@gotcos/glasses-server 6.39.2 → 6.40.0

package/CHANGELOG.md

@@ -1,3 +1,39 @@
+## 6.40.0
+
+Local thinking now follows the effort you asked for.
+
+6.39.3 turned thinking off wholesale. This release makes it a ladder: the
+default effort keeps local turns instant, and raising the effort raises the
+thinking budget with it -- xhigh maps to a high thinking level, max and
+ultracode to the maximum. That is what the escalation means: in a controlled
+flag benchmark against an Opus 5 reference (10/10), a local Qwen model WITH
+thinking also scored 10/10, where the same model without it scored 6-7. The
+daemon accepts level strings on Qwen-class models and silently tolerates
+think:false on models without the capability (both probed live).
+
+COS_OLLAMA_THINK becomes a PIN rather than the only switch: "1"/"true"
+always think, "0"/"false" never think, an explicit level forces that level,
+and unset defers to the effort map. Pairs with COS Control 0.5.79, which
+adds a local-model picker writing COS_OLLAMA_MODEL so multi-model daemons
+stop depending on pull order.
+
+## 6.39.3
+
+Local turns answer in seconds again: thinking is off by default.
+
+The Ollama bridge sent no `think` field, so a thinking-class model spent its
+full hidden reasoning chain before the first visible token. Measured on
+qwen3.5:35b the day it became the default local model: 6,265 generated tokens
+for a 150-word answer, 98.6 seconds of dead screen against 2.2 seconds with
+thinking disabled, at the same visible answer quality -- and a hard prompt
+could out-run the 180-second wall clock entirely.
+
+Every request now sends `think: false` unless COS_OLLAMA_THINK opts back in
+("1"/"true", or an explicit budget: low, medium, high, max). The daemon
+silently tolerates `think: false` on models without the thinking capability
+(verified live), so no capability gating is needed. COS Control 0.5.77
+allowlists the new key so a pinned preference survives Update Server.
+
 ## 6.39.2
 
 Every Claude fork failed; the copy it left behind was real.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gotcos/glasses-server",
-  "version": "6.39.2",
+  "version": "6.40.0",
   "description": "COS Glasses \u2014 self-hosted AI heads-up-display server for Even G2 smart glasses, powered by Claude Code, Codex, Cursor Agent CLI, or local Ollama",
   "type": "module",
   "bin": {

package/server/lib/ollama-bridge.ts

@@ -26,10 +26,46 @@ import {
   startOllamaRun,
 } from './ollama-run-ledger.js'
 import { notifyExchange, notifySessionStart } from './telegram-notify.js'
-import { OLLAMA_MODEL } from '../../shared/model-preference.js'
+import { OLLAMA_MODEL, type EffortPreference } from '../../shared/model-preference.js'
 
 const INACTIVITY_MS = 60_000
 const WALL_MAX_MS = 180_000
+
+/**
+ * Thinking follows the REQUESTED EFFORT, not a blanket switch.
+ *
+ * The default effort ('high' -- every ordinary lens turn) keeps thinking
+ * OFF: measured 2026-08-26 on qwen3.5:35b, a hidden chain turned a
+ * 2.2-second answer into 98.6 seconds of dead screen at the same visible
+ * quality, and a hard prompt can out-run WALL_MAX_MS entirely. Raising the
+ * effort raises the thinking budget with it (xhigh -> 'high', max and
+ * ultracode -> 'max'), because that is what the escalation MEANS -- the
+ * flag benchmark showed a thinking local model matching Opus 5 (10/10)
+ * where the same model without thinking scored 6-7. Level strings are
+ * accepted by the daemon on Qwen-class models (probed live), and
+ * `think: false` is silently tolerated by models WITHOUT the thinking
+ * capability (verified against llama3.2:1b -- HTTP 200), so no capability
+ * gate is needed.
+ *
+ * COS_OLLAMA_THINK, when set, PINS the behavior regardless of effort:
+ * "1"/"true" always think, "0"/"false" never think, or an explicit level.
+ * Anything unrecognized reads as unset and defers to the effort map.
+ */
+export function resolveOllamaThink(
+  raw: string | undefined,
+  effort?: EffortPreference,
+): boolean | string {
+  const value = (raw ?? '').trim().toLowerCase()
+  if (value === '1' || value === 'true') return true
+  if (value === '0' || value === 'false') return false
+  if (value === 'low' || value === 'medium' || value === 'high' || value === 'max') return value
+  switch (effort) {
+    case 'xhigh': return 'high'
+    case 'max':
+    case 'ultracode': return 'max'
+    default: return false
+  }
+}
 const HISTORY_LIMIT = 20
 
 type OllamaChatMessage = { role: 'system' | 'user' | 'assistant'; content: string }
@@ -207,6 +243,7 @@ export async function callOllamaStreaming(
         model: catalog.model,
         messages,
         stream: true,
+        think: resolveOllamaThink(process.env.COS_OLLAMA_THINK, options?.effort),
       }),
       signal: abort.signal,
     })