polymath-agent 0.2.0 → 0.3.0

package/README.md

@@ -93,13 +93,38 @@ poly usage                                # cost by date + model
 | `poly recommend <goal>` | Pre-run recommendation: cheapest / best-value / best-quality model combos + savings. |
 | `poly models` | Browse the catalog with pricing, tier, tool support. Filters: `--tier`, `--tools`, `--search`. |
 | `poly usage` | Recorded usage & cost grouped by **date + model**. `--today`, `--since`, `--sync`. |
-| `poly analyze` | **Which approach reaches the goal with the fewest tokens** — best model per task type, objective × achievement, usage per command. |
-| `poly sync` | Push the analytics ledger to Firebase ([Data Connect SQL](dataconnect/) and/or Firestore). |
-| `poly config show\|set\|firestore\|dataconnect` | View/change settings. |
+| `poly analyze` | **Which approach reaches the goal with the fewest tokens** — efficiency playbook, best model per task type, objective × achievement, usage per command. |
+| `poly sync` | Push **distilled efficiency insights** to Firebase ([Data Connect SQL](dataconnect/) / Firestore). Raw logs stay local unless `--raw`. |
+| `poly config show\|set\|firestore\|dataconnect\|local` | View/change settings. |
 
 After each `poly run`, rate the result 0–9 (one keypress) — your goal-achievement
 rating joins the auto score (completed/planned steps) to power `poly analyze`.
 
+### The efficiency playbook (learned routing)
+
+Everything is captured locally (SQLite). `poly analyze` distills it into a **playbook**
+of *notably* efficient approaches — a (task, model) pair qualifies only with ≥3
+successful runs, ≥70% success, and **≥20% fewer tokens than the median** of its
+competitors. The playbook then **boosts routing**: proven-efficient models get
+preferred under the `value` objective (`reason: proven 54% fewer tokens on edit`).
+`poly sync` uploads *only* the playbook by default — your goals and raw logs never
+leave the machine unless you pass `--raw`.
+
+### Local LLMs (Ollama / LM Studio) — $0 routing
+
+```bash
+ollama serve                                # or LM Studio's local server
+poly config local on                        # default base: http://localhost:11434/v1
+poly config local on --base http://localhost:1234/v1   # LM Studio
+poly models -s local/                       # local models join the catalog at $0
+poly run "..."                              # cheapest objective → local wins what it can
+```
+
+Local models appear as `local/<name>`, cost $0, and need **no API key** — with
+`local on` and no OpenRouter key, Polymath runs fully offline on your machine.
+Tokens are still tracked, so the playbook learns when your local model is the
+most efficient approach.
+
 ### Routing objectives
 
 Routing is **skill-aware**: each task type maps to a skill (coding / reasoning /

package/dist/cli.js

@@ -49,6 +49,11 @@ var DEFAULT_CONFIG = {
     enabled: false,
     location: "us-east4",
     serviceId: "polymath"
+  },
+  local: {
+    enabled: false,
+    baseUrl: "http://localhost:11434/v1"
+    // Ollama default; LM Studio: http://localhost:1234/v1
   }
 };
 function loadConfig() {
@@ -60,7 +65,8 @@ function loadConfig() {
       ...DEFAULT_CONFIG,
       ...raw,
       firestore: { ...DEFAULT_CONFIG.firestore, ...raw.firestore ?? {} },
-      dataconnect: { ...DEFAULT_CONFIG.dataconnect, ...raw.dataconnect ?? {} }
+      dataconnect: { ...DEFAULT_CONFIG.dataconnect, ...raw.dataconnect ?? {} },
+      local: { ...DEFAULT_CONFIG.local, ...raw.local ?? {} }
     };
   } catch {
     return { ...DEFAULT_CONFIG };
@@ -81,6 +87,7 @@ function resolveApiKey(config) {
 
 // src/providers/openrouter.ts
 var BASE = globalThis.process?.env?.OPENROUTER_BASE_URL?.replace(/\/$/, "") || "https://openrouter.ai/api/v1";
+var LOCAL_PREFIX = "local/";
 var OpenRouterError = class extends Error {
   status;
   constructor(message, status) {
@@ -93,10 +100,12 @@ var OpenRouterClient = class {
   apiKey;
   referer;
   title;
+  localBaseUrl;
   constructor(opts = {}) {
     this.apiKey = opts.apiKey;
     this.referer = opts.referer ?? "https://github.com/polymath-agent";
     this.title = opts.title ?? "Polymath";
+    this.localBaseUrl = opts.localBaseUrl?.replace(/\/$/, "");
   }
   headers(json = true) {
     const h = {
@@ -107,6 +116,24 @@ var OpenRouterClient = class {
     if (json) h["Content-Type"] = "application/json";
     return h;
   }
+  /** Resolve where a model's request goes: the local server for `local/*`, else OpenRouter. */
+  target(modelId) {
+    if (this.localBaseUrl && modelId.startsWith(LOCAL_PREFIX)) {
+      return { base: this.localBaseUrl, model: modelId.slice(LOCAL_PREFIX.length), isLocal: true };
+    }
+    return { base: BASE, model: modelId, isLocal: false };
+  }
+  requireKeyFor(isLocal) {
+    if (!isLocal && !this.apiKey) throw new OpenRouterError("No API key set. Run `poly login`.");
+  }
+  /** List models from the local OpenAI-compatible server (Ollama / LM Studio). */
+  async listLocalRawModels() {
+    if (!this.localBaseUrl) return [];
+    const res = await fetch(`${this.localBaseUrl}/models`);
+    if (!res.ok) throw new OpenRouterError(`Local server: failed to list models (${res.status})`, res.status);
+    const json = await res.json();
+    return json.data ?? [];
+  }
   /** Raw /models payload (no auth required). */
   async listRawModels() {
     const res = await fetch(`${BASE}/models`, { headers: this.headers(false) });
@@ -126,24 +153,28 @@ var OpenRouterClient = class {
     const d = json.data ?? {};
     return { label: d.label, usage: d.usage, limit: d.limit };
   }
-  buildBody(req, stream) {
+  buildBody(req, stream, modelOverride, isLocal) {
     return {
-      model: req.model,
+      model: modelOverride,
       messages: req.messages.map(serializeMessage),
       ...req.tools && req.tools.length ? { tools: req.tools, tool_choice: "auto" } : {},
       temperature: req.temperature ?? 0.2,
       ...req.maxTokens ? { max_tokens: req.maxTokens } : {},
       stream,
-      usage: { include: true }
+      // OpenRouter-specific accounting param; local servers may reject unknown fields.
+      ...isLocal ? {} : { usage: { include: true } },
+      // OpenAI-compat way to get token usage in the final stream chunk (Ollama/LM Studio).
+      ...isLocal && stream ? { stream_options: { include_usage: true } } : {}
     };
   }
   /** Non-streaming completion. costUsd is computed from `pricing` (deterministic). */
   async complete(req, pricing) {
-    if (!this.apiKey) throw new OpenRouterError("No API key set. Run `poly login`.");
-    const res = await fetch(`${BASE}/chat/completions`, {
+    const t = this.target(req.model);
+    this.requireKeyFor(t.isLocal);
+    const res = await fetch(`${t.base}/chat/completions`, {
       method: "POST",
       headers: this.headers(),
-      body: JSON.stringify(this.buildBody(req, false))
+      body: JSON.stringify(this.buildBody(req, false, t.model, t.isLocal))
     });
     if (!res.ok) {
       const text = await res.text().catch(() => "");
@@ -164,8 +195,10 @@ var OpenRouterClient = class {
       content: typeof msg.content === "string" ? msg.content : "",
       toolCalls: parseToolCalls(msg.tool_calls),
       usage,
-      model: json.model ?? req.model,
-      costUsd: computeCost(usage, pricing, json.usage?.cost),
+      // Keep the prefixed id for local models so the ledger stays consistent.
+      model: t.isLocal ? req.model : json.model ?? req.model,
+      // Local inference is free regardless of what the server claims to report.
+      costUsd: computeCost(usage, pricing, t.isLocal ? void 0 : json.usage?.cost),
       finishReason: choice.finish_reason ?? null
     };
   }
@@ -174,11 +207,12 @@ var OpenRouterClient = class {
    * Tool-call deltas are accumulated and surfaced in the final result.
    */
   async *stream(req, pricing) {
-    if (!this.apiKey) throw new OpenRouterError("No API key set. Run `poly login`.");
-    const res = await fetch(`${BASE}/chat/completions`, {
+    const t = this.target(req.model);
+    this.requireKeyFor(t.isLocal);
+    const res = await fetch(`${t.base}/chat/completions`, {
       method: "POST",
       headers: this.headers(),
-      body: JSON.stringify(this.buildBody(req, true))
+      body: JSON.stringify(this.buildBody(req, true, t.model, t.isLocal))
     });
     if (!res.ok || !res.body) {
       const text = await res.text().catch(() => "");
@@ -212,7 +246,7 @@ var OpenRouterClient = class {
         if (evt?.error) {
           throw new OpenRouterError(evt.error.message ?? "Stream provider error", evt.error.code);
         }
-        if (evt.model) model = evt.model;
+        if (evt.model && !t.isLocal) model = evt.model;
         if (evt.usage) usageJson = evt.usage;
         const choice = evt.choices?.[0];
         if (!choice) continue;
@@ -239,17 +273,17 @@ var OpenRouterClient = class {
       completionTokens: usageJson?.completion_tokens ?? 0,
       totalTokens: usageJson?.total_tokens ?? 0
     };
-    const toolCalls = [...toolAcc.values()].filter((t) => t.name).map((t) => ({
-      id: t.id || `call_${t.name}`,
+    const toolCalls = [...toolAcc.values()].filter((t2) => t2.name).map((t2) => ({
+      id: t2.id || `call_${t2.name}`,
       type: "function",
-      function: { name: t.name, arguments: t.args || "{}" }
+      function: { name: t2.name, arguments: t2.args || "{}" }
     }));
     return {
       content,
       toolCalls,
       usage,
       model,
-      costUsd: computeCost(usage, pricing, usageJson?.cost),
+      costUsd: computeCost(usage, pricing, t.isLocal ? void 0 : usageJson?.cost),
       finishReason
     };
   }
@@ -387,6 +421,38 @@ async function getModels(client2, opts = {}) {
   return models;
 }
 
+// src/models/local.ts
+function parseLocalModels(raw) {
+  const out = [];
+  for (const m of raw) {
+    if (!m?.id) continue;
+    const name = String(m.id);
+    out.push({
+      id: LOCAL_PREFIX + name,
+      name: `${name} (local)`,
+      provider: "local",
+      contextLength: m.context_length ?? 8192,
+      pricing: { promptUsdPerMTok: 0, completionUsdPerMTok: 0 },
+      tier: classifyTier(name, 0),
+      capabilities: {
+        // OpenAI-compatible local servers pass tool schemas through; models that
+        // can't call tools simply reply with text, which the agent loop handles.
+        tools: true,
+        vision: /llava|vision|vl\b|moondream/i.test(name)
+      }
+    });
+  }
+  return out;
+}
+async function getLocalModels(client2) {
+  try {
+    const raw = await client2.listLocalRawModels();
+    return parseLocalModels(raw);
+  } catch {
+    return [];
+  }
+}
+
 // src/auth/onboarding.ts
 import readline from "node:readline";
 
@@ -723,8 +789,11 @@ var HEADLINE_SKILLS = ["coding", "reasoning", "retrieval", "speed"];
 function projectCost(m, est) {
   return est.promptTokens / 1e6 * m.pricing.promptUsdPerMTok + est.completionTokens / 1e6 * m.pricing.completionUsdPerMTok;
 }
-function taskValue(m, taskType) {
-  return taskStrength(m, taskType) / Math.max(blendedPrice(m), 0.01);
+function taskValue(m, taskType, empirical) {
+  const base = taskStrength(m, taskType) / Math.max(blendedPrice(m), 0.01);
+  const savings = empirical?.[`${taskType}:${m.id}`];
+  const boost = savings ? 1 + Math.min(savings, 100) / 100 : 1;
+  return base * boost;
 }
 function candidatesFor(taskType, models, policy, est) {
   const spec = TASK_SPECS[taskType];
@@ -753,7 +822,9 @@ function rank(models, policy, taskType) {
       break;
     case "value":
     default:
-      sorted.sort((a, b) => taskValue(b, taskType) - taskValue(a, taskType));
+      sorted.sort(
+        (a, b) => taskValue(b, taskType, policy.empirical) - taskValue(a, taskType, policy.empirical)
+      );
       break;
   }
   return sorted;
@@ -771,7 +842,8 @@ function route(taskType, models, policy, est = { promptTokens: 4e3, completionTo
   const ranked = rank(cands, policy, taskType);
   const chosen = ranked[0];
   const skill = TASK_SKILL[taskType];
-  const reason = policy.objective === "cheapest" ? `cheapest model that covers ${skill}` : policy.objective === "quality" ? `strongest at ${skill}` : `best ${skill}-per-dollar`;
+  const proven = policy.empirical?.[`${taskType}:${chosen.id}`];
+  const reason = policy.objective === "cheapest" ? `cheapest model that covers ${skill}` : policy.objective === "quality" ? `strongest at ${skill}` : proven ? `proven ${Math.round(proven)}% fewer tokens on ${taskType} (playbook)` : `best ${skill}-per-dollar`;
   return { model: chosen, reason, estCostUsd: projectCost(chosen, est) };
 }
 
@@ -1002,6 +1074,23 @@ function getDb() {
       synced INTEGER NOT NULL DEFAULT 0
     );
     CREATE INDEX IF NOT EXISTS idx_cmd_date ON command_runs(date);
+
+    -- Distilled efficiency insights: ONLY the notably cost-efficient approaches.
+    -- This is what syncs to the cloud by default (raw logs stay local).
+    CREATE TABLE IF NOT EXISTS insights (
+      id TEXT PRIMARY KEY,            -- "<task_type>__<model>"
+      computed_at INTEGER NOT NULL,
+      task_type TEXT NOT NULL,
+      model TEXT NOT NULL,
+      provider TEXT NOT NULL,
+      samples INTEGER NOT NULL,       -- successful steps observed
+      success_rate REAL NOT NULL,
+      avg_tokens REAL NOT NULL,       -- per successful step
+      baseline_tokens REAL NOT NULL,  -- median across qualified competitors
+      savings_pct REAL NOT NULL,      -- vs baseline (the "\uC720\uB3C5" margin)
+      avg_cost_usd REAL NOT NULL,
+      synced INTEGER NOT NULL DEFAULT 0
+    );
   `);
   const cols = db.prepare(`PRAGMA table_info(usage_log)`).all();
   if (!cols.some((c2) => c2.name === "command")) {
@@ -1304,9 +1393,62 @@ function unsyncedCommandRuns() {
 }
 function markTableSynced(table2, ids) {
   if (!ids.length) return;
-  const stmt = getDb().prepare(`UPDATE ${table2} SET synced=1 WHERE ${table2 === "sessions" ? "id" : "id"}=?`);
+  const stmt = getDb().prepare(`UPDATE ${table2} SET synced=1 WHERE id=?`);
   for (const id of ids) stmt.run(id);
 }
+function upsertInsight(i) {
+  getDb().prepare(
+    `INSERT INTO insights (id, computed_at, task_type, model, provider, samples, success_rate,
+         avg_tokens, baseline_tokens, savings_pct, avg_cost_usd, synced)
+       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 0)
+       ON CONFLICT(id) DO UPDATE SET
+         computed_at=excluded.computed_at, samples=excluded.samples,
+         success_rate=excluded.success_rate, avg_tokens=excluded.avg_tokens,
+         baseline_tokens=excluded.baseline_tokens, savings_pct=excluded.savings_pct,
+         avg_cost_usd=excluded.avg_cost_usd, synced=0`
+  ).run(
+    i.id,
+    i.computedAt,
+    i.taskType,
+    i.model,
+    i.provider,
+    i.samples,
+    i.successRate,
+    i.avgTokens,
+    i.baselineTokens,
+    i.savingsPct,
+    i.avgCostUsd
+  );
+}
+function deleteInsightsExcept(validIds) {
+  const all = getDb().prepare(`SELECT id FROM insights`).all();
+  const keep = new Set(validIds);
+  const del = getDb().prepare(`DELETE FROM insights WHERE id=?`);
+  for (const r of all) if (!keep.has(String(r.id))) del.run(String(r.id));
+}
+function listInsights() {
+  const rows = getDb().prepare(`SELECT * FROM insights ORDER BY savings_pct DESC`).all();
+  return rows.map(mapInsight);
+}
+function unsyncedInsights() {
+  const rows = getDb().prepare(`SELECT * FROM insights WHERE synced=0`).all();
+  return rows.map(mapInsight);
+}
+function mapInsight(r) {
+  return {
+    id: String(r.id),
+    computedAt: Number(r.computed_at),
+    taskType: String(r.task_type),
+    model: String(r.model),
+    provider: String(r.provider),
+    samples: Number(r.samples),
+    successRate: Number(r.success_rate),
+    avgTokens: Number(r.avg_tokens),
+    baselineTokens: Number(r.baseline_tokens),
+    savingsPct: Number(r.savings_pct),
+    avgCostUsd: Number(r.avg_cost_usd)
+  };
+}
 
 // src/usage/report.ts
 function renderUsageReport(filter = {}) {
@@ -1349,6 +1491,80 @@ function renderUsageReport(filter = {}) {
   ].join("\n");
 }
 
+// src/usage/insights.ts
+var MIN_SAMPLES = 3;
+var MIN_SUCCESS = 0.7;
+var MIN_MARGIN = 0.2;
+function median(xs) {
+  const s = [...xs].sort((a, b) => a - b);
+  const mid = Math.floor(s.length / 2);
+  return s.length % 2 ? s[mid] : (s[mid - 1] + s[mid]) / 2;
+}
+function distillInsights(now = Date.now()) {
+  const eff = modelTaskEfficiency();
+  const byTask = /* @__PURE__ */ new Map();
+  for (const r of eff) {
+    const list = byTask.get(r.taskType) ?? [];
+    list.push(r);
+    byTask.set(r.taskType, list);
+  }
+  const valid = [];
+  for (const [taskType, list] of byTask) {
+    const qualified = list.filter(
+      (r) => r.steps >= MIN_SAMPLES && r.successRate >= MIN_SUCCESS && r.avgTokensPerSuccess > 0
+    );
+    if (qualified.length < 2) continue;
+    const baseline = median(qualified.map((r) => r.avgTokensPerSuccess));
+    for (const r of qualified) {
+      const savings = 1 - r.avgTokensPerSuccess / baseline;
+      if (savings >= MIN_MARGIN) {
+        const id = `${taskType}__${r.model}`;
+        valid.push(id);
+        upsertInsight({
+          id,
+          computedAt: now,
+          taskType,
+          model: r.model,
+          provider: r.model.split("/")[0] ?? "unknown",
+          samples: r.steps,
+          successRate: r.successRate,
+          avgTokens: r.avgTokensPerSuccess,
+          baselineTokens: baseline,
+          savingsPct: savings * 100,
+          avgCostUsd: r.avgCostPerSuccess
+        });
+      }
+    }
+  }
+  deleteInsightsExcept(valid);
+  return listInsights();
+}
+function insightBoostMap(insights) {
+  const map = {};
+  for (const i of insights) map[`${i.taskType}:${i.model}`] = i.savingsPct;
+  return map;
+}
+function renderPlaybook(insights) {
+  if (!insights.length) {
+    return c.bold("Efficiency playbook") + "\n" + c.dim(
+      `Nothing distilled yet \u2014 needs \u22652 models with \u2265${MIN_SAMPLES} successful steps on the same task type,
+where one beats the median by \u2265${MIN_MARGIN * 100}% tokens. Keep running tasks (vary models with -o / pins).`
+    );
+  }
+  return c.bold("Efficiency playbook") + c.dim("  (the notably efficient approaches \u2014 this is what `poly sync` uploads)") + "\n" + table(
+    ["Task", "Model", "Avg tok", "Baseline", "Savings", "Success", "n"],
+    insights.map((i) => [
+      i.taskType,
+      c.green(i.model),
+      tokens(Math.round(i.avgTokens)),
+      tokens(Math.round(i.baselineTokens)),
+      c.green(`-${i.savingsPct.toFixed(0)}%`),
+      `${Math.round(i.successRate * 100)}%`,
+      String(i.samples)
+    ])
+  );
+}
+
 // src/usage/analyze.ts
 var MIN_SUCCESS_RATE = 0.5;
 function renderAnalysis(filter = {}) {
@@ -1359,6 +1575,9 @@ function renderAnalysis(filter = {}) {
   if (!byModelTask.length && !byObjective.length && !byCommand.length) {
     return c.dim('No analytics yet. Run `poly run "<task>"` a few times (and rate the result) first.');
   }
+  const insights = distillInsights();
+  out.push(renderPlaybook(insights));
+  out.push("");
   if (byModelTask.length) {
     const byTask = /* @__PURE__ */ new Map();
     for (const r of byModelTask) {
@@ -1445,7 +1664,7 @@ function renderAnalysis(filter = {}) {
 }
 
 // src/usage/firestoreSync.ts
-async function syncUsage(config) {
+async function syncUsage(config, opts = {}) {
   if (!config.firestore.enabled) {
     return { synced: 0, message: "Firestore sync is disabled (enable with `poly config firestore on`)." };
   }
@@ -1473,28 +1692,62 @@ async function syncUsage(config) {
     }
   }
   const fdb = fsMod.getFirestore();
+  distillInsights();
+  const insights = unsyncedInsights();
+  if (insights.length) {
+    const batch = fdb.batch();
+    const col = fdb.collection("polymath_insights");
+    for (const i of insights) {
+      batch.set(col.doc(i.id), {
+        computedAt: i.computedAt,
+        taskType: i.taskType,
+        model: i.model,
+        provider: i.provider,
+        samples: i.samples,
+        successRate: i.successRate,
+        avgTokens: i.avgTokens,
+        baselineTokens: i.baselineTokens,
+        savingsPct: i.savingsPct,
+        avgCostUsd: i.avgCostUsd
+      });
+    }
+    await batch.commit();
+    markTableSynced("insights", insights.map((i) => i.id));
+  }
+  if (!opts.raw) {
+    return {
+      synced: insights.length,
+      message: insights.length ? `Synced ${insights.length} efficiency insight(s) to polymath_insights. Raw logs stayed local (use --raw to push).` : "No new insights to sync \u2014 raw logs stay local by default (use --raw to push them)."
+    };
+  }
   const rows = unsyncedRows();
-  if (!rows.length) return { synced: 0, message: "Nothing to sync \u2014 all rows already pushed." };
-  const batch = fdb.batch();
-  const col = fdb.collection(config.firestore.collection);
-  for (const r of rows) {
-    const ref = col.doc(`${r.date}__${r.id}`);
-    batch.set(ref, {
-      ts: r.ts,
-      date: r.date,
-      provider: r.provider,
-      model: r.model,
-      taskType: r.taskType,
-      promptTokens: r.promptTokens,
-      completionTokens: r.completionTokens,
-      totalTokens: r.totalTokens,
-      costUsd: r.costUsd,
-      sessionId: r.sessionId ?? null
-    });
+  if (!rows.length && !insights.length) return { synced: 0, message: "Nothing to sync \u2014 all rows already pushed." };
+  if (rows.length) {
+    const batch = fdb.batch();
+    const col = fdb.collection(config.firestore.collection);
+    for (const r of rows) {
+      const ref = col.doc(`${r.date}__${r.id}`);
+      batch.set(ref, {
+        ts: r.ts,
+        date: r.date,
+        provider: r.provider,
+        model: r.model,
+        taskType: r.taskType,
+        command: r.command ?? "run",
+        promptTokens: r.promptTokens,
+        completionTokens: r.completionTokens,
+        totalTokens: r.totalTokens,
+        costUsd: r.costUsd,
+        sessionId: r.sessionId ?? null
+      });
+    }
+    await batch.commit();
+    markSynced(rows.map((r) => r.id));
   }
-  await batch.commit();
-  markSynced(rows.map((r) => r.id));
-  return { synced: rows.length, message: `Synced ${rows.length} rows to ${config.firestore.collection}.` };
+  return {
+    synced: insights.length + rows.length,
+    message: `Synced ${insights.length} insights + ${rows.length} raw rows to Firestore.`
+  };
 }
 
 // src/usage/dataconnect.ts
@@ -1539,14 +1792,56 @@ async function executeGraphql(cfg2, token, query, variables) {
   }
 }
 var iso = (ms) => new Date(ms).toISOString();
-async function syncDataConnect(config) {
+async function syncDataConnect(config, opts = {}) {
   const dc = config.dataconnect;
   if (!dc?.enabled) {
-    return { sessions: 0, steps: 0, commands: 0, calls: 0, message: "Data Connect sync is disabled (enable with `poly config dataconnect on`)." };
+    return { insights: 0, sessions: 0, steps: 0, commands: 0, calls: 0, message: "Data Connect sync is disabled (enable with `poly config dataconnect on`)." };
   }
   const projectId = config.firestore.projectId;
   const token = await adminAccessToken(projectId);
   const cfg2 = { projectId, location: dc.location, serviceId: dc.serviceId };
+  distillInsights();
+  const insights = unsyncedInsights();
+  for (const i of insights) {
+    await executeGraphql(
+      cfg2,
+      token,
+      `mutation UpsertInsight($id: String!, $computedAt: Timestamp!, $taskType: String!,
+         $model: String!, $provider: String!, $samples: Int!, $successRate: Float!,
+         $avgTokens: Float!, $baselineTokens: Float!, $savingsPct: Float!, $avgCostUsd: Float!) {
+         insight_upsert(data: {
+           id: $id, computedAt: $computedAt, taskType: $taskType, model: $model,
+           provider: $provider, samples: $samples, successRate: $successRate,
+           avgTokens: $avgTokens, baselineTokens: $baselineTokens,
+           savingsPct: $savingsPct, avgCostUsd: $avgCostUsd
+         })
+       }`,
+      {
+        id: i.id,
+        computedAt: iso(i.computedAt),
+        taskType: i.taskType,
+        model: i.model,
+        provider: i.provider,
+        samples: i.samples,
+        successRate: i.successRate,
+        avgTokens: i.avgTokens,
+        baselineTokens: i.baselineTokens,
+        savingsPct: i.savingsPct,
+        avgCostUsd: i.avgCostUsd
+      }
+    );
+  }
+  markTableSynced("insights", insights.map((i) => i.id));
+  if (!opts.raw) {
+    return {
+      insights: insights.length,
+      sessions: 0,
+      steps: 0,
+      commands: 0,
+      calls: 0,
+      message: `Synced ${insights.length} efficiency insight(s) to Data Connect (${cfg2.serviceId}@${cfg2.location}). Raw logs stayed local \u2014 use \`poly sync --raw\` to push everything.`
+    };
+  }
   const sessions = unsyncedSessions();
   for (const s of sessions) {
     await executeGraphql(
@@ -1679,11 +1974,12 @@ async function syncDataConnect(config) {
   }
   markSynced(calls.map((c2) => c2.id));
   return {
+    insights: insights.length,
     sessions: sessions.length,
     steps: steps.length,
     commands: commands.length,
     calls: calls.length,
-    message: `Synced ${sessions.length} sessions, ${steps.length} steps, ${commands.length} commands, ${calls.length} calls to Data Connect (${cfg2.serviceId}@${cfg2.location}).`
+    message: `Synced ${insights.length} insights + raw: ${sessions.length} sessions, ${steps.length} steps, ${commands.length} commands, ${calls.length} calls (${cfg2.serviceId}@${cfg2.location}).`
   };
 }
 
@@ -2274,21 +2570,30 @@ function truncate2(s, n) {
 
 // src/index.ts
 var program = new Command();
-program.name("poly").description("Polymath \u2014 cost-optimized, multi-model TUI coding agent").version("0.2.0");
+program.name("poly").description("Polymath \u2014 cost-optimized, multi-model TUI coding agent").version("0.3.0");
 function client(config) {
   return new OpenRouterClient({
     apiKey: resolveApiKey(config),
     referer: config.referer,
-    title: config.title
+    title: config.title,
+    localBaseUrl: config.local.enabled ? config.local.baseUrl : void 0
   });
 }
 function buildPolicy(config, opts) {
   const objective = opts.objective || config.defaultObjective;
   const maxCost = opts.maxCost != null ? parseFloat(opts.maxCost) : config.maxCostPerCallUsd;
+  let empirical;
+  try {
+    empirical = insightBoostMap(listInsights());
+    if (!Object.keys(empirical).length) empirical = void 0;
+  } catch {
+    empirical = void 0;
+  }
   return {
     objective,
     maxCostPerCallUsd: Number.isFinite(maxCost) ? maxCost : void 0,
-    pinned: config.pinned
+    pinned: config.pinned,
+    empirical
   };
 }
 function localDate3(d = /* @__PURE__ */ new Date()) {
@@ -2315,9 +2620,23 @@ function trackCommand(opts) {
   }
 }
 async function loadCatalog(config, refresh = false) {
-  const models = await getModels(client(config), { refresh });
+  const cl = client(config);
+  const hasKey = !!resolveApiKey(config);
+  let models = [];
+  try {
+    models = await getModels(cl, { refresh });
+  } catch (e) {
+    if (!config.local.enabled) throw e;
+  }
+  if (config.local.enabled) {
+    const local = await getLocalModels(cl);
+    if (!local.length) {
+      console.error(c.yellow(`Local server (${config.local.baseUrl}) returned no models \u2014 is it running?`));
+    }
+    models = hasKey ? [...local, ...models] : local;
+  }
   if (!models.length) {
-    console.error(c.red("Could not load the model catalog. Check your connection."));
+    console.error(c.red("No models available. Check your connection, or `poly config local on` with a running Ollama/LM Studio."));
     process.exit(1);
   }
   return models;
@@ -2328,10 +2647,12 @@ program.command("login").description("Connect Polymath to OpenRouter (set/replac
 program.command("run", { isDefault: true }).description("Launch the interactive agent (TUI)").argument("[goal...]", "what to do (optional; prompts if omitted)").option("-o, --objective <name>", "routing objective: cheapest | value | quality").option("--max-cost <usd>", "exclude models whose projected per-call cost exceeds this").option("-w, --write", "allow the agent to write files (confined to --cwd)", false).option("-x, --commands", "DANGER: let the model run arbitrary shell commands in --cwd", false).option("-C, --cwd <dir>", "working directory", process.cwd()).action(async (goalParts, opts) => {
   const startedAt = Date.now();
   const config = loadConfig();
-  const key = await ensureApiKey(config);
-  if (!key) {
-    console.error(c.red("No API key \u2014 cannot run. Try `poly login`."));
-    process.exit(1);
+  if (!config.local.enabled || resolveApiKey(config)) {
+    const key = await ensureApiKey(config);
+    if (!key && !config.local.enabled) {
+      console.error(c.red("No API key \u2014 cannot run. Try `poly login`, or `poly config local on` for a local LLM."));
+      process.exit(1);
+    }
   }
   const reloaded = loadConfig();
   const models = await loadCatalog(reloaded);
@@ -2434,21 +2755,22 @@ program.command("usage").description("Show recorded usage & cost by date + model
 program.command("analyze").description("Which approach reaches the goal with the FEWEST tokens \u2014 per model, task, objective, command").option("--since <date>", "YYYY-MM-DD inclusive").option("--until <date>", "YYYY-MM-DD inclusive").action(async (opts) => {
   console.log(renderAnalysis({ since: opts.since, until: opts.until }));
 });
-program.command("sync").description("Push the local analytics ledger to Firebase (Data Connect SQL and/or Firestore)").action(async () => {
+program.command("sync").description("Push DISTILLED efficiency insights to Firebase (raw logs stay local unless --raw)").option("--raw", "also push the full raw ledger (sessions/steps/calls/commands)", false).action(async (opts) => {
   const config = loadConfig();
   let pushed = false;
   if (config.dataconnect.enabled) {
     pushed = true;
     try {
-      const res = await syncDataConnect(config);
-      console.log(res.sessions + res.steps + res.commands + res.calls > 0 ? c.green(res.message) : c.dim(res.message));
+      const res = await syncDataConnect(config, { raw: !!opts.raw });
+      const n = res.insights + res.sessions + res.steps + res.commands + res.calls;
+      console.log(n > 0 ? c.green(res.message) : c.dim(res.message));
     } catch (e) {
       console.error(c.red(`Data Connect sync failed: ${e?.message ?? e}`));
     }
   }
   if (config.firestore.enabled) {
     pushed = true;
-    const res = await syncUsage(config);
+    const res = await syncUsage(config, { raw: !!opts.raw });
     console.log(res.synced > 0 ? c.green(res.message) : c.dim(res.message));
   }
   if (!pushed) {
@@ -2499,6 +2821,17 @@ cfg.command("firestore").description("Enable/disable Firestore sync: on | off").
   saveConfig(config);
   console.log(c.green(`Firestore sync ${config.firestore.enabled ? "enabled" : "disabled"}.`));
 });
+cfg.command("local").description("Enable/disable a local LLM server (Ollama/LM Studio): on | off [--base <url>]").argument("<state>").option("--base <url>", "OpenAI-compatible base URL (default http://localhost:11434/v1)").action((state, opts) => {
+  const config = loadConfig();
+  config.local.enabled = /^on|true|1$/i.test(state);
+  if (opts.base) config.local.baseUrl = String(opts.base).replace(/\/$/, "");
+  saveConfig(config);
+  console.log(
+    c.green(
+      `Local LLM ${config.local.enabled ? "enabled" : "disabled"} (${config.local.baseUrl}). Models appear as local/<name> with $0 cost.`
+    )
+  );
+});
 cfg.command("dataconnect").description("Enable/disable Firebase Data Connect (SQL) sync: on | off [--location <loc>] [--service <id>]").argument("<state>").option("--location <loc>", "Data Connect location (default us-east4)").option("--service <id>", "Data Connect service id (default polymath)").action((state, opts) => {
   const config = loadConfig();
   config.dataconnect.enabled = /^on|true|1$/i.test(state);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polymath-agent",
-  "version": "0.2.0",
+  "version": "0.3.0",
   "description": "Polymath — a cost-optimized, multi-model TUI coding agent. Decomposes work into typed tasks, routes each task to the cheapest capable model via OpenRouter, and logs real usage/cost by date + model.",
   "type": "module",
   "bin": {