = lib.postProcessAnswer(text, evidence);
-  const totalMs = Math.round(performance.now() - t0);
-  self.postMessage({
-    type: "done",
-    requestId,
-    answer: processed.answer,
-    citations: processed.citations,
-    nohit: processed.nohit,
-    raw: text,
-    usage: {
-      ttftMs: first ? Math.round(first - t0) : totalMs,
-      totalMs,
-      tps: usage && usage.extra && typeof usage.extra.decode_tokens_per_s === "number" ? Math.round(usage.extra.decode_tokens_per_s) : null,
-      completionTokens: usage ? usage.completion_tokens : null,
-    },
-  });
-}
-
-function abort(msg) {
-  console.debug("eddie agent worker: abort", msg.requestId, active ? active.requestId : null);
-  if (!active) return;
-  if (msg.requestId != null && msg.requestId !== active.requestId) return;
-  active.aborted = true;
-  try {
-    if (engine) engine.interruptGenerate();
-  } catch (err) {
-    console.warn("eddie agent: interrupt failed", err);
+
+  function describe(err) {
+    if (err == null) return "unknown error";
+    if (typeof err === "string") return err;
+    return err.message || String(err);
   }
-}
 
-function postError(requestId, message) {
-  self.postMessage({ type: "error", requestId: requestId == null ? undefined : requestId, message });
-}
+  /** True for the engine's "not loaded" replies (the client re-runs load). */
+  function isModelNotLoadedMessage(message) {
+    return typeof message === "string" && message.indexOf(NOT_LOADED) === 0;
+  }
+
+  return { createAgentEngine, AGENT_NOT_LOADED: NOT_LOADED, isModelNotLoadedMessage };
+});
+
+// SPDX-License-Identifier: GPL-3.0-only
+
+// Eddie agent worker (module worker, created on the first "Ask"): the
+// fallback host when the service worker (eddie-sw.js) is unavailable, same
+// protocol either way. The engine is widget/src/lib/agent-engine.js; this
+// file binds it to a dedicated module worker and loads WebLLM with a dynamic
+// import(). widget/build.sh concatenates widget/src/lib/agent.js and
+// agent-engine.js ahead of this file (EddieLib).
+//
+// Protocol: see widget/README.md ("Worker protocol", agent section).
+
+"use strict";
+
+const WEBLLM_URL = "https://esm.run/@mlc-ai/web-llm@0.2.84";
+
+const lib = EddieLib;
 
-function describe(err) {
-  if (err == null) return "unknown error";
-  if (typeof err === "string") return err;
-  return err.message || String(err);
-}
+const engine = lib.createAgentEngine({
+  post: (message) => self.postMessage(message),
+  loadWebLLM: () => import(WEBLLM_URL),
+});
+
+self.onmessage = function (e) {
+  engine.handle(e.data || {}, (message) => self.postMessage(message));
+};