Skip to content
DuDuClaw OSDuDuClaw OSDuDuClaw OS Docs

DuDuClaw Architecture Overview

Architecture Overview (reviewed against v1.67.0)

Section titled “Architecture Overview (reviewed against v1.67.0)”

DuDuClaw is a Multi-Runtime AI Agent Platform — supporting Claude Code / Codex / Antigravity / Grok CLI (plus OpenAI-compatible APIs) as AI backends via a unified AgentRuntime trait with auto-detection and per-agent configuration (Gemini CLI has been removed; Antigravity remains supported). DuDuClaw is not a standalone LLM product; it is the plumbing layer that turns one (or many) AI CLIs into long-running agents with channel routing, session memory, self-evolution, multi-account rotation, local LLM inference, browser automation, and IDE integration.

Key Architectural Decisions

Section titled “Key Architectural Decisions”

Runtime & Transport

Section titled “Runtime & Transport”
  • Multi-Runtime (AgentRuntime trait) — 12 runtime ids in runtime_catalog.rs: eleven CLI backends (Claude, Codex, Antigravity (agy), Grok, Qwen Code, Kimi Code, GitHub Copilot CLI, Kiro, Cursor, Mistral Vibe, OpenCode) plus OpenAI-compat HTTP. Four CLIs and OpenAI-compat have their own modules; the other seven CLIs share one generic print-mode runtime. RuntimeRegistry auto-detection, per-agent config in agent.toml [runtime] provider.
  • Cross-runtime failover model substitution (failover.rs): when [runtime] fallback sends a call to a different provider, the fallback runtime no longer inherits the primary’s model id (a codex agent’s gpt-5.4 used to be spawned against the Claude runtime). The model is resolved in four ordered branches — ① the first agent.toml [model] fallbacks entry whose family confidently belongs to the fallback runtime (qualified provider/model entries are unqualified), ② keep the requested model when that runtime already serves it (this is how openai_compat, which declares no model family, keeps proxying arbitrary ids), ③ the runtime catalog’s own first model for that backend, ④ otherwise refuse to spawn and report no model configured for fallback runtime <P>, recorded as a failed attempt. Every substitution emits a warn! carrying agent / from_runtime / to_runtime / from_model / to_model.
  • MCP Server (stdio) (duduclaw mcp-server) exposes channel, memory, agent, skill, task, shared wiki, and autopilot tools to AI Runtime via JSON-RPC 2.0 over stdin/stdout. Registered at the agent level in <agent>/.mcp.json (v1.8.5 reverted v1.8.4’s global registration because Claude CLI -p --dangerously-skip-permissions only reads project-level .mcp.json). Gateway startup auto-creates/repairs .mcp.json for all agents.
  • MCP Server (HTTP/SSE) (duduclaw http-server --bind 127.0.0.1:8765, v1.9.4) — Bearer-authenticated POST /mcp/v1/call (single JSON-RPC tool call), GET /mcp/v1/stream (SSE long-lived event stream, Bearer / ?api_key=), POST /mcp/v1/stream/call (async + SSE result push), GET /healthz (no auth). Token bucket rate limit (60 req/min). mcp_sse_store.rs manages SSE connections with broadcast channels. Complements stdio for external HTTP clients.
  • ACP and A2A — duduclaw acp (= duduclaw acp client) implements Agent Client Protocol v1 over stdio for Zed / JetBrains / Neovim agent panels. duduclaw acp server (formerly acp-server, removed in v1.69.0) serves A2A: message/send appends to bus_queue.jsonl, tasks/get maps bus observations back to A2A states, and the Agent Card is at /.well-known/agent-card.json (legacy /agent.json alias).
  • Agent directories are Claude Code compatible: each contains .claude/, .mcp.json, SOUL.md, CLAUDE.md, CONTRACT.toml, agent.toml, wiki/, SKILLS/, memory/, tasks/, state/.

Channels (11)

Section titled “Channels (11)”
  • Telegram (long polling) — file/photo/sticker/voice, forums/topics, mention-only, voice transcription through the OpenAI Whisper API.
  • LINE (webhook) — HMAC-SHA256 signature, sticker catalog, per-chat settings.
  • Discord (Gateway WebSocket) — tokio::select! heartbeat, slash commands, auto-thread; voice channels (Songbird) only in builds with the non-default discord-voice feature, which release binaries do not include. v1.9.2 hardened: real op 6 RESUME (persists session_id + resume_gateway_url + sequence), stall watchdog (break if no traffic for 2× heartbeat interval), heartbeat capacity 1→16 with try_send, op 9 jitter 1-5s, RESUMED dispatch handling, backoff cap 60s.
  • Slack (Socket Mode), WhatsApp (Cloud API webhook, signature fail-closed), Feishu (Open Platform v2), Google Chat (webhook, JWT-verified), Microsoft Teams (Azure Bot / Connector v3, JWT-verified), WeCom (HMAC-SHA1 + AES-256-CBC), DingTalk (HMAC-SHA256 + time window), WebChat (/ws/chat + React frontend).
  • The generic POST /webhook/{agent_id} endpoint was never mounted and was removed in v1.66; inbound webhooks belong to individual channels and to Odoo (POST /webhook/odoo).
  • Channel hot-start/stop: Dashboard channels.add / channels.remove launches/aborts the channel task without gateway restart.
  • Media pipeline: image auto-resize (max 1568px) + MIME detection + Vision integration.

Sub-Agent Orchestration

Section titled “Sub-Agent Orchestration”
  • create_agent / spawn_agent / list_agents MCP tools with reports_to hierarchy.
  • System prompt auto-injects “## Your Team” sub-agent roster.
  • Structured handoff: DelegationEnvelope (context / constraints / task_chain / expected_output) with Raw fallback.
  • TaskSpec workflow: multi-step task planning with dependency-aware scheduling, auto-retry (3x), replan (2x), persistence.
  • Long-response splitting: Sub-agent replies wider than the channel byte budget are split via channel_format::split_text with 📨 **agent** 的回報 (1/N) / (續 2/N) labels (Discord 1900 / Telegram 4000 / LINE 4900 / Slack 3900).
  • Orphan response recovery: reconcile_orphan_responses atomically replays bus_queue.jsonl entries left behind by crash / Ctrl+C / hotswap.

Session Memory Stack

Section titled “Session Memory Stack”
  • Native multi-turn: Claude CLI --resume <session-id> with SHA-256 deterministic session ID; auto-fallback to history-in-prompt when --resume fails (stale handle, account rotation, unknown stream-json error).
  • Turn trimming (>800 chars → head 300 + tail 200 + [trimmed N chars], CJK-safe).
  • Direct API prompt cache (“system_and_3” breakpoint strategy; no measured hit rate is published).
  • Compression summaries injected into system prompt (not conversation turns) at 50k token threshold.
  • Instruction Pinning (v1.8.6 P0) — first user turn → async Haiku extraction of core task → stored in sessions.pinned_instructions → injected at system prompt tail (U-shaped attention). Clarification answers accumulate (≤1000 chars).
  • Snowball Recap (v1.8.6 P0) — every turn prepends <task_recap> to the user message. Zero LLM cost.
  • P2 Key-Fact Accumulator (v1.8.6) — per substantive turn, Haiku extracts 2-4 key facts → key_facts table with FTS5 → top-3 relevant facts injected into system prompt. ~100-150 tokens vs MemGPT’s 6,500 (−87%).
  • CLI lightweight path — call_claude_cli_lightweight() with --effort medium --max-turns 1 --no-session-persistence --tools "" for metadata tasks. 25-40% cost reduction.
  • Stabilization flags — --strict-mcp-config (MCP isolation) + --exclude-dynamic-system-prompt-sections (cross-turn prompt stability, 10-15% token reduction). --bare was removed in v1.8.11 (broke OS keychain credential lookup).

Evolution

Section titled “Evolution”
  • Prediction-driven engine: Active Inference + Dual Process Theory; by design most conversations end without an LLM call (no measured share is published). Moderate errors are stored as episodic memory; Significant errors start an AEE evolution round; Critical errors (or three Significant in a row) start an emergency round; a small exploration share of Negligible errors (ε, hard floor 5%) also starts one.
  • MetaCognition: self-calibrating error thresholds every 100 predictions.
  • MistakeNotebook: cross-loop error memory prevents regression; entries now carry deterministic TrajectoryEvidence (which tool/assertion failed) so reflection consolidation stops trusting unverified self-reported diagnosis (Evolution v3).
  • SOUL.md is read-only for agents (Evolution v3, WP1.1 — operator/dashboard writes only). The legacy Generator→Verifier→Updater rewrite path, SOUL.md versioning, the 24h observation window and automatic rollback were removed on 2026-09-29 (S11): with nothing able to write the file on an agent’s behalf, they guarded a path that no longer existed.
  • AEE (Agentic Evolution Engine, the only evolution engine): the evolution target is not SOUL.md but a playbook of small, independently-retirable gene-shaped entries (category/signals/eval-case-linked), evolved through a Gate (deterministic, veto-keeping) / Measure (scored, no veto) split, a champion + matches-or-improves commit gate, and entry-level (not whole-file) observation windows. See evolution-engine.md ch.12 and docs/features/38-aee-playbook-evolution.md.
  • Agent-as-Evaluator: independent Evaluator Agent (Haiku cost control) for adversarial verification with structured JSON verdicts.
  • ConversationOutcome: zero-LLM conversation result detection (TaskType / Satisfaction / Completion) in zh-TW + en.
  • External factors: user feedback, security events, channel metrics, Odoo business context, peer agent signals feed into the prediction engine and evolution rounds.

Wiki Knowledge Layer (v1.8.9)

Section titled “Wiki Knowledge Layer (v1.8.9)”
  • 4-layer architecture (Vault-for-LLM inspired): L0 Identity / L1 Core / L2 Context / L3 Deep.
  • Trust weighting (trust 0.0-1.0 frontmatter) — search results ranked by trust-weighted score.
  • Auto-injection: build_system_prompt() auto-injects L0+L1 pages into WIKI_CONTEXT across CLI / channel reply / dispatcher paths — unified across the Claude / Codex / Antigravity / Grok / OpenAI-compat runtimes.
  • FTS5 index (unicode61 tokenizer) — auto-syncs on every write/delete, manual rebuild via wiki_rebuild_fts.
  • Knowledge graph: wiki_graph MCP tool exports BFS-limited Mermaid diagrams; node shapes by layer.
  • Dedup detection: wiki_dedup detects duplicate pages by title match + tag Jaccard similarity (≥0.8).
  • Reverse backlink index: scans related frontmatter + body markdown links for bidirectional mapping.
  • Search filters: wiki_search (either scope) supports min_trust, layer, expand (1-hop backlink expansion).
  • Shared Wiki: ~/.duduclaw/shared/wiki/ for cross-agent SOPs, policies, product specs. Visibility controlled via wiki_visible_to capability.

Memory System

Section titled “Memory System”
  • Cognitive memory (optional): SqliteMemoryEngine with episodic/semantic separation and Generative Agents 3D-weighted retrieval (Recency × Importance × Relevance).
  • Memory decay daily scheduler: background task runs duduclaw_memory::decay::run_decay every 24h. Low-importance + 30 days old → archived. Archived + 90 days → permanent delete.
  • Cognitive memory MCP tools: memory_search_by_layer (episodic/semantic filter), memory_successful_conversations, memory_episodic_pressure, memory_consolidation_status.
  • MemGPT 3-layer system (Core Memory, Recall Memory, Archival Bridge, Budget Manager, Consolidation Pipeline, 6 MCP tools) was removed in v1.8.1 (−1,985 LOC) — the prompt injection caused 6,500 token bloat per prompt and “lost in the middle” attention degradation.

Local Inference

Section titled “Local Inference”
  • Unified InferenceBackend trait (duduclaw-inference crate): OpenAI-compatible HTTP (llama-server/Ollama/vLLM/SGLang/llamafile). The in-process llama.cpp, mistral.rs and MLX backends were removed in 2026-09 — no release binary ever compiled them; run a local OpenAI-compatible server instead.
  • Confidence Router: three-tier LocalFast / LocalStrong / CloudAPI routing, CJK-aware token estimation.
  • InferenceManager: auto-switching state machine — llamafile → Direct backend → OpenAI-compat → Cloud API.
  • llamafile manager: subprocess lifecycle, health monitoring, OpenAI-compatible API on localhost.
  • MCP tools: model_list, model_load, model_unload, inference_status, hardware_info, route_query, inference_mode, llamafile_start/stop/list.

Token Compression

Section titled “Token Compression”
  • Reply-path budget pipeline (gateway/prompt_compression.rs): TurnTrim → DropOldestToolEchoes → BisectAndSummarize, cost-pressure aware, CJK-safe token estimation; skipped when recent cache efficiency is above 50% and the budget overshoot is below 15%.
  • The earlier Meta-Token (LTSC) / LLMLingua-2 / StreamingLLM compressor and its compress_text / decompress_text tools were removed in v1.33.

Voice Pipeline

Section titled “Voice Pipeline”
  • HTTP endpoints: POST /api/stt (an OpenAI-compatible transcription API or a local command template; 501 when no STT provider is configured), POST /api/tts over Piper (local) / Edge TTS / MiniMax T2A / OpenAI TTS.
  • Telegram voice: transcription through the OpenAI Whisper API and Edge TTS replies, both hardcoded; the dashboard voice settings do not reach this path.
  • Not in release binaries: in-process Whisper (whisper feature), ONNX embedding (onnx feature) and Discord voice (discord-voice feature).
  • SenseVoice, Deepgram, Silero VAD, symphonia decoding and LiveKit voice rooms were listed here before; none of them has code. See docs/features/14-voice-pipeline.md.

Security

Section titled “Security”
  • Claude Code PreToolUse hooks (installed per agent by agent_hook_installer into <agent_dir>/.claude/settings.json): duduclaw hook agent-file-guard (Rust subcommand, matcher Write|Edit|MultiEdit|NotebookEdit|Bash, command carries --agent and --home — agent-structure files outside the canonical tree, own-SOUL.md / own-CONTRACT.toml writes, cross-agent writes, and for an employee caller any write under the DuDuClaw home outside its own agent directory and attachments/, judged on the real path after symbolic links; plus the org_field_guard field-level freeze on reports_to / department / name / [capabilities] / [delegation] / [acp] and, for employee callers, every section of their own agent.toml outside the editable list; the Bash lane is a speed bump and Read is not covered — rules and limits in 05-security-defense.md) and duduclaw hook data-file-guard (RFC-23 §14.4, matcher Read|Bash, armed only when redaction is active; a Bash filename heuristic, not a sandbox — H10 2026-09 replaced the shell script that was inert on Windows). The 2026-04 three-phase shell-script defense and its GREEN/YELLOW/RED threat-level state machine were removed in ba015a48 — see docs/features/05-security-defense.md.
  • SOUL.md drift detection (SHA-256 fingerprint, ≤10 versioned backups in .soul_history/).
  • Prompt injection scanner (input_guard, 11 rule categories, block threshold 60, NFKC-normalized, en + zh-TW patterns, XML delimiter protection).
  • Secret leak scanner — 19 secret patterns (Anthropic / OpenAI / AWS / GitHub / GitLab / Slack / Stripe / Google / SendGrid / JWT / PEM keys and key or password assignments) plus a high-entropy check, used by the skill security scanner.
  • CONTRACT.toml — must_not / must_always boundaries, auto-injected into system prompt; duduclaw test red-team CLI (9 built-in scenarios).
  • Unified multi-source audit log: audit.unified_log merges security_audit.jsonl / tool_calls.jsonl / channel_failures.jsonl / feedback.jsonl into common envelope (timestamp / source / event_type / agent_id / severity / summary / details) with Logs page filter chips.
  • AES-256-GCM at rest — per-agent key isolation.
  • Dashboard / WebSocket auth: JWT account login (Argon2id-hashed passwords in users.db) or the gateway admin token. An earlier Ed25519 challenge-response path has been removed from the gateway; no configuration could ever enable it. Ed25519 is still used for licence signatures, update verification and the relay device protocol.
  • Container sandbox — two separate paths. The per-agent task sandbox (agent.toml [container] sandbox_enabled) runs a delegated task’s AI CLI in a read-only, non-root, resource-limited Docker container (Docker only, needs network_access = true, fails closed when it cannot run; see Task sandbox guide). The script sandbox used by PTC execute_program and the secaudit PoC step runs on Docker (WSL2 first on Windows) with --network=none, a read-only root and only a private read-only script directory mounted; when it cannot run, PTC refuses the script unless [container.sandbox] script_when_unavailable = "run_unsandboxed", and the PoC never runs on the host.
  • Browser automation & computer use — two fetch tools and an optional browser server the agent chooses between, no auto-router: L1 web_fetch_cached (SSRF-gated, cached HTTP), L2 web_extract (CSS selector scrape). L5 computer use runs in a container started by computer_use_orchestrator (image ghcr.io/zhixuli0406/duduclaw-computer-use:v<version>, never pulled automatically, actions via xdotool) and is driven by the agent through eight computer_* MCP tools, which the duduclaw mcp-server process forwards to gateway-owned sessions over a signed loopback route (POST /api/internal/computer-use, computer_use_sessions/; network only to the per-agent allowed_domains hosts, pinned at session start). The chat-triggered loop and the native host-desktop mode were removed. L3 headless is an optional per-agent Playwright/Browserbase MCP server (.mcp.json). Deny-by-default via CapabilitiesConfig (computer_use / browser_via_bash / allowed_tools / denied_tools). The dead browser_router.rs 5-layer router and its “L4 Sandbox Browser” tier were removed in 2026-09.
  • CJK-safe byte slicing: duduclaw_core::truncate_bytes / truncate_chars replaced 31 unsafe s[..s.len().min(N)] sites (fixed v1.8.11 multi-byte codepoint panics).

Accounts & Cost

Section titled “Accounts & Cost”
  • Per-agent model routing (SDK-first): agent.toml [model] — preferred (Claude SDK model), local.model, local.use_router, api_mode (cli/direct/auto), account_pool (see below).
  • Multi-OAuth account rotation: OAuth sessions (Claude Pro/Team/Max via claude auth status + CLAUDE_CODE_OAUTH_TOKEN for setup-token accounts) + API keys. 4 strategies (Priority/LeastCost/Failover/RoundRobin). Rate-limit cooldown (2min), billing-exhaustion cooldown (24h), budget enforcement, token expiry tracking (30d/7d warnings).
  • Per-agent account pool (agent.toml [model] account_pool): restricts which rotator accounts an agent may use. Applied to the candidate set — after the provider/health/cooldown/budget filters and before the strategy runs — so all four strategies keep their exact semantics over a narrower set. Entries match an account id or its dashboard label (exact, trimmed, ASCII-case-insensitive; never a substring). Fail-open: a pool matching no available account (stale ids, everything cooling down) logs a warn and falls back to the full set — a stale pool must never leave an agent with no account. Unset/empty ⇒ rotation unchanged. Entry points: AccountRotator::select_with_pool / select_for_provider_with_pool.
  • Dual dispatch path: both sub-agent dispatcher (claude_runner::call_with_rotation) and user-facing channel reply (channel_reply::call_claude_cli_rotated → rotate_cli_spawn_with_pool) go through the rotator, each threading the answering agent’s account_pool.
  • FailureReason classification — RateLimited / Billing / Timeout / BinaryMissing / SpawnError / EmptyResponse / NoAccounts / Unknown — with category-specific zh-TW user messages and channel_failures.jsonl audit records.
  • Binary discovery: which_claude() / which_claude_in_home() probe Homebrew (Intel + Apple Silicon), Bun, Volta, npm-global, .claude/bin, .local/bin, asdf shims, NVM version directories — fixes launchd-launched gateway binary discovery when PATH is empty.
  • CostTelemetry: SQLite-backed token usage tracking with cache efficiency analytics (cache_read / (input + cache_read + cache_creation)), 200K price cliff warning, adaptive routing (cache_eff <30% → local). MCP tools: cost_summary, cost_agents, cost_recent.
  • Per-model cost rollup (CostTelemetry::summary_by_model): token_usage has carried a model column since the first schema, but every rollup grouped by agent / user / day — “which model is the money going to?” was unanswerable. summary_by_model(agent_id: Option<&str>, since_unix) groups by model (costliest first; rows with no recorded model id bucket under "(unknown)", never guessed) and reports requests / input_tokens / output_tokens / cache_read_tokens / cache_creation_tokens / cost_millicents + cost_usd / cache_efficiency. Cost is the summed stored per-row cost_millicents — the same single pricing path (cost_for, applied once at record time) as every sibling rollup, never re-derived; cost_usd is a pure unit conversion of it. Surfaced additively: the MCP cost_summary and cost_agents responses gain a by_model array for the same window (cost_agents’ agent rows move under an agents key, since a top-level JSON array cannot carry a named sibling), and the dashboard RPC cost.by_model (params agent_id?, days? — default 7, clamped 1–365) returns the rollup under the same admin gate as its cost.* siblings. A rollup failure inside cost_summary/cost_agents degrades to an empty by_model rather than failing the call the caller actually made.
  • Direct API client (direct_api.rs): bypasses Claude CLI for pure chat, cache_control: ephemeral on system prompt (no measured hit rate is published). Singleton reqwest::Client with 120s timeout; used as fallback when all OAuth accounts are cooling.

Scheduling

Section titled “Scheduling”
  • HeartbeatScheduler: per-agent unified scheduling — bus polling + GVU silence breaker + cron, max_concurrent_runs semaphore.
  • CronScheduler: reads cron_tasks.jsonl (+ cron_tasks.db since v1.8.12), fires tasks on cron expression. list_cron_tasks returns all tasks (no longer filters by default_agent, v1.8.3). Schedules are created with tasks_create + schedule; the older schedule_task MCP tool was removed in v1.69.0.
  • ReminderScheduler: one-shot reminders (relative 5m/2h/1d or ISO 8601), direct static message or agent_callback wake-up mode.

Skill Ecosystem

Section titled “Skill Ecosystem”
  • 6-stage lifecycle: Activation → Compression (three-layer progressive loading) → Extraction → Distillation → Diagnosis → Gap Analysis. The former Reconstruction stage had no caller and was removed in 2026-09.
  • GitHub live indexing — Search API with 24h local cache, weighted search.
  • Skill auto-synthesis (Phase 3-4): gap accumulator detects repeated domain gaps → synthesizes skills from episodic memory (Voyager-inspired) → sandbox trial with TTL → cross-agent graduation; off by default (agent.toml [evolution] skill_synthesis_enabled). MCP tools: skill_security_scan, skill_graduate, skill_synthesis_status.
  • Rust-native Skill security scanner (skill_lifecycle::security_scanner) — no Python subprocess; backs the dashboard vet, the MCP skill_security_scan tool, and the sandbox-trial gate.

Task & Knowledge

Section titled “Task & Knowledge”
  • Task Board: SQLite-backed task management with status/priority/assignment tracking + real-time Activity Feed WebSocket. Dashboard RPC tasks.list/create/update/remove/assign, activity.list; agent MCP tools tasks_list, tasks_create, tasks_update, tasks_claim, tasks_complete, tasks_block, activity_list, activity_post. An AI employee changing a task that is not its own (not assignee, claimer or creator; for complete/block, not assignee or claimer) needs the delegation relationship with the assignee, and cannot add, remove or reorder the control tags outcome:…, grant:…, auto-research; on a goal task title, description and acceptance_criteria are frozen for AI employees.
  • Shared Knowledge Base: ~/.duduclaw/shared/wiki/ with Wiki target classification (agent/shared/both). MCP tools: wiki_ls/read/write/search/stats/lint with scope="shared" (the six shared_wiki_* spellings were removed in v1.69.0), plus shared_wiki_delete and wiki_share.
  • Autopilot rule engine: automated delegation/notifications/skill execution. Triggers a new rule may use (12): task_created, task_updated, task_status_changed, activity_new, channel_message, agent_idle, run_at_risk, os_file, os_frontmost, tick, security_event, odoo_event; cron_tick is never emitted and refused on create since v1.67.1 (23-autopilot-engine).

Integrations

Section titled “Integrations”
  • Odoo ERP bridge (duduclaw-odoo crate): JSON-RPC middleware supporting CE/EE, 17 MCP tools (CRM/Sales/Inventory/Accounting), EditionGate auto-detection, event polling + POST /webhook/odoo (both off by default) feeding odoo_event autopilot rules. Per-agent credential isolation via OdooConnectorPool (RFC-21 §2, v1.11.0). Dashboard test-before-save: odoo.test RPC accepts inline params (v1.13.1) — credential omitted falls back to stored secret; same SSRF/HTTPS/db-name validators as odoo.configure; scrub_odoo_error() caps connector errors at 240 chars to prevent HTML / URL leakage.
  • Prometheus metrics: GET /metrics on gateway HTTP — failover, wiki trust, decision continuity, prompt compression, resident sensing (tick_*), goal-loop and live-fork counters. The request/token/duration/session/channel/budget series were never incremented and were removed in v1.66.
  • RL trajectory collector: writes per-agent trajectories to ~/.duduclaw/rl_trajectories.jsonl during channel interactions. The duduclaw rl CLI that exported them was removed in 2026-09.
  • BroadcastLayer tracing layer streams real-time logs to WebSocket subscribers.
  • Dashboard WebSocket heartbeat: server Ping every 30s, close idle sockets after 60s without Pong. Client ping application-level RPC every 25s (browsers can’t issue control frames).

Reliability & Governance (v1.9.4)

Section titled “Reliability & Governance (v1.9.4)”
  • duduclaw-durability crate (🗑️ Removed in commit b0639b96 on 2026-07-04) — This crate contained a five-pillar durability framework (idempotency, retry, circuit_breaker, checkpoint, dlq). It was removed after determining it had zero call sites in the production codebase. The gateway’s LLM fallback chain relies on other mechanisms (see gateway/failover.rs). Historical references to checkpoint save/rewind/fork are not available in the current codebase.
  • duduclaw-governance crate (🗑️ removed in b0639b96) — its rate / permission / quota / lifecycle policies had no enforcer, and the dashboard Governance page and its RPCs were removed in v1.66. Rate limits, delegation policy + MCP scopes, and license quotas are enforced separately.
  • LLM fallback chain (gateway/failover.rs, module failover::model) — layer 2 of the three-layer failover stack (account → model → runtime, all under one module tree since 2026-09-29): primary timeout / 503 / 429 / overloaded auto-switches to the lighter fallback model, never on a billing error. is_llm_fallback_error / should_attempt_model_fallback are pure functions with unit tests; FailoverManager::model_fallback_for is the combined decision every dispatch path calls. UTF-8-safe truncation via char_indices.
  • Evolution Events system (gateway/evolution_events/) — 30+ event schema, async batch+retry emitter, query interface, reliability guarantees. HTTP endpoints exposed on gateway and surfaced in Web ReliabilityPage.

Memory Evaluation (v1.9.4 / W21)

Section titled “Memory Evaluation (v1.9.4 / W21)”
  • LOCOMO evaluation (python/duduclaw/memory_eval/) — retrieval_accuracy, retention_rate, locomo_integrity_check. cron_runner is a manual CLI entry point (python -m memory_eval.cron_runner smoke_test|weekly_kpis|monthly_locomo); nothing in the repo schedules it. 5-minute smoke_test P0 verifies basic memory functions. build_golden_qa.py builds the gold QA set; data/golden_qa_set.jsonl carries the first 200 entries. duduclaw-memory engine adds batch query API for evaluation.
  • Python agents/ + mcp/ modules — agents/capabilities/ (manifest + matcher), agents/routing/ (router + resolution + memory_resolver). mcp/auth/ (API Key with key masking), mcp/tools/memory/ (store / read / search / namespace / quota with strict scope enforcement at execute() entry — patches a v1.9.3 auth gap where any valid API key bypassed scope limits).

Web Dashboard

Section titled “Web Dashboard”
  • Tech stack: React 19 + TypeScript + Tailwind CSS 4 + Base UI + the shared mds component library.
  • Real-time log streaming (BroadcastLayer → WebSocket).
  • OrgChart (D3.js interactive agent hierarchy).
  • Memory page with Key Insights tab (key_facts cards with access_count badges) + Self-Improvement tab (stagnation warnings, rejection statistics, playbook rule cards).
  • Logs page with source filter chips + severity dropdown + severity-colored left borders + JSON detail expansion.
  • Toast notification system (module-scoped event bus, max-5 queue, warm variants).
  • Skill Market 3-tab (Marketplace / Shared Skills / My Skills).
  • Autopilot settings + Session Replay + WikiGraph.
  • Reliability page — evolution-event query and per-agent reliability summary (audit.evolutionQuery, audit.reliabilitySummary).
  • i18n: zh-TW / en / ja-JP (600+ translation keys).
  • Dark/Light theme (system + manual toggle).