mirror of
https://github.com/tinyhumansai/openhuman.git
synced 2026-07-28 05:12:33 +00:00
fb4a88867a
* feat(skills): extend GenerateSkillSpec with full_index_js and update generator - Add optional `full_index_js` field to GenerateSkillSpec so LLM-generated JS source can be written verbatim instead of using the default template - Change generate_alphahuman() return type from PathBuf to Vec<PathBuf> (returns both manifest.json and index.js paths for audit logging) - Add UnifiedSkillRegistry::skills_dir() and engine() helper accessors needed by the self-evolve orchestrator Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat(skills): add isolated QuickJS skill tester with mock bridge globals Introduces SkillTester::run_isolated() which spins up a fresh rquickjs context (no shared state), injects no-op mock globals for all bridge APIs (db, net, state, platform, cron, skills), loads index.js, and calls init/start plus each tool's execute() with empty args. The test passes only if every exported tool returns without throwing and produces valid output. Used by the self-evolve loop to validate generated code before registering it into the live registry. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat(skills): add LLM code generator using Anthropic Claude API Introduces LlmGenerator with two methods: - generate_spec(task): first-pass generation from a natural-language task description; produces a complete GenerateSkillSpec including full_index_js - fix_spec(task, prev_code, error): retry pass that feeds the previous code and the test error back to the model for correction Uses claude-3-5-haiku-20241022 via direct reqwest POST to the Anthropic messages API. System prompt teaches the model the exact QuickJS index.js format, available bridge globals, and synchronous net.fetch semantics. Strips markdown code fences from the response before use. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat(skills): add self-evolve orchestrator with iteration loop and audit log Introduces SkillEvolver::evolve() which implements the full pipeline: 1. LLM generates a GenerateSkillSpec from a task description 2. generator::generate_alphahuman() writes manifest.json + index.js 3. SkillTester::run_isolated() validates the code in a fresh sandbox 4. On failure: feed error back to LLM and regenerate (up to max_iterations) 5. On success: discover_skills() + start_skill() to register live, then execute() to run the real task and capture final output 6. Returns SelfEvolveResult with full audit log, files_created, and the final UnifiedSkillResult Wraps the entire evolve() call in tokio::time::timeout (default 120s). Cleans up written skill directory on permanent failure so no partial skill is left registered. Key types: SelfEvolveRequest — task_description, max_iterations, timeout_secs, optional anthropic_api_key override SelfEvolveResult — skill_id, success, iterations_used, audit_log, files_created, final_result, failure_reason IterationLog — per-iteration code, test output, pass/fail flag Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat(tauri): register unified_self_evolve_skill command with progress events Adds unified_self_evolve_skill Tauri command that: - Accepts a SelfEvolveRequest (task_description + optional limits) - Emits skill:evolve:progress event after each iteration so the frontend can show live progress without polling - Delegates to SkillEvolver::evolve() and returns SelfEvolveResult - Mobile stub returns an informative error (feature requires desktop runtime) Registers the command in generate_handler![] in lib.rs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat(ui): add SelfEvolveModal and Auto-Generate button to SkillsGrid SelfEvolveModal (src/components/skills/SelfEvolveModal.tsx): - createPortal modal with 4 states: idle / running / success / failed - Text input for natural-language task description - Calls unified_self_evolve_skill via invoke() on submit - Subscribes to skill:evolve:progress Tauri events to show live iteration updates (iteration number, pass/fail indicator) - On success: displays audit log accordion with per-iteration code, test output, and final task result - On failure: shows failure reason and full audit trail for debugging - Calls onSkillCreated() callback to trigger grid refresh SkillsGrid changes: - Add selfEvolveOpen state + "Auto-Generate" button in the toolbar - Extract refreshSkills as a named async function reused by both the initial load and the SelfEvolveModal's onSkillCreated callback - Add activeSkillType state so SkillSetupModal shows the correct badge when opened from either the Connect or Manage flow Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(skills): fix alphahuman skill template and generation pipeline - generator.rs: use bare `tools = []` (not const) for globalThis access, add `execute: async function(args)` method, fix `input_schema` casing - mod.rs: call start_skill() after generate() so skills are immediately executable - llm_generator.rs: update model to claude-haiku-4-5-20251001, sync system prompt Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(skills): fix tools not loading for generated skills and add robustness fixes - llm_generator: require 'var tools = [...]' in system prompt (const/let are block-scoped and do not attach to globalThis, causing extract_tools() to find an empty array in the production QuickJS engine) - llm_generator: add smart_name() — deterministic display name from task description (strips preambles + lead fillers + stop words, title-cases 3 tokens); replaces brittle LLM-dependent naming - llm_generator: improve sandbox testing rules in system prompt so placeholder values prevent test failures when net.fetch returns empty JSON - skill_tester: detect {error: '...'} return values as test failures (both sync and Promise paths) so the iteration loop attempts a fix instead of treating them as success - manager.ts: dispatch setSkillSetupComplete(true) for skills without a setup flow so deriveConnectionStatus() returns "connected" instead of "connecting" Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>