fix(serving): unify UI latency/output and stop /no_think echoes

[konjoinfinity]

Summary

The SquishBar chat UI and the VS Code extension were far slower to respond than the web UI on the same model, and they leaked raw reasoning / tool-call JSON into the chat. Separately, models echoed the /no_think soft-switch directive into their replies. This makes every client behave like the web UI — fast, streamed, and clean — and guarantees /no_think-style directives never reach a client.

Root causes

• VS Code (slow + leaked JSON/think): every message ran through the tool-calling loop, which always sent tools. The server forces stream=False whenever tools are present (server.py), so the client waited for the entire generation before seeing anything. The web UI's plain chat sends no tools → streams → fast.
• SquishBar (slow + raw <think> blocks): the streaming chat path did no <think> handling, dumping the model's chain-of-thought straight into the bubble (the web UI hides it). It also sent no max_tokens cap.
• /no_think echo: reasoning was disabled by literally injecting the string /no_think into the system message; models that don't honour the soft-switch parroted it back.

Fixes

• VS Code: new squish.agentMode setting (default off). Off = plain streaming chat (fast TTFT, matches web UI / SquishBar); on = the multi-round tool loop. Text-mode tool-call detection is gated behind agent mode so a JSON-looking reply is never reinterpreted as a tool call.
• SquishBar: a streaming ThinkFilter strips <think>…</think> from the token stream (mirroring the web UI), and the request now sends max_tokens/temperature.
• Server: disable reasoning via the chat template's enable_thinking=False flag (_apply_chat_template) instead of injecting literal text. New strip_think_directives() in serving/tool_calling.py is applied on the /v1/chat/completions (streaming + non-streaming) and Ollama-compat (/api/generate, /api/chat) responses, so /think / /no_think / /nothink never reach any client regardless of model or endpoint.

Type of change

• Bug fix
• New feature
• Performance improvement
• Documentation update
• Refactor / cleanup

Checklist

• pytest tests/ passes locally (serving + ollama + openai-compat + server-unit suites green; full MLX suite needs Apple Silicon)
• ruff check reports no errors on changed files (ruff format clean on new files)
• No hardcoded absolute paths
• No model weights, eval output files, or log files staged for commit
• Changes are scoped to one logical concern (UI parity + directive leak)
• Performance-sensitive changes include a before/after squish bench run — N/A; the speedup is the server no longer forcing non-streaming for VS Code and reasoning no longer being dumped/echoed (no kernel-path change). Best validated live on Apple Silicon.

Validation

• New tests/serving/test_strip_think_directives.py (directive removal + false-positive guards).
• New VS Code test: plain-chat mode streams without tools.
• VS Code: tsc compiles clean; 180/180 jest tests pass.

Related issues

🤖 Generated with Claude Code

---

Generated by Claude Code