feat: multi-agent orchestration — GOAP, AgentGraph, ManagerAgent, Swarm, Memory, WebPipeline#690
Open
cluo wants to merge 15 commits intoHmbown:mainfrom
Open
feat: multi-agent orchestration — GOAP, AgentGraph, ManagerAgent, Swarm, Memory, WebPipeline#690cluo wants to merge 15 commits intoHmbown:mainfrom
cluo wants to merge 15 commits intoHmbown:mainfrom
Conversation
… Thread metadata - protocol: Add AgentActivityStatus 6-state enum (idle/thinking/tool_calling/ speaking/error/spawning) + AgentZone 5-zone enum (desk/hot_desk/lounge/ meeting/corridor) - protocol: Add Thread.zone + Thread.activity fields - protocol: Add new EventFrame variants: ThreadActivityChanged, ThreadZoneChanged, ThreadSpawned, CollaborationLink - state: SQL migration ALTER TABLE threads ADD COLUMN zone/activity TEXT - state: Full CRUD support (SELECT/INSERT/UPDATE/row_to_thread) - core: spawn_thread_with_history sets zone=Corridor, activity=Spawning - core: to_protocol_thread serializes zone/activity from persisted metadata Absorbed from openclaw-office L4 strategy (iteration-001).
…arm, GOAP, A2A, profiles Phase 3 — Agent orchestration: - AgentMiddleware trait + MiddlewareChain with @Next/@Prev anchors - AgentRouter (keyword registry + LLM fallback capability routing) - CompactionGuard middleware (80% threshold auto-compaction signal) - ModelRouter (task-type → optimal model selection) Phase 4 — Multi-provider LLM: - ProviderKind expanded with Anthropic, Google, Microsoft - Unified call_llm/stream_llm wrappers with RetryConfig - ThinkingCapability (5 modes) + ModelMetadata - ProxyAgent with dual-base-URL + auto-failover Phase 5 — Multi-agent coordination: - SwarmTopology (Queen/Mesh/Adaptive) + SwarmConfig - ConsensusProtocol (Raft/Byzantine/Gossip) - DebateRole + DebateConfig for two-agent adversarial reasoning Phase 6 — Planning & profiles: - GoapPlanner with A* search (GOAP) - A2aMessage + A2aAdapter protocol - AgentProfile + ProfileRegistry (6 built-in profiles) - SubAgentLifecycle, pause/resume, spawn dry_run/concurrency - Plan checkpoint/restore with Paused/Interrupted states - NoteTool enhanced with JSON Lines + SHA-256 dedup
- Rename ThinkingMode::Adaptive → Auto - Add ReasoningEffort (Off/Low/Medium/High/Max) with to_thinking_mode() - Add TaskComplexity classifier (Lookup/CodeGen/Architecture/Debug) - Add AutoReasoningSelector evaluating 4 signals: 1. Task complexity → base effort + model 2. Cache pressure (< 30% hit → drop one tier) 3. Cost budget (> 80% spent → throttle to Low) 4. Sub-agent fanout → always route to flash, cap at Medium - Explicit override always wins (manual Shift+Tab toggle respected) - Audit-logging via tracing events on each auto selection - 12 tests covering classification, throttling, override, fanout
Add HandoffPrompt — structured agent context transfer as a default replacement for routine compaction: - HandoffPrompt captures summary, progress, blockers, open_loops, working_set, model, timestamp, and token count - render() produces a stable cache-prefix-ready prompt for the successor agent - Builder pattern: with_summary, add_completed, add_in_progress, add_pending, add_blocker, add_open_loop, add_file, with_pct_complete - Serialize/deserialize for session persistence - Integrated with CompactionGuard: when context threshold crossed, handoff prompt can be generated instead of compacting - Preserves V4 prefix-cache discount (~90%) vs compaction's ~10× cost - 3 tests + round-trip serialization Also export HandoffProgress via subagent module.
Gap 1 — 128-token boundary alignment: - Add estimate_tokens() and pad_to_128_token_boundary() to prompts - Conservative 4-char/token heuristic for English prose - Newline-padding so sections don't shift across cache boundaries - 4 deterministic alignment tests Gap 4 — Strict tool mode capability: - Add StrictToolMode to ToolCapability enum in crates/tools - Tools can opt-in for schema-validated tool-call JSON - Recommended for mutating tools (apply_patch, exec_shell, write_file)
Add fim_edit tool: - Takes path + prefix_anchor + suffix_anchor - Splits file at anchors, builds FIM prompt for hole-filling - Supports line numbers and text-content anchors - Falls back to apply_patch when FIM endpoint unavailable - StrictToolMode capability for validated tool-call JSON - 4 tests: anchor finding, prompt structure, schema validation Also fix: unused import warnings in goap.rs and subagent/mod.rs
Add Auto variant to the TUI's ReasoningEffort enum: - Shift+Tab cycle: Off → High → Max → Auto → Off - api_value() returns None for Auto — engine fills via selector - config_ui ReasoningEffortValue with Auto conversions - ReasoningEffort::from_setting accepts 'auto' Now if reasoning_effort = "auto" in config, the app starts in Auto mode and the engine delegates to AutoReasoningSelector from llm::reasoning for per-turn effort + model selection.
…peline tools - agent_graph: LangGraph-style state machine with conditional routing, checkpoints, streaming events. Register as graph_execute tool. - agent_memory: Short-term (FIFO), long-term (persistent), entity (KV) memory. Register as memory_store + memory_search tools. - manager_agent: CrewAI-style hierarchical delegation with create → assign → dispatch → record → synthesize lifecycle. Register as manager_delegate tool. - web_pipeline: Deep crawl (BFS/DFS/BestFirst), BM25 content filter, smart chunking (sentence/paragraph/heading/fixed). Register as web_pipeline tool. - goap: add GoapPlanTool (ToolSpec) with 12 default dev actions - swarm: add tool module wiring All 18 unit tests pass. #![allow(dead_code)] applied to library-style API surface.
…gent runtime - engine.rs: shared token_counter, build_middleware_chain(), update_token_counter(), build_handoff_prompt() with plan/todo/ working-set population - subagent runtime: AgentRouter for auto-routing spawn type, ModelRouter for per-type model selection, ProfileRegistry for named profiles (code-reviewer, architect, etc.) - middleware: CompactionGuard reads shared token_counter, flags compaction_needed in MiddlewareContext after each step, emits progress event when threshold crossed - profiles: resolve at spawn time, inject posture_prompt into agent's first message - token counter updates on every API response (engine + sub-agents) Fixes Hmbown#664 (context-limit handoff).
- base.md Toolbox: add goap_plan, swarm_init, manager_delegate, memory_store/memory_search, graph_execute, web_pipeline entries - AGENTS.md: new Multi-Agent Orchestration section with composition pattern, key files, and hierarchical layer table
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
Wraps the agent-browser CLI (vercel-labs/agent-browser) as a first-class DeepSeek-TUI tool. The tool is only registered when agent-browser is found on PATH (checked once, cached via OnceLock). Actions covered: - navigate: open a URL - snapshot: accessibility tree with @en refs (interactive + compact) - click, fill, type_text: interact with elements by ref/selector - screenshot: PNG/JPEG with optional full_page + annotate flags - get_text, get_url, get_title: extract page data - wait: element/URL/text/load/JS condition - evaluate: run arbitrary JavaScript via agent-browser eval --stdin - press_key, scroll, close Output parsing: JSON (data.* nested + top-level) + fallback text extraction for screenshot paths. For advanced agent-browser commands (tabs, network interception, auth vault, React DevTools), the model falls back to exec_shell. Docs: base.md Toolbox + AGENTS.md Browser Automation section added. Tool registered in Plan and Agent modes via with_browser_tool(). 9 unit tests pass.
Wraps the obscura CLI (h4ckf0r0day/obscura) as a ToolSpec for fast, stealthy JS-rendered page extraction. The tool is only registered when obscura is found on PATH (cached via OnceLock). Actions covered: - fetch: JS-rendered page dump (text/html/links) with optional eval - scrape: parallel multi-URL extraction with concurrency control - serve: background CDP WebSocket server for Puppeteer/Playwright Key features: - 30 MB memory / instant start vs Chrome's 200 MB / ~2s - Built-in stealth mode: fingerprint randomization + 3,520-domain tracker blocking - DOM->Markdown via LP.getMarkdown CDP domain - Complementary to agent-browser: obscura for stealth/scrape, browser tool for interactive clicks/forms/screenshots Docs: base.md Toolbox + AGENTS.md Headless Browser section with comparison table. Registered in Plan and Agent modes. 4 unit tests pass.
…ive_default) Auto-applied via . - agent_graph: #[derive(Default)] for GraphState, is_some_and() - browser: simplify .map_err(|e| ToolError::execution_failed(e))? to .map_err(ToolError::execution_failed)? - obscura: same .map_err simplification - subagent/mod: if let chains with && let, #[derive(Default)] SubAgentLifecycle - subagent/router: remove ref from pattern - fim_edit: if let chain - shell: if let chains in NoteTool - agent_memory: if let chains
…agency_agents()
Integrates the msitarzewski/agency-agents library (92.9k stars, 144+
specialized AI agent personalities) with DeepSeek-TUI's profile system.
Added:
- ProfileRegistry::import_agency_agents(dir) — bulk import from directory
of .md files with YAML frontmatter
- parse_yaml_frontmatter() — zero-dependency YAML frontmatter parser
- slugify() — converts 'Frontend Developer' → 'frontend-developer'
- format_agency_posture() — formats body as agent posture prompt
- infer_agent_type() — maps division/content keywords to SubAgentType
Division mapping: Engineering→Implementer, Testing→Verifier,
Design→Implementer, Review→Review, Strategy/Product/PM→Plan, others→General
Usage:
git clone agency-agents ~/.deepseek/agency-agents/
agent_spawn { profile: "frontend-developer", prompt: "..." }
agent_spawn { profile: "security-engineer", prompt: "..." }
Docs: AGENTS.md Agent Profiles section + base.md Toolbox reference
14 unit tests pass (8 existing + 6 new).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds six new orchestration tools that compose hierarchically:
New tools
goap_plangraph_executeswarm_initmanager_delegatememory_store/memory_searchweb_pipelineSub-agent runtime enhancements
Key files
crates/tui/src/tools/agent_graph.rs— newcrates/tui/src/tools/agent_memory.rs— newcrates/tui/src/tools/manager_agent.rs— newcrates/tui/src/tools/web_pipeline.rs— newcrates/tui/src/core/engine.rs— token counter, handoff prompt, middleware wiringcrates/tui/src/tools/subagent/mod.rs— router/profile/middleware wiringcrates/tui/src/tools/registry.rs— builder methodscrates/tui/src/core/engine/tool_setup.rs— all 6 tools registered in Plan + Agent modescrates/tui/src/prompts/base.md— Toolbox section updatedAGENTS.md— Multi-Agent Orchestration section addedTests