Major restructuring of the agent stack to support content-type skills,
data-driven runtime budgets, and the new translate skill.
LLM/MLX/
- Split MLXLocalLLMService.swift (1500 → 700 lines) into Core/, Gemma4/,
Installation/ subdirectories
- New BundledModel: LinearBudgetFormula replaces tier tables, model
parameters fully data-driven via ModelRuntimeProfile
- New RuntimeBudgets: pure functions for text/thinking/multimodal budgets,
zero hardcoded model IDs
- Drop adjustedTextOutputTokens prepared-subtraction (chunked prefill made
it obsolete and triggered false-positive truncation on long prompts)
Skills/
- Add SkillType (.device / .content) metadata, parsed from SKILL.md
frontmatter
- New translate skill (type: content, no tools) with anaphora handling
- Trim translate (4563→1804), device (2357→1680), calendar (2169→1790)
- contacts SKILL.md adds multi-turn disambiguation + cancel guidance,
semantics owned by body not framework
- SkillRegistry validates allowed-tools against ToolRegistry at load time
Agent/
- ToolCallParser becomes registry-aware: drops tool_call names not in
ToolRegistry or protocol set (load_skill / list_skills); ToolRegistry
is now the single white list, validation moves to parser boundary
- Router sticky routing uses semantic user-turn boundary instead of
fixed suffix(4); correctly handles long agent loops
- AgentEngine system prompt template differentiates device vs content
skills with explicit invocation rules
- SYSPROMPT.md auto-migration: old single-placeholder template backed up
to .bak when new dual-placeholder is detected
- PromptBuilder.buildLoadedSkillPrompt receives availableTools list,
explicitly tells model which tools (or none) are callable
Tools/
- ToolRegistry split into Helpers/Stores/AppPermissions/Handlers/* for
separation of concerns
- Mark Tools/Handlers/* with explicit allowed argument signatures
UI/
- Add long-press copy contextMenu on user and assistant bubbles
- Enable .textSelection on user bubble
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The 'optimization' broke model behavior — Gemma 4 needs
the literal <tool_call> example in system prompt to understand
the correct output format. Reverted to original.
- Remove literal <tool_call> block from system turn
(was being carried into buildFollowUp systemBlock, confusing
the model into thinking it had already made a tool call)
- Condense instructions from ~110 tokens to ~55 tokens
(halves prefill cost for the system block in every follow-up round)
- Rewrite as bullet rules instead of paragraph prose
