feat: improve task execution, verification and packaging
Add execution incidents, action bursts and shared Flash history. Expose verification and Explorer settings with streamed check results. Support graceful cancellation, recording recovery and bundled UI resources. Add CI quality checks and clean up comments and documentation.
This commit is contained in:
somew1nd committed
1 parent
5ccf06d677
commit
67248c887e
289 files changed
+15883
-6919
No files matched your search
+34
-35
@@ -8,11 +8,11 @@
|
||||
// ------------------------------------------------------------------------------
|
||||
"default": {
|
||||
"provider": "google",
|
||||
"model": "gemini-3.7-flash",
|
||||
"model": "gemini-3.8-flash",
|
||||
"thinking_level": "medium",
|
||||
"fallback": {
|
||||
"provider": "google",
|
||||
"model": "gemini-3.6-flash",
|
||||
"model": "gemini-3.7-flash",
|
||||
"thinking_level": "medium"
|
||||
}
|
||||
},
|
||||
@@ -23,8 +23,8 @@
|
||||
"presets": {
|
||||
"gemini-flagship": {
|
||||
"provider": "google",
|
||||
"model": "gemini-3.7-flash",
|
||||
"fallback": { "provider": "google", "model": "gemini-3.6-flash" }
|
||||
"model": "gemini-3.8-flash",
|
||||
"fallback": { "provider": "google", "model": "gemini-3.7-flash" }
|
||||
},
|
||||
"gemini-flash": {
|
||||
"provider": "google",
|
||||
@@ -89,11 +89,6 @@
|
||||
"thinking_level": "medium"
|
||||
},
|
||||
|
||||
// 🛡️ Validator & Failure Analyzer: Self-healing, action verification & recovery
|
||||
"validator_failure_analyzer": {
|
||||
"thinking_level": "medium"
|
||||
},
|
||||
|
||||
// 📋 Summarizer: Aggregates step history and generates structured outcome
|
||||
"summarizer": {},
|
||||
|
||||
@@ -121,8 +116,15 @@
|
||||
// 🏛️ Committee Council Members
|
||||
"planner_avatar": {},
|
||||
"history_analyzer_expert": {},
|
||||
// 🔎 Checker: Visual subgoal verification and rollback
|
||||
"checker": {}
|
||||
// 🔎 Checker: Plan-declared checkpoint verification and exit final review.
|
||||
// Thoughts are streamed so the verifier's reasoning shows live in the timeline.
|
||||
"checker": {
|
||||
"thinking_level": "medium",
|
||||
"include_thoughts": true
|
||||
}
|
||||
// 🎯 Planner Validation: Advisory review of milestone changes.
|
||||
// Defaults to the same lightweight model as the pixel safety net
|
||||
// (gemini-3.5-flash-lite, temperature 0); add "planner_validation": {...} here to override.
|
||||
},
|
||||
|
||||
// ------------------------------------------------------------------------------
|
||||
@@ -132,11 +134,11 @@
|
||||
// ==========================================================================
|
||||
// ⚡ 4.1 FLASH Profile Options (Reactive Single-Agent Lightweight Loop)
|
||||
// ==========================================================================
|
||||
// Best for: Simple, fast, deterministic tasks (< 30 UI steps).
|
||||
// Best for: Simple, fast, deterministic tasks; runs unbounded by default.
|
||||
// Execution: Observe-Think-Act single LLM reactive loop without graph overhead.
|
||||
"flash": {
|
||||
"max_turns": 30, // Maximum reactive turns before stopping
|
||||
"explorer_mode": "flash", // Perception mode for Flash runner ('flash' 1-shot detection)
|
||||
"max_turns": 0, // Maximum reactive turns before stopping (0 = unlimited; history is bounded by transcript compression, not by a turn cap)
|
||||
"explorer_mode": "flash", // Explorer tier for the Flash runner: 'flash' | 'pro' | 'ultra' (see top-level "explorer")
|
||||
// 🧠 Step Summarizer: Asynchronous visual context compressor
|
||||
// Summarizes intermediate steps in background without blocking the runner,
|
||||
// replacing images only after strictly objective, high-density summaries are ready.
|
||||
@@ -150,17 +152,16 @@
|
||||
// 🚀 4.2 PRO Profile Options (Deep Multi-Agent Graph Closed-Loop)
|
||||
// ==========================================================================
|
||||
// Best for: Complex, exploratory, dynamic multi-branch tasks.
|
||||
// Execution: LangGraph multi-node graph with autonomous planning, validation, and self-healing.
|
||||
// Execution: LangGraph multi-node graph with a living task plan, a pre-execution safety net (single actions) / fast-action bursts (multi-action turns),
|
||||
// execution-incident recovery handled by the Operator itself (no separate repair agent), and a read-only Checker for checkpoint verification + exit final review.
|
||||
"pro": {
|
||||
// 🧭 Explorer: Perception sub-agent under Pro profile ('flash', 'pro', or 'ultra')
|
||||
"explorer": {
|
||||
"mode": "flash", // Default mode for Pro nodes (Operator/Validator)
|
||||
"caching": true // Context caching for multi-turn Pro/Ultra visual reasoning
|
||||
"mode": "flash" // Explorer tier for Pro nodes (Operator/Validator): 'flash' | 'pro' | 'ultra'
|
||||
},
|
||||
// 🎯 Planner Validation: Async verification when task plan milestones change
|
||||
"planner_validation": {
|
||||
"enabled": true, // On by default: ratchet-baseline milestone validation
|
||||
"similarity_threshold": 0.85
|
||||
"enabled": true // On by default: every milestone text change is reviewed (advisory: hint + reason only, never rolled back)
|
||||
},
|
||||
// 🏛️ Committee: Multi-Agent Council debate tool for Operator
|
||||
"committee": {
|
||||
@@ -222,7 +223,7 @@
|
||||
"chunking": {
|
||||
"max_steps": 12,
|
||||
"target_source_tokens": 2000,
|
||||
"model": "gemini-3.7-flash",
|
||||
"model": "gemini-3.8-flash",
|
||||
"max_chunks": 8
|
||||
// "max_eras": 8 // independent era cap; omitted = follows max_chunks
|
||||
},
|
||||
@@ -245,25 +246,23 @@
|
||||
"enabled": true, // Whether Outputter is mounted to synthesize final outputs
|
||||
"force_synthesis": false // Force Outputter synthesis even without explicit output schema
|
||||
},
|
||||
// 🧭 Global / Per-Agent Explorer Version Overrides
|
||||
// 🧭 Explorer tier: the perception sub-agent behind `ask_explorer`.
|
||||
// Tiers are a user setting; calling agents never see which tier is active.
|
||||
// flash - one-shot visual detection on the screenshot (fastest)
|
||||
// pro - 3-turn reasoning loop: UI-tree search + coordinate audit + detection
|
||||
// ultra - 8-turn loop adding zoom, OCR and pixel-level image processing (slowest)
|
||||
// `flash.explorer_mode` and `pro.explorer.mode` above are the per-profile knobs;
|
||||
// `caching` (Gemini explicit context caching) defaults per tier: off for pro, on for ultra.
|
||||
"explorer": {
|
||||
"default_version": "flash", // Fallback version when not specified
|
||||
"flash_mode": "flash", // Version for Flash runner
|
||||
"pro_mode": "flash", // Version for Pro graph runner
|
||||
"caching": true
|
||||
},
|
||||
"explorer_versions": {
|
||||
"operator": "flash",
|
||||
"validator": "flash"
|
||||
"default_version": "flash" // Fallback tier when no profile knob applies
|
||||
// "caching": true, // Uncomment to force context caching for pro/ultra
|
||||
},
|
||||
// Advanced: per-agent tier override, e.g. {"validator": "ultra"}. Takes
|
||||
// precedence over the profile knobs; leave empty to let them decide.
|
||||
"explorer_versions": {},
|
||||
// 🚫 Per-agent Tool Denylists
|
||||
"denylisted_tools": {
|
||||
"explorer": [
|
||||
"ask_image_processor",
|
||||
"get_ocr_list",
|
||||
"inspect_region"
|
||||
]
|
||||
"explorer": [] // Perception tools to withhold from the Explorer, e.g. "ask_image_processor"
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
Reference in new issue
Block a user