@houtini/lm 2.9.0 → 2.11.0

package/dist/index.js

@@ -8,7 +8,7 @@
 import { Server } from '@modelcontextprotocol/sdk/server/index.js';
 import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
 import { CallToolRequestSchema, ListToolsRequestSchema, ListResourcesRequestSchema, ReadResourceRequestSchema, } from '@modelcontextprotocol/sdk/types.js';
-import { profileModelsAtStartup, getCachedProfile, toModelProfile as cachedToProfile, getHFEnrichmentLine, getPromptHints, getThinkingSupport, } from './model-cache.js';
+import { profileModelsAtStartup, getCachedProfile, toModelProfile as cachedToProfile, getHFEnrichmentLine, getPromptHints, getThinkingSupport, recordPerformance, getAllPerformance, getLifetimeTotals, } from './model-cache.js';
 import { readFile } from 'node:fs/promises';
 import { isAbsolute, basename } from 'node:path';
 const LM_BASE_URL = process.env.LM_STUDIO_URL || 'http://localhost:1234';
@@ -19,7 +19,9 @@ const DEFAULT_TEMPERATURE = 0.3;
 const CONNECT_TIMEOUT_MS = 5000;
 const INFERENCE_CONNECT_TIMEOUT_MS = 30_000; // generous connect timeout for inference
 const SOFT_TIMEOUT_MS = 300_000; // 5 min — progress notifications reset MCP client timeout, so this is a safety net not the primary limit
-const READ_CHUNK_TIMEOUT_MS = 30_000; // max wait for a single SSE chunk
+const READ_CHUNK_TIMEOUT_MS = 30_000; // max wait for a single SSE chunk mid-stream
+const PREFILL_TIMEOUT_MS = 180_000; // max wait for the FIRST chunk — prompt prefill on slow hardware with big inputs can legitimately take 1-2 min
+const PREFILL_KEEPALIVE_MS = 10_000; // fire a progress notification every N ms while waiting for prefill to finish
 const FALLBACK_CONTEXT_LENGTH = parseInt(process.env.LM_CONTEXT_WINDOW || '100000', 10);
 // ── Session-level token accounting ───────────────────────────────────
 // Tracks cumulative tokens offloaded to the local LLM across all calls
@@ -32,37 +34,145 @@ const session = {
     /** Per-model performance tracking for routing insights */
     modelStats: new Map(),
 };
+// Lifetime mirror — kept in sync with the SQLite `model_performance` table
+// so the footer/discover path stays synchronous. Hydrated once at startup
+// from `getAllPerformance()`, then updated in-memory alongside every DB
+// write in `recordUsage`. Also updated after the async DB write completes
+// so counters can only ever run a tick behind, never ahead.
+const lifetime = {
+    totalCalls: 0,
+    totalTokens: 0,
+    modelsUsed: 0,
+    firstSeenAt: null,
+    /** Per-model lifetime stats — same shape as session.modelStats for easy formatting. */
+    modelStats: new Map(),
+};
+async function hydrateLifetimeFromDb() {
+    try {
+        const totals = await getLifetimeTotals();
+        lifetime.totalCalls = totals.totalCalls;
+        lifetime.totalTokens = totals.totalTokens;
+        lifetime.modelsUsed = totals.modelsUsed;
+        lifetime.firstSeenAt = totals.firstSeenAt;
+        const rows = await getAllPerformance();
+        lifetime.modelStats.clear();
+        for (const r of rows) {
+            lifetime.modelStats.set(r.modelId, {
+                calls: r.totalCalls,
+                ttftCalls: r.ttftCalls,
+                perfCalls: r.perfCalls,
+                totalTtftMs: r.totalTtftMs,
+                totalTokPerSec: r.totalTokPerSec,
+                totalPromptTokens: r.totalPromptTokens,
+                firstSeenAt: r.firstSeenAt,
+                lastUsedAt: r.lastUsedAt,
+            });
+        }
+    }
+    catch (err) {
+        process.stderr.write(`[houtini-lm] Lifetime hydration failed (stats will build from this session): ${err}\n`);
+    }
+}
 function recordUsage(resp) {
     session.calls++;
+    const promptTokens = resp.usage?.prompt_tokens ?? 0;
+    let completionTokens = resp.usage?.completion_tokens ?? 0;
+    const reasoningTokens = resp.usage?.completion_tokens_details?.reasoning_tokens ?? 0;
     if (resp.usage) {
-        session.promptTokens += resp.usage.prompt_tokens;
-        session.completionTokens += resp.usage.completion_tokens;
+        session.promptTokens += promptTokens;
+        session.completionTokens += completionTokens;
     }
     else if (resp.content.length > 0) {
         // Estimate when usage is missing (truncated responses)
-        session.completionTokens += Math.ceil(resp.content.length / 4);
+        const est = Math.ceil(resp.content.length / 4);
+        completionTokens = est;
+        session.completionTokens += est;
     }
-    // Track per-model perf stats
+    // Tok/s used by both session and lifetime stats
+    const tokPerSec = resp.usage && resp.generationMs > 50
+        ? (resp.usage.completion_tokens / (resp.generationMs / 1000))
+        : 0;
+    // Session per-model (unchanged behaviour)
     if (resp.model) {
-        const existing = session.modelStats.get(resp.model) || { calls: 0, perfCalls: 0, totalTtftMs: 0, totalTokPerSec: 0 };
+        const existing = session.modelStats.get(resp.model) || { calls: 0, ttftCalls: 0, perfCalls: 0, totalTtftMs: 0, totalTokPerSec: 0 };
         existing.calls++;
-        if (resp.ttftMs)
+        if (resp.ttftMs) {
             existing.totalTtftMs += resp.ttftMs;
-        const tokPerSec = resp.usage && resp.generationMs > 50
-            ? (resp.usage.completion_tokens / (resp.generationMs / 1000))
-            : 0;
+            existing.ttftCalls++;
+        }
         if (tokPerSec > 0) {
             existing.perfCalls++;
             existing.totalTokPerSec += tokPerSec;
         }
         session.modelStats.set(resp.model, existing);
     }
+    // Lifetime mirror + SQLite write — fire-and-forget so a DB hiccup can't
+    // stall a tool response. The in-memory mirror is updated synchronously so
+    // the footer and discover output reflect this call immediately.
+    if (resp.model && (promptTokens > 0 || completionTokens > 0)) {
+        const now = Date.now();
+        const wasFirstEver = !lifetime.modelStats.has(resp.model);
+        const lExisting = lifetime.modelStats.get(resp.model) || {
+            calls: 0, ttftCalls: 0, perfCalls: 0, totalTtftMs: 0, totalTokPerSec: 0, totalPromptTokens: 0,
+            firstSeenAt: now, lastUsedAt: now,
+        };
+        lExisting.calls++;
+        if (resp.ttftMs) {
+            lExisting.totalTtftMs += resp.ttftMs;
+            lExisting.ttftCalls++;
+        }
+        if (tokPerSec > 0) {
+            lExisting.perfCalls++;
+            lExisting.totalTokPerSec += tokPerSec;
+        }
+        lExisting.totalPromptTokens += promptTokens;
+        lExisting.lastUsedAt = now;
+        lifetime.modelStats.set(resp.model, lExisting);
+        lifetime.totalCalls++;
+        lifetime.totalTokens += promptTokens + completionTokens;
+        if (wasFirstEver) {
+            lifetime.modelsUsed++;
+            if (lifetime.firstSeenAt === null)
+                lifetime.firstSeenAt = now;
+        }
+        recordPerformance(resp.model, {
+            ttftMs: resp.ttftMs,
+            tokPerSec: tokPerSec > 0 ? tokPerSec : undefined,
+            promptTokens,
+            completionTokens,
+            reasoningTokens,
+        }).catch((err) => {
+            process.stderr.write(`[houtini-lm] Performance write failed (continuing): ${err}\n`);
+        });
+    }
 }
 function sessionSummary() {
     const total = session.promptTokens + session.completionTokens;
-    if (session.calls === 0)
+    if (session.calls === 0 && lifetime.totalCalls === 0)
         return '';
-    return `Session: ${total.toLocaleString()} tokens offloaded across ${session.calls} call${session.calls === 1 ? '' : 's'}`;
+    const callWord = (n) => (n === 1 ? 'call' : 'calls');
+    const sessionPart = session.calls > 0
+        ? `this session: ${total.toLocaleString()} tokens / ${session.calls} ${callWord(session.calls)}`
+        : 'this session: 0 tokens';
+    // Lifetime numbers only show once there's something in the DB — avoids a
+    // confusing "lifetime: 0" on a truly fresh install.
+    if (lifetime.totalCalls > 0) {
+        return `💰 Claude quota saved — ${sessionPart} · lifetime: ${lifetime.totalTokens.toLocaleString()} tokens / ${lifetime.totalCalls} ${callWord(lifetime.totalCalls)}`;
+    }
+    return `💰 Claude quota saved ${sessionPart}`;
+}
+/**
+ * Return true when this response is the first one with measurable perf stats
+ * for its model in the current session. Used to surface a one-off "benchmarked"
+ * line so Claude sees the real speed of the local model on a genuine task,
+ * not an artificial warmup.
+ */
+function isFirstBenchmarkedCall(modelId, tokPerSec) {
+    if (!modelId || tokPerSec <= 0)
+        return false;
+    const stats = session.modelStats.get(modelId);
+    // After recordUsage has run, perfCalls === 1 means this was the first measured call.
+    return !!stats && stats.perfCalls === 1;
 }
 function apiHeaders() {
     const h = { 'Content-Type': 'application/json' };
@@ -365,6 +475,10 @@ async function chatCompletionStreamingInner(messages, options = {}) {
         messages,
         temperature: options.temperature ?? DEFAULT_TEMPERATURE,
         max_tokens: effectiveMaxTokens,
+        // Send max_completion_tokens alongside max_tokens for OpenAI reasoning-model
+        // compatibility (OpenAI spec distinguishes total generation cap from visible
+        // output cap). Backends that don't understand it ignore unknown fields.
+        max_completion_tokens: effectiveMaxTokens,
         stream: true,
         stream_options: { include_usage: true },
     };
@@ -375,23 +489,36 @@ async function chatCompletionStreamingInner(messages, options = {}) {
         body.response_format = options.responseFormat;
     }
     // Handle thinking/reasoning models.
-    // Some models (Gemma 4, Qwen3, DeepSeek) have extended thinking that consumes
-    // part of the max_tokens budget for invisible reasoning before producing content.
-    // Strategy: try to disable thinking via enable_thinking=false, BUT also inflate
-    // max_tokens as a safety net since some models (Gemma 4) hardcode thinking=true
-    // in their Jinja template and ignore the API parameter.
+    // Some models (Gemma 4, Qwen3, DeepSeek R1, Nemotron, gpt-oss) have extended
+    // thinking that consumes part of the max_tokens budget for invisible reasoning
+    // before producing content. Strategy:
+    //   1. reasoning_effort=<family-specific value> to minimise reasoning
+    //   2. enable_thinking:false — Qwen3 vendor param (ignored elsewhere)
+    //   3. inflate max_tokens 4× — safety net when both flags are ignored
+    //      (e.g. Gemma 4 hardcodes enable_thinking=true in its Jinja template)
+    //
+    // IMPORTANT: reasoning_effort values are NOT standard. OpenAI/gpt-oss use
+    // 'low'|'medium'|'high'; Ollama adds 'none'; LM Studio's Nemotron adapter
+    // only accepts 'on'|'off'. Sending 'low' to Nemotron causes LM Studio to
+    // silently fall back to 'on' — maximising reasoning, the OPPOSITE of intent.
+    // Hence the family-specific mapping below. When uncertain, we omit the
+    // field entirely rather than risk a bad-value fallback.
     const modelId = (options.model || LM_MODEL || '').toString();
     if (modelId) {
         const thinking = await getThinkingSupport(modelId);
         if (thinking?.supportsThinkingToggle) {
             body.enable_thinking = false;
-            // Safety net: inflate max_tokens to account for reasoning budget.
-            // Gemma 4 ignores enable_thinking=false (hardcoded in template),
-            // so the model will think regardless. Without inflation, reasoning
-            // consumes all tokens and content comes back empty.
-            const requestedTokens = (options.maxTokens ?? DEFAULT_MAX_TOKENS);
-            body.max_tokens = Math.max(requestedTokens * 4, requestedTokens + 2000);
-            process.stderr.write(`[houtini-lm] Thinking model ${modelId}: enable_thinking=false, max_tokens inflated ${requestedTokens} → ${body.max_tokens}\n`);
+            const reasoningValue = getReasoningEffortValue(modelId);
+            if (reasoningValue !== null) {
+                body.reasoning_effort = reasoningValue;
+            }
+            // Inflation uses effectiveMaxTokens (the context-aware value), not
+            // DEFAULT_MAX_TOKENS — otherwise big-context models get sized down.
+            const beforeInflation = effectiveMaxTokens;
+            const inflated = Math.max(beforeInflation * 4, beforeInflation + 2000);
+            body.max_tokens = inflated;
+            body.max_completion_tokens = inflated;
+            process.stderr.write(`[houtini-lm] Thinking model ${modelId}: reasoning_effort=${reasoningValue ?? '(omitted)'}, enable_thinking=false, max_tokens inflated ${beforeInflation} → ${inflated}\n`);
         }
     }
     const startTime = Date.now();
@@ -406,13 +533,39 @@ async function chatCompletionStreamingInner(messages, options = {}) {
     const reader = res.body.getReader();
     const decoder = new TextDecoder();
     let content = '';
-    let chunkCount = 0;
+    let reasoning = '';
+    let progressSeq = 0;
     let model = '';
     let usage;
     let finishReason = '';
     let truncated = false;
+    let prefillStall = false;
     let buffer = '';
     let ttftMs;
+    let firstChunkReceived = false;
+    // Prefill keep-alive — /v1/chat/completions gives no SSE events during
+    // prompt processing, so the MCP client clock ticks uninterrupted on a slow
+    // backend with a big input. Fire a progress notification every 10s until
+    // the first chunk arrives to keep the client from timing out at 60s.
+    const sendProgress = (message) => {
+        if (options.progressToken === undefined)
+            return;
+        progressSeq++;
+        server.notification({
+            method: 'notifications/progress',
+            params: {
+                progressToken: options.progressToken,
+                progress: progressSeq,
+                message,
+            },
+        }).catch(() => { });
+    };
+    const keepAliveTimer = setInterval(() => {
+        if (firstChunkReceived)
+            return;
+        const waitedMs = Date.now() - startTime;
+        sendProgress(`Waiting for model... (${(waitedMs / 1000).toFixed(0)}s, still in prefill)`);
+    }, PREFILL_KEEPALIVE_MS);
     try {
         while (true) {
             // Check soft timeout before each read
@@ -422,17 +575,24 @@ async function chatCompletionStreamingInner(messages, options = {}) {
                 process.stderr.write(`[houtini-lm] Soft timeout at ${elapsed}ms, returning ${content.length} chars of partial content\n`);
                 break;
             }
-            // Read with per-chunk timeout (handles stalled generation)
+            // Split prefill vs mid-stream timeouts. Prefill on slow hardware with
+            // a 7k-token input can legitimately take 1-2 min; mid-stream stalls
+            // should surface much faster. Track firstChunkReceived to switch.
             const remaining = SOFT_TIMEOUT_MS - elapsed;
-            const chunkTimeout = Math.min(READ_CHUNK_TIMEOUT_MS, remaining);
+            const perChunkCeiling = firstChunkReceived ? READ_CHUNK_TIMEOUT_MS : PREFILL_TIMEOUT_MS;
+            const chunkTimeout = Math.min(perChunkCeiling, remaining);
             const result = await timedRead(reader, chunkTimeout);
             if (result === 'timeout') {
                 truncated = true;
-                process.stderr.write(`[houtini-lm] Chunk read timeout, returning ${content.length} chars of partial content\n`);
+                prefillStall = !firstChunkReceived;
+                process.stderr.write(`[houtini-lm] ${prefillStall ? 'Prefill' : 'Mid-stream'} timeout, returning ${content.length} chars of partial content\n`);
                 break;
             }
             if (result.done)
                 break;
+            if (!firstChunkReceived) {
+                firstChunkReceived = true;
+            }
             buffer += decoder.decode(result.value, { stream: true });
             // Parse SSE lines
             const lines = buffer.split('\n');
@@ -448,41 +608,20 @@ async function chatCompletionStreamingInner(messages, options = {}) {
                     if (json.model)
                         model = json.model;
                     const delta = json.choices?.[0]?.delta;
-                    // Track reasoning/thinking tokens — models like Gemma 4, Qwen3, DeepSeek
-                    // emit reasoning_content during their thinking phase before producing
-                    // visible content. We must send progress notifications during this phase
-                    // to prevent MCP client timeout.
-                    if (delta?.reasoning_content) {
-                        chunkCount++;
-                        if (options.progressToken !== undefined) {
-                            server.notification({
-                                method: 'notifications/progress',
-                                params: {
-                                    progressToken: options.progressToken,
-                                    progress: chunkCount,
-                                    message: `Thinking... (${chunkCount} chunks)`,
-                                },
-                            }).catch(() => { });
-                        }
+                    // Reasoning channel. LM Studio (with "Separate reasoning_content"
+                    // dev setting), DeepSeek R1, Ollama OpenAI-compat, Nemotron etc.
+                    // stream reasoning via delta.reasoning_content — we MUST capture it
+                    // so the safety net below can return something when the model
+                    // burns its entire budget before emitting a single content token.
+                    if (typeof delta?.reasoning_content === 'string' && delta.reasoning_content.length > 0) {
+                        reasoning += delta.reasoning_content;
+                        sendProgress(`Thinking... (${reasoning.length} chars of reasoning)`);
                     }
-                    if (delta?.content) {
+                    if (typeof delta?.content === 'string' && delta.content.length > 0) {
                         if (ttftMs === undefined)
                             ttftMs = Date.now() - startTime;
                         content += delta.content;
-                        chunkCount++;
-                        // Send progress notification to reset MCP client timeout.
-                        // Each notification resets the 60s clock, giving slow models
-                        // unlimited time as long as they're actively generating.
-                        if (options.progressToken !== undefined) {
-                            server.notification({
-                                method: 'notifications/progress',
-                                params: {
-                                    progressToken: options.progressToken,
-                                    progress: chunkCount,
-                                    message: `Streaming... ${content.length} chars`,
-                                },
-                            }).catch(() => { });
-                        }
+                        sendProgress(`Streaming... ${content.length} chars`);
                     }
                     const reason = json.choices?.[0]?.finish_reason;
                     if (reason)
@@ -506,7 +645,10 @@ async function chatCompletionStreamingInner(messages, options = {}) {
                     if (json.model)
                         model = json.model;
                     const delta = json.choices?.[0]?.delta;
-                    if (delta?.content) {
+                    if (typeof delta?.reasoning_content === 'string' && delta.reasoning_content.length > 0) {
+                        reasoning += delta.reasoning_content;
+                    }
+                    if (typeof delta?.content === 'string' && delta.content.length > 0) {
                         if (ttftMs === undefined)
                             ttftMs = Date.now() - startTime;
                         content += delta.content;
@@ -525,50 +667,117 @@ async function chatCompletionStreamingInner(messages, options = {}) {
         }
     }
     finally {
-        // Release the reader — don't await cancel() as it can hang
-        reader.releaseLock();
+        clearInterval(keepAliveTimer);
+        // Best-effort cancel with a short timeout — cancel() can hang if the upstream
+        // connection is wedged, so we race it against a 500ms timer. This frees the
+        // underlying socket sooner on abrupt client disconnects without blocking the
+        // tool response path.
+        try {
+            await Promise.race([
+                reader.cancel().catch(() => { }),
+                new Promise((resolve) => setTimeout(resolve, 500)),
+            ]);
+        }
+        catch { /* never propagate cleanup errors */ }
+        try {
+            reader.releaseLock();
+        }
+        catch { /* already released */ }
     }
     const generationMs = Date.now() - startTime;
     // Strip <think>...</think> reasoning blocks from models that always emit them
-    // (e.g. GLM Flash, Nemotron). Claude doesn't need the model's internal reasoning.
-    // Handle both closed blocks and unclosed ones (model ran out of tokens mid-think,
-    // or grammar-constrained output forced content before the closing tag).
+    // inline on the content channel (e.g. GLM Flash). Claude doesn't need the
+    // model's internal reasoning. Handle both closed and unclosed blocks.
     let cleanContent = content.replace(/<think>[\s\S]*?<\/think>\s*/g, ''); // closed blocks
     cleanContent = cleanContent.replace(/^<think>\s*/, ''); // orphaned opening tag
     cleanContent = cleanContent.trim();
-    // Safety net on top of the thinking-model max_tokens inflation: some MLX/GGUF
-    // quants still exhaust their budget inside an unclosed <think> block despite
-    // `enable_thinking:false` and the 4× inflation. If stripping leaves nothing but
-    // raw output exists, return the raw reasoning so the caller sees *something*
-    // rather than an empty body + lone footer (issue #6).
+    // Safety nets for empty visible output. Try in order:
+    //   1. thinkStripFallback: stripping <think> left nothing, but raw content had text
+    //   2. reasoningFallback: no visible content AT ALL, but reasoning_content was streamed
+    //      (this is the Nemotron/DeepSeek-R1/LM-Studio-dev-toggle case — previously
+    //      produced silent empty bodies because reasoning was discarded)
     let thinkStripFallback = false;
-    if (!cleanContent && content.trim()) {
-        thinkStripFallback = true;
-        cleanContent = content.trim();
+    let reasoningFallback = false;
+    if (!cleanContent) {
+        if (content.trim()) {
+            thinkStripFallback = true;
+            cleanContent = content.trim();
+        }
+        else if (reasoning.trim()) {
+            reasoningFallback = true;
+            cleanContent =
+                '[No visible output — the model spent its entire output budget on reasoning_content before emitting any content. ' +
+                    'Raw reasoning below so you can see what it was doing:]\n\n' +
+                    reasoning.trim();
+        }
     }
-    return { content: cleanContent, rawContent: content, model, usage, finishReason, truncated, ttftMs, generationMs, thinkStripFallback };
+    return {
+        content: cleanContent,
+        rawContent: content,
+        reasoningContent: reasoning || undefined,
+        model,
+        usage,
+        finishReason,
+        truncated,
+        ttftMs,
+        generationMs,
+        thinkStripFallback,
+        reasoningFallback,
+        prefillStall,
+    };
+}
+let detectedBackend = null;
+function getBackend() {
+    return detectedBackend ?? 'openai-compat';
 }
 /**
- * Fetch models from LM Studio's native v0 API first (richer metadata),
- * falling back to the OpenAI-compatible v1 endpoint for non-LM-Studio hosts.
+ * Fetch models with backend-aware probing.
+ *   1. LM Studio /api/v0/models — richest metadata, sets backend='lmstudio'
+ *   2. Ollama /api/tags           — native list, sets backend='ollama', maps to ModelInfo
+ *   3. OpenAI-compatible /v1/models — generic fallback (DeepSeek, vLLM, llama.cpp, OpenRouter)
  */
 async function listModelsRaw() {
-    // Try v0 API first — returns type, arch, publisher, quantization, state
+    // Try LM Studio's v0 API first — returns type, arch, publisher, quantization, state
     try {
         const v0 = await fetchWithTimeout(`${LM_BASE_URL}/api/v0/models`, { headers: apiHeaders() });
         if (v0.ok) {
             const data = (await v0.json());
+            detectedBackend = 'lmstudio';
             return data.data;
         }
     }
     catch {
-        // v0 not available — fall through to v1
+        // v0 not available — fall through
     }
-    // Fallback: OpenAI-compatible v1 endpoint (works with Ollama, vLLM, llama.cpp)
+    // Try Ollama's /api/tags next. Shape differs from OpenAI: returns
+    // { models: [{ name, model, size, details: { family, parameter_size, ... } }] }
+    try {
+        const tags = await fetchWithTimeout(`${LM_BASE_URL}/api/tags`, { headers: apiHeaders() });
+        if (tags.ok) {
+            const data = (await tags.json());
+            if (Array.isArray(data.models)) {
+                detectedBackend = 'ollama';
+                return data.models.map((m) => ({
+                    id: m.name,
+                    object: 'model',
+                    type: 'llm',
+                    arch: m.details?.family,
+                    quantization: m.details?.quantization_level,
+                    state: 'loaded', // Ollama loads on-demand; treat all listed as available
+                    publisher: m.name.includes('/') ? m.name.split('/')[0] : undefined,
+                }));
+            }
+        }
+    }
+    catch {
+        // Not Ollama — fall through
+    }
+    // Fallback: OpenAI-compatible v1 endpoint (DeepSeek, vLLM, llama.cpp, OpenRouter)
     const res = await fetchWithTimeout(`${LM_BASE_URL}/v1/models`, { headers: apiHeaders() });
     if (!res.ok)
         throw new Error(`Failed to list models: ${res.status}`);
     const data = (await res.json());
+    detectedBackend = 'openai-compat';
     return data.data;
 }
 function getContextLength(model) {
@@ -579,6 +788,81 @@ function getContextLength(model) {
 function getMaxContextLength(model) {
     return model.max_context_length;
 }
+/**
+ * Map model family / backend → reasoning_effort value that minimises reasoning.
+ *
+ * The `reasoning_effort` field exists across OpenAI, Ollama, LM Studio and
+ * DeepSeek, but the accepted values differ per vendor. Verified empirically
+ * from the LM Studio error response: "Supported values: none, minimal, low,
+ * medium, high, xhigh" (that's the set the LM Studio adapter accepts).
+ *
+ *   OpenAI (gpt-5, o-series)        : 'low' | 'medium' | 'high' (spec)
+ *   Ollama                          : 'low' | 'medium' | 'high' | 'none'
+ *   LM Studio (all models)          : 'none' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh'
+ *
+ * We want the HARDEST off-switch we can portably send:
+ *   - LM Studio / Ollama: 'none'  (no reasoning budget at all)
+ *   - Generic OpenAI-compat: 'low' (OpenAI's minimum, safe to send)
+ *
+ * An unsupported value is a hard 400 error on LM Studio (not a silent
+ * fallback), so this function is conservative — it returns null for
+ * unknown backends and we omit the field rather than risk a 400.
+ */
+function getReasoningEffortValue(_modelId) {
+    const backend = getBackend();
+    // LM Studio accepts 'none' as an explicit reasoning-off switch for
+    // every thinking model (Nemotron, DeepSeek R1, Gemma 4, gpt-oss, ...).
+    if (backend === 'lmstudio')
+        return 'none';
+    // Ollama likewise documents 'none' as valid.
+    if (backend === 'ollama')
+        return 'none';
+    // Generic OpenAI-compatible — 'low' is the minimum OpenAI accepts per spec.
+    // DeepSeek's own API treats 'low' as minimum too.
+    return 'low';
+}
+/** Rough chars→tokens ratio used for pre-flight estimates. Matches the ratio
+ * we already use to estimate completion_tokens when usage is missing. */
+const CHARS_PER_TOKEN = 4;
+/** Conservative default prefill rate when no per-model measurement exists.
+ * Slower than real hardware so we err toward letting the call run — a false
+ * refusal is much worse than a false-ok that eventually times out. */
+const DEFAULT_PREFILL_TOK_PER_SEC = 300;
+/** Hard ceiling for when we refuse to send the call. Leaves ~15s of
+ * generation headroom inside the ~60s MCP-client request-timeout budget. */
+const PREFILL_REFUSE_THRESHOLD_SEC = 45;
+/** Soft warning threshold — we proceed but log a stderr warning. */
+const PREFILL_WARN_THRESHOLD_SEC = 25;
+/**
+ * Estimate how long prompt prefill will take, using measured per-model data
+ * from the SQLite cache when available. `totalTtftMs` is very close to pure
+ * prefill time for a streaming call (first-content-delta arrives right after
+ * prefill finishes), so `totalPromptTokens / totalTtftMs` gives a usable
+ * prefill-tok/s rate for that specific (model, hardware) pair.
+ */
+function estimatePrefill(inputChars, modelId) {
+    const inputTokens = Math.ceil(inputChars / CHARS_PER_TOKEN);
+    const stats = lifetime.modelStats.get(modelId);
+    let prefillTokPerSec = DEFAULT_PREFILL_TOK_PER_SEC;
+    let basis = 'default';
+    if (stats && stats.ttftCalls >= 2 && stats.totalTtftMs > 0 && stats.totalPromptTokens > 0) {
+        // Only trust measured data after >=2 TTFT samples — single samples on a
+        // cold model run are noisy. Average prompt tokens per call approximated
+        // as totalPromptTokens / calls, then divided by average TTFT seconds.
+        const avgPromptTokens = stats.totalPromptTokens / stats.calls;
+        const avgTtftSec = (stats.totalTtftMs / stats.ttftCalls) / 1000;
+        if (avgTtftSec > 0) {
+            prefillTokPerSec = avgPromptTokens / avgTtftSec;
+            basis = 'measured';
+        }
+    }
+    return {
+        inputTokens,
+        estimatedSeconds: inputTokens / prefillTokPerSec,
+        prefillTokPerSec,
+        basis,
+    };
+}
 async function routeToModel(taskType) {
     let models;
     try {
@@ -645,9 +929,11 @@ function assessQuality(resp, rawContent) {
         : null;
     return {
         truncated: resp.truncated,
+        prefillStall: resp.prefillStall ?? false,
         finishReason: resp.finishReason || 'unknown',
         thinkBlocksStripped: hadThinkBlocks,
         thinkStripFallback: resp.thinkStripFallback ?? false,
+        reasoningFallback: resp.reasoningFallback ?? false,
         estimatedTokens: estimated,
         contentLength: resp.content.length,
         generationMs: resp.generationMs,
@@ -656,9 +942,13 @@ function assessQuality(resp, rawContent) {
 }
 function formatQualityLine(quality) {
     const flags = [];
-    if (quality.truncated)
+    if (quality.prefillStall)
+        flags.push('PREFILL-STALL (no tokens received — input may be too large for this model/hardware)');
+    else if (quality.truncated)
         flags.push('TRUNCATED');
-    if (quality.thinkStripFallback)
+    if (quality.reasoningFallback)
+        flags.push('reasoning-only (model exhausted output budget before emitting visible content — showing raw reasoning)');
+    else if (quality.thinkStripFallback)
         flags.push('think-strip-empty (showing raw reasoning — model ignored enable_thinking:false)');
     else if (quality.thinkBlocksStripped)
         flags.push('think-blocks-stripped');
@@ -672,6 +962,12 @@ function formatQualityLine(quality) {
 }
 /**
  * Format a footer line for streaming results showing model, usage, and truncation status.
+ *
+ * Layout:
+ *   ---
+ *   Model: ... | prompt→completion tokens | perf | extra | quality
+ *   📊 [first-call benchmark line, only on the first measured call per model]
+ *   💰 Claude quota saved this session: ...
  */
 function formatFooter(resp, extra) {
     // Record usage for session tracking before formatting
@@ -680,7 +976,17 @@ function formatFooter(resp, extra) {
     if (resp.model)
         parts.push(`Model: ${resp.model}`);
     if (resp.usage) {
-        parts.push(`${resp.usage.prompt_tokens}→${resp.usage.completion_tokens} tokens`);
+        // OpenAI-spec reasoning-tokens split — when present, show it so the user
+        // sees how much of the completion budget went to hidden reasoning vs
+        // visible output. Diagnoses "empty body + hit-max-tokens" immediately.
+        const reasoningTokens = resp.usage.completion_tokens_details?.reasoning_tokens;
+        if (typeof reasoningTokens === 'number' && reasoningTokens > 0) {
+            const visible = resp.usage.completion_tokens - reasoningTokens;
+            parts.push(`${resp.usage.prompt_tokens}→${resp.usage.completion_tokens} tokens (${reasoningTokens} reasoning / ${visible} visible)`);
+        }
+        else {
+            parts.push(`${resp.usage.prompt_tokens}→${resp.usage.completion_tokens} tokens`);
+        }
     }
     else if (resp.content.length > 0) {
         // Estimate when usage is missing (truncated responses where final SSE chunk was lost)
@@ -691,8 +997,9 @@ function formatFooter(resp, extra) {
     const perfParts = [];
     if (resp.ttftMs !== undefined)
         perfParts.push(`TTFT: ${resp.ttftMs}ms`);
+    let tokPerSec = 0;
     if (resp.usage && resp.generationMs > 50) {
-        const tokPerSec = resp.usage.completion_tokens / (resp.generationMs / 1000);
+        tokPerSec = resp.usage.completion_tokens / (resp.generationMs / 1000);
         perfParts.push(`${tokPerSec.toFixed(1)} tok/s`);
     }
     if (resp.generationMs)
@@ -708,40 +1015,43 @@ function formatFooter(resp, extra) {
         parts.push(qualityLine);
     if (resp.truncated)
         parts.push('⚠ TRUNCATED (soft timeout — partial result)');
+    const benchmarkLine = isFirstBenchmarkedCall(resp.model, tokPerSec)
+        ? `📊 First measured call on ${resp.model}: ${tokPerSec.toFixed(1)} tok/s${resp.ttftMs !== undefined ? `, ${resp.ttftMs}ms to first token` : ''} — use this to gauge whether to delegate longer tasks.`
+        : '';
     const sessionLine = sessionSummary();
+    if (parts.length === 0 && !benchmarkLine && !sessionLine)
+        return '';
+    const lines = [`\n\n---${parts.length > 0 ? `\n${parts.join(' | ')}` : ''}`];
+    // First-call speed benchmark — surfaced once per model per session, based on
+    // the real task just completed (not a synthetic warmup). Gives Claude honest
+    // speed data to calibrate future delegation decisions.
+    if (benchmarkLine)
+        lines.push(benchmarkLine);
+    // Session savings — on its own line so it reads as value, not as accounting.
     if (sessionLine)
-        parts.push(sessionLine);
-    return parts.length > 0 ? `\n\n---\n${parts.join(' | ')}` : '';
+        lines.push(sessionLine);
+    return lines.join('\n');
 }
 // ── MCP Tool definitions ─────────────────────────────────────────────
 const TOOLS = [
     {
         name: 'chat',
-        description: 'Send a task to a local LLM running on a separate machine. This is a FREE, parallel worker — ' +
-            'use it to offload bounded work while you continue doing other things. The local LLM runs independently ' +
-            'and does not consume your tokens or rate limits.\n\n' +
-            'PLANNING: When you start a large task (refactoring, migrations, test suites, documentation), ' +
-            'break it into steps and identify which ones are bounded grunt work you can delegate here. ' +
-            'The more you offload, the more tokens you save. The session footer tracks cumulative savings.\n\n' +
-            'WHEN TO USE (delegate generously — it costs nothing):\n' +
-            '• Explain or summarise code/docs you just read\n' +
+        description: 'Send a task to a local LLM — a sidekick running on the user\'s hardware or a configured OpenAI-compatible endpoint. ' +
+            'It does not consume the user\'s Claude quota. Trades latency for tokens: local inference is typically 3-30× slower than frontier models, so delegation wins when the task is bounded and self-contained.\n\n' +
+            'Good fit:\n' +
+            '• Explain or summarise code/docs you already have in context\n' +
             '• Generate boilerplate, test stubs, type definitions, mock data\n' +
             '• Answer factual questions about languages, frameworks, APIs\n' +
             '• Draft commit messages, PR descriptions, comments\n' +
             '• Translate or reformat content (JSON↔YAML, snake_case↔camelCase)\n' +
-            '• Brainstorm approaches before you commit to one\n' +
-            '• Any self-contained subtask that does not need tool access\n\n' +
-            'PROMPT QUALITY (the local model is highly capable — results depend on your prompt):\n' +
-            '(1) Always send COMPLETE code/context — never truncate, the local LLM cannot access files.\n' +
+            '• Brainstorm approaches before committing to one\n\n' +
+            'Less good when: the task needs tool access, depends on multi-file context you have not captured, or is quick enough for you to answer directly before the round-trip completes.\n\n' +
+            'Prompt tips (local models take instructions literally):\n' +
+            '(1) Send COMPLETE context — the local LLM cannot read files.\n' +
             '(2) Be explicit about output format ("respond as a JSON array", "return only the function").\n' +
-            '(3) Set a specific persona in the system field — "Senior TypeScript dev" beats "helpful assistant".\n' +
-            '(4) State constraints: "no preamble", "reference line numbers", "max 5 bullet points".\n' +
-            '(5) For code generation, include the surrounding context (imports, types, function signatures).\n\n' +
-            'QA: Always review the local LLM\'s output before using it. Verify correctness, check edge cases, ' +
-            'and fix any issues. You are the architect — the local model is a fast drafter, not the final authority.\n\n' +
-            'ROUTING: If multiple models are loaded, houtini-lm automatically picks the best one for the task. ' +
-            'If a better model is downloaded but not loaded, you\'ll see a suggestion in the response footer. ' +
-            'Call discover to see what\'s available.',
+            '(3) Specific system persona beats generic — "Senior TypeScript dev" not "helpful assistant".\n' +
+            '(4) State constraints — "no preamble", "reference line numbers", "max 5 bullets".\n\n' +
+            'Routing picks the best loaded model automatically. Call `discover` to see what is loaded and, after the first real call, its measured speed. The footer shows cumulative tokens kept in the user\'s quota.',
         inputSchema: {
             type: 'object',
             properties: {
@@ -759,7 +1069,7 @@ const TOOLS = [
                 },
                 max_tokens: {
                     type: 'number',
-                    description: 'Max response tokens. Default 2048. Use higher for code generation, lower for quick answers.',
+                    description: 'Max response tokens. Defaults to 25% of the loaded model\'s context window (fallback 16,384). Pass a number to cap it tighter for quick answers.',
                 },
                 json_schema: {
                     type: 'object',
@@ -772,20 +1082,18 @@ const TOOLS = [
     {
         name: 'custom_prompt',
         description: 'Structured analysis via the local LLM with explicit system/context/instruction separation. ' +
-            'This 3-part format prevents context bleed and gets the best results from local models.\n\n' +
-            'USE THIS for complex tasks where prompt structure matters — it consistently outperforms ' +
-            'stuffing everything into a single message. The separation helps the local model focus.\n\n' +
-            'WHEN TO USE:\n' +
+            'The 3-part format prevents context bleed in smaller models — the local LLM acknowledges the context in a fake assistant turn before receiving the instruction.\n\n' +
+            'Good fit when prompt structure matters:\n' +
             '• Code review — paste full source, ask for bugs/improvements\n' +
             '• Comparison — paste two implementations, ask which is better and why\n' +
             '• Refactoring suggestions — paste code, ask for a cleaner version\n' +
             '• Content analysis — paste text, ask for structure/tone/issues\n' +
             '• Any task where separating context from instruction improves clarity\n\n' +
-            'PROMPT STRUCTURE (each field has a job — keep them focused):\n' +
-            '• System: persona + constraints, under 30 words. "Expert Python developer focused on performance and correctness."\n' +
-            '• Context: COMPLETE data. Full source code, full logs, full text. NEVER truncate or summarise.\n' +
-            '• Instruction: exactly what to produce, under 50 words. Specify format: "Return a JSON array of {line, issue, fix}."\n\n' +
-            'QA: Review the output. The local model is a capable drafter — verify its analysis before acting on it.',
+            'Field guidance (each has a job — keep them focused):\n' +
+            '• system: persona + constraints, under 30 words. "Expert Python developer focused on performance and correctness."\n' +
+            '• context: COMPLETE data — full source, full logs, full text. Never truncate.\n' +
+            '• instruction: exactly what to produce, under 50 words. Specify format: "Return a JSON array of {line, issue, fix}."\n\n' +
+            'Review the output before acting on it — local model capability varies.',
         inputSchema: {
             type: 'object',
             properties: {
@@ -807,7 +1115,7 @@ const TOOLS = [
                 },
                 max_tokens: {
                     type: 'number',
-                    description: 'Max response tokens. Default 2048.',
+                    description: 'Max response tokens. Defaults to 25% of the loaded model\'s context window (fallback 16,384).',
                 },
                 json_schema: {
                     type: 'object',
@@ -819,21 +1127,19 @@ const TOOLS = [
     },
     {
         name: 'code_task',
-        description: 'Send a code analysis task to the local LLM. Wraps the request with an optimised code-review system prompt.\n\n' +
-            'This is the fastest way to offload code-specific work. Temperature is locked to 0.2 for ' +
-            'focused, deterministic output. The system prompt is pre-configured for code review.\n\n' +
-            'WHEN TO USE:\n' +
+        description: 'Send a code-specific task to the local LLM, wrapped with an optimised code-review system prompt. Temperature is locked low (0.2 or the routed model\'s hint) for deterministic output.\n\n' +
+            'Good fit:\n' +
             '• Explain what a function/class does\n' +
             '• Find bugs or suggest improvements\n' +
             '• Generate unit tests or type definitions for existing code\n' +
             '• Add error handling, logging, or validation\n' +
             '• Convert between languages or patterns\n\n' +
-            'GETTING BEST RESULTS:\n' +
-            '• Provide COMPLETE source code — the local LLM cannot read files.\n' +
+            'For best results:\n' +
+            '• Provide COMPLETE source — the local LLM cannot read files.\n' +
             '• Include imports and type definitions so the model has full context.\n' +
-            '• Be specific in the task: "Write 3 Jest tests for the error paths in fetchUser" beats "Write tests".\n' +
+            '• Be specific: "Write 3 Jest tests for the error paths in fetchUser" beats "Write tests".\n' +
             '• Set the language field — it shapes the system prompt and improves accuracy.\n\n' +
-            'QA: Always verify generated code compiles, handles edge cases, and follows project conventions.',
+            'Verify generated code compiles, handles edge cases, and follows project conventions before committing.',
         inputSchema: {
             type: 'object',
             properties: {
@@ -851,7 +1157,7 @@ const TOOLS = [
                 },
                 max_tokens: {
                     type: 'number',
-                    description: 'Max response tokens. Default 2048.',
+                    description: 'Max response tokens. Defaults to 25% of the loaded model\'s context window (fallback 16,384).',
                 },
             },
             required: ['code', 'task'],
@@ -859,18 +1165,19 @@ const TOOLS = [
     },
     {
         name: 'code_task_files',
-        description: 'Like code_task, but the local LLM reads the files directly from disk — the contents never pass through the MCP client\'s context window.\n\n' +
-            'USE THIS instead of code_task when you want the LLM to review multiple files or a single large file, without copying source into the chat.\n\n' +
-            'HOW IT WORKS:\n' +
-            '• Provide absolute paths to the files you want analysed.\n' +
-            '• The server reads each file (Promise.allSettled — one unreadable file does not sink the call).\n' +
-            '• Files are concatenated with `=== <filename> ===` headers, then sent to the same code-review pipeline as code_task.\n' +
-            '• Read failures are surfaced inline (with the reason) so the LLM can still reason about what it did receive.\n\n' +
-            'WHEN TO USE:\n' +
-            '• Reviewing multiple related files (module + its tests, client + server pair)\n' +
+        description: 'Like code_task, but the local LLM reads files directly from disk — source never passes through the MCP client\'s context window. Use when reviewing multiple files or a single large file.\n\n' +
+            'How it works:\n' +
+            '• Provide absolute paths. Relative paths are rejected.\n' +
+            '• Files are read in parallel (Promise.allSettled) — one unreadable file does not sink the call.\n' +
+            '• Files are concatenated with `=== filename ===` headers and sent to the same code-review pipeline as code_task.\n' +
+            '• Read failures are surfaced inline with the reason so the LLM can still reason about the rest.\n' +
+            '• Pre-flight prefill estimate: if measured per-model data shows the input would exceed the MCP client\'s ~60s request timeout during prompt processing, the call is refused early with a diagnostic instead of hanging. Split or trim when this fires.\n\n' +
+            'Good fit:\n' +
+            '• Reviewing related files together (module + its tests, client + server pair)\n' +
             '• Auditing a single large file too big to paste comfortably\n' +
-            '• Any code_task where saving MCP client tokens matters\n\n' +
-            'QA: Same rules as code_task — verify the output before acting on it.',
+            '• Any code_task where keeping source out of the Claude context window matters\n\n' +
+            'Size guidance: on slow hardware (< 25 tok/s generation), keep total input under ~8,000 tokens (~32,000 chars) to stay safely under the client timeout. Faster hardware handles much more — the pre-flight estimator adapts once you\'ve done a few calls and real per-model timings are in the SQLite cache.\n\n' +
+            'Same review discipline as code_task — verify the output before acting on it.',
         inputSchema: {
             type: 'object',
             properties: {
@@ -931,9 +1238,33 @@ const TOOLS = [
             required: ['input'],
         },
     },
+    {
+        name: 'stats',
+        description: 'Show user stats: tokens offloaded, calls made, per-model performance — for the current session AND ' +
+            'lifetime (persisted in SQLite at ~/.houtini-lm/model-cache.db). Unlike `discover` which includes the ' +
+            'model catalog, `stats` returns just the numbers in a compact markdown table — cheap to call repeatedly ' +
+            'to see the 💰 Claude-quota savings counter climb. Useful for quantifying how much work the local model ' +
+            'is genuinely doing, and for noticing when a model\'s reasoning-token ratio is drifting.',
+        inputSchema: {
+            type: 'object',
+            properties: {
+                model: {
+                    type: 'string',
+                    description: 'Optional: filter output to a single model ID. Omit to see all models this workstation has used.',
+                },
+            },
+        },
+    },
 ];
 // ── MCP Server ───────────────────────────────────────────────────────
-const server = new Server({ name: 'houtini-lm', version: '2.9.0' }, { capabilities: { tools: {}, resources: {} } });
+// Session-level sidekick framing. MCP clients surface this to the model
+// at initialisation, so it sets the baseline expectation for when to delegate
+// rather than relying on per-tool descriptions being re-read on every call.
+const SIDEKICK_INSTRUCTIONS = `Houtini-lm is a local LLM sidekick. It runs on the user's hardware (or a configured OpenAI-compatible endpoint) and handles bounded work without consuming the user's Claude quota.\n\n` +
+    `When to reach for it: bounded, self-contained tasks you can describe in one message — explanations, boilerplate, test stubs, code review of pasted or file-loaded source, translations, commit messages, format conversion, brainstorming. Trades wall-clock time for tokens (typically 3-30× slower than frontier models).\n\n` +
+    `When not to: tasks that need tool access, cross-file reasoning you haven't captured, or work fast enough to answer directly before the delegation round-trip completes.\n\n` +
+    `Call \`discover\` in delegation-heavy sessions to see what model is loaded, its capability profile, and — after the first real call — its measured speed. The response footer reports cumulative tokens kept in the user's quota.`;
+const server = new Server({ name: 'houtini-lm', version: '2.11.0' }, { capabilities: { tools: {}, resources: {} }, instructions: SIDEKICK_INSTRUCTIONS });
 // ── MCP Resources ─────────────────────────────────────────────────────
 // Exposes session performance metrics as a readable resource so Claude can
 // proactively check offload efficiency and make smarter delegation decisions.
@@ -954,7 +1285,7 @@ server.setRequestHandler(ReadResourceRequestSchema, async (request) => {
         for (const [modelId, stats] of session.modelStats) {
             modelStats[modelId] = {
                 calls: stats.calls,
-                avgTtftMs: stats.calls > 0 ? Math.round(stats.totalTtftMs / stats.calls) : 0,
+                avgTtftMs: stats.ttftCalls > 0 ? Math.round(stats.totalTtftMs / stats.ttftCalls) : 0,
                 avgTokPerSec: stats.perfCalls > 0 ? parseFloat((stats.totalTokPerSec / stats.perfCalls).toFixed(1)) : null,
             };
         }
@@ -1114,6 +1445,32 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                     ? ` ${route.hints.outputConstraint}`
                     : '';
                 const combined = sections.join('\n\n');
+                // Pre-flight prefill estimate. Huge inputs can legitimately exceed
+                // the MCP client's ~60s request timeout during prompt processing, and
+                // progress notifications don't reset that timeout on Claude Desktop.
+                // If measured per-model data in the SQLite cache shows this input
+                // would obviously overrun, refuse with a concrete diagnostic so the
+                // caller knows to split or trim instead of waiting for a silent hang.
+                const estimate = estimatePrefill(combined.length, route.modelId);
+                if (estimate.basis === 'measured' && estimate.estimatedSeconds > PREFILL_REFUSE_THRESHOLD_SEC) {
+                    const prefillRate = Math.round(estimate.prefillTokPerSec);
+                    const estSec = Math.round(estimate.estimatedSeconds);
+                    return {
+                        content: [{
+                                type: 'text',
+                                text: `Error: estimated prefill time exceeds the ~60s MCP client timeout.\n\n` +
+                                    `• Input size: ~${estimate.inputTokens.toLocaleString()} tokens across ${successCount} file(s)\n` +
+                                    `• Measured prefill rate on ${route.modelId}: ~${prefillRate} tok/s (from ${lifetime.modelStats.get(route.modelId)?.ttftCalls ?? 0} prior calls)\n` +
+                                    `• Estimated prefill: ~${estSec}s (threshold: ${PREFILL_REFUSE_THRESHOLD_SEC}s)\n\n` +
+                                    `Options: split the files into smaller groups, trim the largest file, or use \`code_task\` with a focused excerpt. ` +
+                                    `If you know this workstation can handle it, pass fewer files or run the task again when the measured rate improves.`,
+                            }],
+                        isError: true,
+                    };
+                }
+                if (estimate.estimatedSeconds > PREFILL_WARN_THRESHOLD_SEC) {
+                    process.stderr.write(`[houtini-lm] Large input warning: ~${estimate.inputTokens} tokens, est prefill ~${Math.round(estimate.estimatedSeconds)}s (${estimate.basis}). Proceeding.\n`);
+                }
                 const codeMessages = [
                     {
                         role: 'system',
@@ -1171,14 +1528,49 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                 const primary = loaded[0] || models[0];
                 const ctx = getContextLength(primary);
                 const primaryProfile = await getModelProfileAsync(primary);
-                const sessionStats = session.calls > 0
-                    ? `\nSession stats: ${(session.promptTokens + session.completionTokens).toLocaleString()} tokens offloaded across ${session.calls} call${session.calls === 1 ? '' : 's'}`
-                    : '\nSession stats: no calls yet — delegate tasks to start saving tokens';
+                // Use sessionSummary() so discover matches the footer format and
+                // automatically picks up the lifetime line when the SQLite cache has
+                // cross-session data.
+                const summary = sessionSummary();
+                const sessionStats = session.calls > 0 || lifetime.totalCalls > 0
+                    ? `\n${summary}`
+                    : `\n💰 Claude quota saved this session: 0 tokens — no calls yet. Measured speed for each model will appear here after the first real call.`;
+                // Measured speed line for the active model. Discover intentionally does
+                // not run a synthetic warmup — speed is captured from real tasks, so the
+                // numbers reflect actual workload rather than a contrived benchmark.
+                // Shows session stats when this session has measured calls; otherwise
+                // falls back to workstation lifetime stats so Claude sees historical
+                // perf from call 1 instead of "not yet benchmarked".
+                const primaryStats = session.modelStats.get(primary.id);
+                const primaryLifetime = lifetime.modelStats.get(primary.id);
+                let speedLine = '';
+                if (primaryStats && primaryStats.perfCalls > 0) {
+                    const avgTtft = primaryStats.ttftCalls > 0 ? Math.round(primaryStats.totalTtftMs / primaryStats.ttftCalls) : 0;
+                    const avgTokSec = (primaryStats.totalTokPerSec / primaryStats.perfCalls).toFixed(1);
+                    speedLine = `Measured speed (session): ${avgTokSec} tok/s · TTFT ${avgTtft}ms (${primaryStats.perfCalls} call${primaryStats.perfCalls === 1 ? '' : 's'})\n`;
+                    if (primaryLifetime && primaryLifetime.perfCalls > primaryStats.perfCalls) {
+                        const lAvgTtft = primaryLifetime.ttftCalls > 0 ? Math.round(primaryLifetime.totalTtftMs / primaryLifetime.ttftCalls) : 0;
+                        const lAvgTokSec = (primaryLifetime.totalTokPerSec / primaryLifetime.perfCalls).toFixed(1);
+                        speedLine += `Measured speed (lifetime on this workstation): ${lAvgTokSec} tok/s · TTFT ${lAvgTtft}ms (${primaryLifetime.perfCalls} calls)\n`;
+                    }
+                }
+                else if (primaryLifetime && primaryLifetime.perfCalls > 0) {
+                    const lAvgTtft = primaryLifetime.ttftCalls > 0 ? Math.round(primaryLifetime.totalTtftMs / primaryLifetime.ttftCalls) : 0;
+                    const lAvgTokSec = (primaryLifetime.totalTokPerSec / primaryLifetime.perfCalls).toFixed(1);
+                    speedLine = `Measured speed (lifetime on this workstation): ${lAvgTokSec} tok/s · TTFT ${lAvgTtft}ms (${primaryLifetime.perfCalls} calls, last used ${new Date(primaryLifetime.lastUsedAt).toISOString().slice(0, 10)})\n`;
+                }
+                else {
+                    speedLine = `Measured speed: not yet benchmarked — will be captured on the first real call.\n`;
+                }
+                const backendLabel = getBackend() === 'lmstudio' ? 'LM Studio'
+                    : getBackend() === 'ollama' ? 'Ollama'
+                        : 'OpenAI-compatible';
                 let text = `Status: ONLINE\n` +
-                    `Endpoint: ${LM_BASE_URL}\n` +
-                    `Latency: ${ms}ms\n` +
+                    `Endpoint: ${LM_BASE_URL} (${backendLabel})\n` +
+                    `Connection latency: ${ms}ms (does not reflect inference speed)\n` +
                     `Active model: ${primary.id}\n` +
-                    `Context window: ${ctx.toLocaleString()} tokens\n`;
+                    `Context window: ${ctx.toLocaleString()} tokens\n` +
+                    speedLine;
                 if (primaryProfile) {
                     text += `Family: ${primaryProfile.family}\n`;
                     text += `Description: ${primaryProfile.description}\n`;
@@ -1200,11 +1592,25 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                 if (session.modelStats.size > 0) {
                     text += `\n\nPerformance (this session):\n`;
                     for (const [modelId, stats] of session.modelStats) {
-                        const avgTtft = stats.calls > 0 ? Math.round(stats.totalTtftMs / stats.calls) : 0;
+                        const avgTtft = stats.ttftCalls > 0 ? Math.round(stats.totalTtftMs / stats.ttftCalls) : 0;
                         const avgTokSec = stats.perfCalls > 0 ? (stats.totalTokPerSec / stats.perfCalls).toFixed(1) : '?';
                         text += `  ${modelId}: ${stats.calls} calls, avg TTFT ${avgTtft}ms, avg ${avgTokSec} tok/s\n`;
                     }
                 }
+                // Workstation lifetime stats — built from SQLite, persists across restarts.
+                // Only shown when there's lifetime data beyond this session, so a first-run
+                // user doesn't see a duplicate of the session block above.
+                const hasLifetimeBeyondSession = Array.from(lifetime.modelStats.entries())
+                    .some(([id, l]) => l.calls > (session.modelStats.get(id)?.calls ?? 0));
+                if (hasLifetimeBeyondSession) {
+                    text += `\nPerformance (lifetime on this workstation):\n`;
+                    for (const [modelId, stats] of lifetime.modelStats) {
+                        const avgTtft = stats.ttftCalls > 0 ? Math.round(stats.totalTtftMs / stats.ttftCalls) : 0;
+                        const avgTokSec = stats.perfCalls > 0 ? (stats.totalTokPerSec / stats.perfCalls).toFixed(1) : '?';
+                        const lastUsed = new Date(stats.lastUsedAt).toISOString().slice(0, 10);
+                        text += `  ${modelId}: ${stats.calls} calls, avg TTFT ${avgTtft}ms, avg ${avgTokSec} tok/s (last used ${lastUsed})\n`;
+                    }
+                }
                 text += `${sessionStats}\n\n`;
                 text += `The local LLM is available. You can delegate tasks using chat, custom_prompt, code_task, code_task_files, or embed.`;
                 return { content: [{ type: 'text', text }] };
@@ -1262,6 +1668,89 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                     };
                 });
             }
+            case 'stats': {
+                const { model: filterModel } = args;
+                const backendLabel = getBackend() === 'lmstudio' ? 'LM Studio'
+                    : getBackend() === 'ollama' ? 'Ollama'
+                        : 'OpenAI-compatible';
+                const lines = [];
+                lines.push(`## Houtini LM stats`);
+                lines.push('');
+                lines.push(`**Endpoint**: ${LM_BASE_URL} (${backendLabel})`);
+                if (lifetime.firstSeenAt) {
+                    lines.push(`**First call on this workstation**: ${new Date(lifetime.firstSeenAt).toISOString().slice(0, 10)}`);
+                }
+                lines.push('');
+                // Totals block
+                lines.push(`### Totals`);
+                lines.push('');
+                lines.push(`| Scope    | Calls | Prompt tokens | Completion tokens | Total tokens |`);
+                lines.push(`|----------|------:|--------------:|------------------:|-------------:|`);
+                lines.push(`| Session  | ${session.calls} | ${session.promptTokens.toLocaleString()} | ${session.completionTokens.toLocaleString()} | ${(session.promptTokens + session.completionTokens).toLocaleString()} |`);
+                lines.push(`| Lifetime | ${lifetime.totalCalls} | — | — | ${lifetime.totalTokens.toLocaleString()} |`);
+                lines.push('');
+                // Per-model block (union of session + lifetime model ids)
+                const modelIds = new Set([
+                    ...session.modelStats.keys(),
+                    ...lifetime.modelStats.keys(),
+                ]);
+                const filtered = filterModel ? [...modelIds].filter((m) => m === filterModel) : [...modelIds];
+                if (filtered.length > 0) {
+                    lines.push(`### Per-model performance`);
+                    lines.push('');
+                    lines.push(`| Model | Scope | Calls | Avg TTFT (ms) | Avg tok/s | Prompt tokens | Last used |`);
+                    lines.push(`|-------|-------|------:|--------------:|----------:|--------------:|-----------|`);
+                    for (const modelId of filtered.sort()) {
+                        const s = session.modelStats.get(modelId);
+                        const l = lifetime.modelStats.get(modelId);
+                        if (s) {
+                            const avgTtft = s.ttftCalls > 0 ? Math.round(s.totalTtftMs / s.ttftCalls) : '—';
+                            const avgTokSec = s.perfCalls > 0 ? (s.totalTokPerSec / s.perfCalls).toFixed(1) : '—';
+                            lines.push(`| ${modelId} | session | ${s.calls} | ${avgTtft} | ${avgTokSec} | — | — |`);
+                        }
+                        if (l) {
+                            const avgTtft = l.ttftCalls > 0 ? Math.round(l.totalTtftMs / l.ttftCalls) : '—';
+                            const avgTokSec = l.perfCalls > 0 ? (l.totalTokPerSec / l.perfCalls).toFixed(1) : '—';
+                            const lastUsed = new Date(l.lastUsedAt).toISOString().slice(0, 10);
+                            lines.push(`| ${modelId} | lifetime | ${l.calls} | ${avgTtft} | ${avgTokSec} | ${l.totalPromptTokens.toLocaleString()} | ${lastUsed} |`);
+                        }
+                    }
+                    lines.push('');
+                }
+                else if (filterModel) {
+                    lines.push(`No history for model: \`${filterModel}\`. Try \`list_models\` to see what's been used.`);
+                    lines.push('');
+                }
+                else {
+                    lines.push(`No calls yet — delegate a task via \`chat\`, \`custom_prompt\`, \`code_task\`, or \`code_task_files\` to start building stats.`);
+                    lines.push('');
+                }
+                // Reasoning-token diagnostic (lifetime only — needs persistence to be meaningful)
+                if (!filterModel) {
+                    // Sum reasoning tokens across all models. We store this per-model
+                    // in SQLite but not in the in-memory mirror, so fetch on demand.
+                    try {
+                        const rows = await getAllPerformance();
+                        const totalReasoning = rows.reduce((sum, r) => sum + (r.totalReasoningTokens || 0), 0);
+                        const totalCompletion = rows.reduce((sum, r) => sum + r.totalCompletionTokens, 0);
+                        if (totalCompletion > 0) {
+                            const pct = ((totalReasoning / totalCompletion) * 100).toFixed(1);
+                            lines.push(`### Reasoning-token overhead (lifetime)`);
+                            lines.push('');
+                            lines.push(`${totalReasoning.toLocaleString()} / ${totalCompletion.toLocaleString()} completion tokens spent on hidden reasoning (${pct}% of generation budget). ` +
+                                (parseFloat(pct) > 30
+                                    ? `**High** — consider loading a non-thinking model, or check that \`reasoning_effort\` is being honoured (see stderr logs).`
+                                    : parseFloat(pct) > 10
+                                        ? `Moderate — normal for thinking-model families.`
+                                        : `Low — reasoning is effectively suppressed.`));
+                            lines.push('');
+                        }
+                    }
+                    catch { /* best-effort — don't fail the tool call */ }
+                }
+                lines.push(`*Stats persist across restarts in \`~/.houtini-lm/model-cache.db\`.*`);
+                return { content: [{ type: 'text', text: lines.join('\n') }] };
+            }
             default:
                 throw new Error(`Unknown tool: ${name}`);
         }
@@ -1282,6 +1771,10 @@ async function main() {
     listModelsRaw()
         .then((models) => profileModelsAtStartup(models))
         .catch((err) => process.stderr.write(`[houtini-lm] Startup profiling skipped: ${err}\n`));
+    // Hydrate the in-memory lifetime mirror from SQLite so the very first
+    // tool call this session shows historical savings + per-model perf.
+    // Non-blocking too; the footer degrades to session-only if this fails.
+    hydrateLifetimeFromDb().catch((err) => process.stderr.write(`[houtini-lm] Lifetime hydration skipped: ${err}\n`));
 }
 main().catch((error) => {
     process.stderr.write(`Fatal error: ${error}\n`);