@gotcos/glasses-server 6.39.1 → 6.39.3

package/CHANGELOG.md

@@ -1,3 +1,38 @@
+## 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.
+
+`buildClaudeForkArgs` still carried `--tools ''`, the third read-only layer the
+attached path retired on 2026-08-16. On the current Claude CLI that flag makes
+a fork-resume attempt a context compaction that fails (`too_few_groups`) and
+then report a synthetic 400 "Prompt is too long" with zero input tokens -- so
+the route classified every claude fork `fork_orphan_possible` even though the
+forked transcript sat on disk complete except for the failed turn. Found the
+hour COS Control 0.5.76 put a Fork button in front of the route; isolated by
+single-flag bisection (plan-only completes, allowedTools-only completes,
+`--tools ''` alone reproduces).
+
+The fork keeps its read-only stance -- `--permission-mode plan` plus an empty
+`--allowedTools` -- and drops only the flag the CLI can no longer carry.
+Pinned by a regression test that goes red if `--tools` returns to the argv.
+
 ## 6.39.1
 
 Durable jobs now remember they ran on Ollama.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gotcos/glasses-server",
-  "version": "6.39.1",
+  "version": "6.39.3",
   "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/fork-thread.ts

@@ -211,11 +211,18 @@ export function forkOrphanPossible(result: ForkResult): boolean {
  * `claude -p --resume <id> --fork-session`, read-only, prompt on stdin.
  *
  * `--fork-session` is documented by the installed CLI as "When resuming, create a
- * new session ID" and is only meaningful alongside `--resume`. The read-only pair
- * (`--permission-mode plan` plus the empty tool lists) is carried over from the
- * attached path unchanged: a fork runs a real model turn, and it does so against a
- * workspace the user did not explicitly hand us, so it gets no more authority than
- * a continuation does.
+ * new session ID" and is only meaningful alongside `--resume`. The fork keeps a
+ * read-only stance — a fork runs a real model turn against a workspace the user
+ * did not explicitly hand us — but that stance is now `--permission-mode plan`
+ * plus an empty `--allowedTools` only. The third layer this used to carry,
+ * `--tools ''`, is GONE: on the current CLI an empty --tools on a fork-resume
+ * triggers a spurious context compaction (`too_few_groups`) followed by a
+ * synthetic 400 "Prompt is too long" with zero input tokens — every claude fork
+ * failed as orphan_possible while the transcript copy sat there complete.
+ * Isolated by single-flag bisection on 2026-08-26: plan-only completed,
+ * allowedTools-only completed, `--tools ''` alone reproduced the failure. (The
+ * attached path dropped all three layers on 2026-08-16 by explicit decision;
+ * the fork deliberately keeps the two that still work.)
  *
  * `stream-json` requires `--verbose`; without it the CLI refuses and we would never
  * see the id we are required to verify.
@@ -228,7 +235,6 @@ export function buildClaudeForkArgs(nativeThreadId: string): string[] {
     '--resume', nativeThreadId,
     '--fork-session',
     '--permission-mode', 'plan',
-    '--tools', '',
     '--allowedTools', '',
   ]
 }

package/server/lib/ollama-bridge.ts

@@ -30,6 +30,25 @@ import { OLLAMA_MODEL } from '../../shared/model-preference.js'
 
 const INACTIVITY_MS = 60_000
 const WALL_MAX_MS = 180_000
+
+/**
+ * Thinking is OFF by default. A thinking-class model spends a hidden
+ * reasoning chain before its first visible token -- measured 2026-08-26 on
+ * qwen3.5:35b: 6,265 generated tokens for a 150-word answer (~20K thinking
+ * chars), 98.6s wall against 2.2s with thinking disabled, at the same
+ * visible answer quality. On the lens that is a dead screen for a minute
+ * and a half, and a hard prompt can out-run WALL_MAX_MS entirely. The
+ * daemon silently tolerates `think: false` on models WITHOUT the thinking
+ * capability (verified live against llama3.2:1b -- HTTP 200), so no
+ * capability gate is needed. COS_OLLAMA_THINK opts back in: "1"/"true"
+ * enables it, or an explicit budget level passes through.
+ */
+export function resolveOllamaThink(raw: string | undefined): boolean | string {
+  const value = (raw ?? '').trim().toLowerCase()
+  if (value === '1' || value === 'true') return true
+  if (value === 'low' || value === 'medium' || value === 'high' || value === 'max') return value
+  return false
+}
 const HISTORY_LIMIT = 20
 
 type OllamaChatMessage = { role: 'system' | 'user' | 'assistant'; content: string }
@@ -207,6 +226,7 @@ export async function callOllamaStreaming(
         model: catalog.model,
         messages,
         stream: true,
+        think: resolveOllamaThink(process.env.COS_OLLAMA_THINK),
       }),
       signal: abort.signal,
     })