feat(local-llm): prompt queue, shell safety, and main-thread performance

[FrostyPhoenix2]

Adds a complete workflow for running local small models (Qwen, Mistral, etc.) through multi-step project builds without browser lock-up or WebContainer overwhelm. Validated end-to-end with Qwen 2.5-coder:7b via Ollama against a 21-prompt React Native / Expo build queue.

Local LLM Settings Panel (new: LocalLLMPanel)

• Per-session toggles: extended stream timeout (3 min vs 45 s), shell command blocking, slim system prompt, dedup file writes, strip old prose, token budget
• Persisted to localStorage via nanostores map (localLLMSettingsStore)
• Flyout panel anchored above the chat input bar with dynamic max-height

Prompt Queue (PromptQueuePanel)

• Run N queued prompts end-to-end unattended; live progress bar
• Adaptive delay between prompts: polls action-runner until all file writes / shell commands reach a terminal state before firing the next prompt — prevents WebContainer being overwhelmed by rapid-fire writes
• Paste or load prompts from a newline-delimited text file

Shell Safety (action-runner)

• Blocklist for hanging commands: expo start, npm install, yarn, pnpm, node, react-native run — throws before execution with a clear message
• 90-second hard timeout on any shell action via Promise.race()
• Toggle via blockHangingCommands setting

Stream + Context Optimisations (api.chat)

• Extended stream timeout: 180 000 ms for local models (isLocalModel flag)
• Context optimisation re-enabled for local models with graceful fallback: LLM pre-pass wrapped in withTimeout(); keyword matching fallback on timeout
• Token limit clamped to model's reported max (fixes Qwen 8 000 token warning)

Main-Thread Performance (Chat.client)

• 4 sequential setMessages() calls in onFinish collapsed into one, wrapped in startTransition() — eliminates "Page Unresponsive" dialogs during queues
• storeMessageHistory deferred to requestIdleCallback (setTimeout fallback) so IndexedDB writes never block the render thread
• Resource-exhaustion errors (ERR_INSUFFICIENT_RESOURCES, Failed to fetch) demoted from toast flood to console.warn

Slim Prompt (prompts/slim)

• Minimal system prompt for 7B–13B models; hard_rules block first with DO NOT run install/server commands, DO NOT truncate files, DO NOT emit code outside boltArtifact — then concise format and behaviour rules
• Registered in PromptLibrary as "Slim Prompt (local models)"

Artifact Render-Loop Fix

• computed() store moved from inline JSX into useRef — fixes "Maximum update depth exceeded" on every workbenchStore update

ZIP Import (.bolt/prompt)

• Detects .bolt/prompt inside imported ZIPs; surfaces content as the auto-fill textarea value so projects can ship a first-message template

Misc

• useChatHistory snapshot save failure: toast → console.warn
• Token limit clamped in stream-text.ts (model.maxTokenAllowed guard)

Summary

Adds a complete, validated workflow for running local small models (Qwen, Mistral, etc.) through multi-step project builds without browser lock-up or WebContainer overwhelm. Tested end-to-end with Qwen 2.5-coder:7b via Ollama against a 21-prompt React Native / Expo build queue — clean completion, responsive UI throughout.

What's included

Local LLM Settings Panel (LocalLLMPanel) — new flyout above the chat input with per-session toggles: extended stream timeout, shell command blocking, slim system prompt, dedup file writes, strip old prose, token budget. All persisted to localStorage.

Prompt Queue (PromptQueuePanel) — paste or load N prompts, run them unattended with a live progress bar. Uses adaptive delay between prompts: polls the action-runner until all file writes/shell commands reach a terminal state before firing the next one, preventing WebContainer overwhelm.

Shell safety — blocklist in action-runner.ts for commands that hang WebContainer (expo start, npm install, yarn, pnpm, node, react-native run). 90-second hard timeout on all shell actions. Toggle in the settings panel.

Stream + context optimisations — 180 s stream timeout for local models; context optimisation re-enabled for local models with timeout wrapper + keyword-matching fallback; token limit clamped to model's reported max (fixes Qwen 8 000 token warning).

Main-thread performance — 4 sequential setMessages() calls in onFinish collapsed into one startTransition()-wrapped call, eliminating "Page Unresponsive" dialogs during long queues. storeMessageHistory deferred to requestIdleCallback so IndexedDB writes never block rendering.

Slim prompt — minimal system prompt for 7B–13B models with hard_rules block first. Registered in PromptLibrary as "Slim Prompt (local models)".

Artifact render-loop fix — computed() store moved from inline JSX to useRef, fixing "Maximum update depth exceeded" on every workbenchStore update.

ZIP import .bolt/prompt — detects a .bolt/prompt file inside imported ZIPs and surfaces its content as the auto-fill textarea value, letting projects ship their own first-message template.

Test plan

• Enable Ollama in provider settings, load a local model
• Open Local LLM panel — verify all toggles persist across page reload
• Import a ZIP with a .bolt/prompt file — textarea should pre-fill
• Queue 5+ prompts via PromptQueuePanel, run to completion — UI stays responsive
• Verify shell commands like npm install are blocked with a clear error message
• Send a chat with a non-local provider — verify nothing regressed