agenticEvolve

A self-evolving AI agent — grows itself daily.

简体中文 | 繁體中文 | 日本語

---

Persistent agent runtime built on claude -p with a Python asyncio gateway. 6-layer memory + cross-layer auto-recall. Closed-loop skill synthesis. Voice I/O. Browser automation. Built-in cron. 2-layer security. Multi-platform (Telegram + Discord + WhatsApp). Interactive CLI REPL with Rich TUI. 39 Telegram commands + 32 CLI commands — your entire dev machine in your pocket.

---

What can you do with it?

Browse the web for you

"Go to the Anthropic docs and find the latest Claude model pricing." The agent opens ABP browser, navigates, extracts the data, and sends you a clean summary. If Cloudflare blocks it, it auto-switches to Brave.

Serve your WhatsApp and Discord groups

/serve → WhatsApp → Groups → toggle your dev group on. Now anyone in that group can talk to your AI agent. It responds to every message, maintains per-group conversation memory, and you control it all from Telegram inline keyboards. Works with Discord channels too — the agent hooks into your desktop app via Chrome DevTools Protocol.

Subscribe to channels and get digests

/subscribe → Discord → pick your favorite channels. Next morning, run /discord and get a clean summary of everything you missed — key discussions, links shared, tools mentioned, action items. Works across Discord channels, WhatsApp groups, and WeChat groups. Never scroll through 500 unread messages again.

Search your own WeChat history

WeChat's built-in search is terrible. The agent reads your local WeChat databases and gives you a searchable export — contacts, messages, groups, favorites. All offline, all on your machine.

Absorb ideas from your group chats overnight

Your /evolve cron at 6 AM doesn't just scan GitHub. It also reads your WeChat tech group chats, summarizes the last 24 hours of discussions — new tools people mentioned, repos shared, techniques debated — and absorbs the best ideas into skills. You wake up with your group's collective intelligence baked in.

Brainstorm business ideas from trending signals

/produce — the agent aggregates today's signals from 11 sources (GitHub Trending, Hacker News, X/Twitter, Reddit, Product Hunt, Lobste.rs, ArXiv, HuggingFace, BestOfJS, WeChat groups, and your starred repos), identifies emerging trends, and brainstorms 5 concrete app/business ideas with revenue models, tech stacks, and MVP scopes. Signal-driven ideation on demand.

Self-improving UX

Every night at 1 AM, the agent reads the day's conversations, finds friction points where you waited too long or got confusing responses, and patches its own code to fix them. You wake up to a better agent.

Ship code from your phone

You're on the subway. You text /do add rate limiting to the API. The agent reads your codebase, writes the middleware, runs the tests, and pushes to git. You get a summary back before your stop.

Absorb any repo in one message

You see a cool repo on Twitter. Screenshot it to the bot. The agent OCRs the image, finds the GitHub URL, clones the repo, maps its architecture, extracts the patterns that matter to your stack, and installs them as skills — all from one photo.

Wake up to new skills you didn't write

The daily /evolve cron fires at 6 AM. By the time you check Telegram, the agent has scanned GitHub trending, found a new testing framework, built a skill for it, passed it through 2-layer security, auto-installed it, and pushed to your repo. You just got smarter overnight.

Talk to your codebase in any language

Send a voice message in English, Cantonese, Mandarin, Japanese, Korean, or any of 40+ supported languages. The agent transcribes locally via whisper.cpp (~500ms), auto-detects the language, responds in text, and reads the answer back to you in the same language via edge-tts.

Deep-dive anything with /learn

/learn https://github.com/some/repo — the agent clones it, reads every file, maps the architecture, evaluates how it could benefit your workflow, gives an ADOPT / ADAPT / SKIP verdict, and optionally builds a skill from it.

---

Capabilities

Capability	Description
CLI REPL	ae — Interactive Rich TUI with streaming output, markdown rendering, tool use spinners, 32 commands with Tab autocomplete, session persistence, auto-recall. No gateway required
Multi-Platform	Telegram (bot API) + Discord (local cache only) + WhatsApp (Baileys v7 bridge). /subscribe to monitor channels for digests, /serve to make the agent respond in any group or DM. WhatsApp supports images, PDFs, text files, and voice messages
Build	Full Claude Code over Telegram or CLI — terminal, file I/O, web search, MCP, 26 skills
Evolve	5-stage pipeline: COLLECT → ANALYZE → BUILD → REVIEW → AUTO-INSTALL. Scans 12 sources: GitHub Trending + HN + X/Twitter + Reddit + Product Hunt + Lobste.rs + ArXiv + HuggingFace + BestOfJS + WeChat groups + Discord cache, synthesizes skills
Absorb	/absorb <url> — clones repo, maps architecture, diffs patterns, implements improvements into your system
Learn	/learn <target> — deep-dive extraction with ADOPT / ADAPT / SKIP verdicts
Voice	Send voice messages → local whisper.cpp transcription (~500ms). /speak → edge-tts with 300+ voices. Auto-detects Cantonese/Mandarin/Japanese/Korean
Browser	ABP (Agent Browser Protocol) for default browsing. Auto-switches to Brave/Chrome via CDP when Cloudflare blocks. Isolated agent profiles
Auto-Recall	unified_search() across 6 memory layers before every response (~400 tokens/msg)
Cron	/loop every 6h /evolve — autonomous growth on a schedule
Security	L1: regex scanner pre-install (reverse shells, credential theft, crypto miners). L2: AgentShield post-install (1282 tests, 102 rules). Auto-rollback on critical findings
Hooks	Typed async event system — message_received, before_invoke, llm_output, tool_call, session_start, session_end
Semantic Recall	TF-IDF cosine similarity search layer augments FTS5 keyword search. 5000-feature vectorizer with bigrams. Corpus rebuilt from sessions, learnings, instincts, memory files. Cached at ~/.agenticEvolve/cache/
Instinct Engine	Behavioural pattern observations scored and routed to instincts table. High-confidence instincts (0.8+ across 2+ projects or 5+ sightings) auto-promote to MEMORY.md
Resilience	Drain-on-shutdown (30s wait for in-flight requests). Typed failure classification (auth/billing/rate-limit). 3-pass context compaction. Hot config reload. Loop detection (warn@3 identical turns, terminate@5). Memory queue read-through (debounced atomic writes, no stale reads). Parallel BUILD stage (ThreadPoolExecutor, 3 isolated workspaces)
Testing	423 automated tests (423 pass, 1 xfail). Covers: 81 command handler integration tests (all 35+ handlers), session DB, FTS5 search, security scanner, signal dedup, semantic search, instinct promotion, cron parser, cost cap, loop detector, context compaction, flag parsing

---

Setup

1. Install

curl -fsSL https://raw.githubusercontent.com/outsmartchad/agenticEvolve/main/scripts/install.sh | bash

The installer handles everything — cloning, dependencies, PATH, and runs the interactive setup wizard. No prerequisites except Python 3 and git.

After installation:

source ~/.zshrc    # reload shell (or: source ~/.bashrc)

Requires: Claude Code (npm install -g @anthropic-ai/claude-code) — the installer checks for it and shows the install command if missing.

2. Interactive Chat (CLI REPL)

ae

That's it. You get a Rich TUI with streaming output, markdown rendering, tool use spinners, and all 32 commands with Tab autocomplete. Resume a previous session with ae --resume <session_id>.

3. Start the Gateway (for Telegram/Discord/WhatsApp)

ae gateway start

Message your bot on Telegram, or let the agent serve your WhatsApp groups and Discord channels.

Useful Commands

Command	What it does
ae	Interactive chat REPL (default)
ae --resume ID	Resume a previous session
ae setup	Re-run the setup wizard
ae doctor	Diagnose issues
ae gateway start	Start the messaging gateway
ae gateway stop	Stop the gateway
ae gateway install	Install as launchd service (auto-start on login)
ae status	System overview
ae cost	Usage and spend

Voice Support (optional)

brew install whisper-cpp ffmpeg
curl -L -o ~/.agenticEvolve/models/ggml-small.bin \
  https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.bin

---

Commands

All commands work in both the CLI REPL (ae) and Telegram. The CLI REPL supports Tab autocomplete with descriptions for all commands.

Core

Command	Function
(any message)	Chat with Claude Code
(voice message)	Auto-transcribe (whisper.cpp) + reply (Telegram only)

Pipelines (LLM-backed)

Command	Function
/evolve [--dry-run]	Scan signals, build and auto-install skills
/absorb <url>	Absorb patterns from any repo
/learn <target>	Deep-dive with ADOPT/STEAL/SKIP verdicts
/produce [--ideas N]	Brainstorm business ideas from all signals
/reflect [--days N]	Self-analysis: patterns, avoidance, next actions
/digest [--days N]	Morning briefing (sessions, signals, cost)
/gc [--dry-run]	Garbage collection (stale sessions, orphans)
/wechat [--hours N]	WeChat group chat digest (reads local DBs)
/discord [--hours N]	Discord channel digest (from stored messages)
/whatsapp [--hours N]	WhatsApp group digest (from stored messages)

Info

Command	Function
/recall <query>	Cross-layer search (all 6 memory layers)
/search <query>	FTS5 search past sessions
/memory	View MEMORY.md + USER.md
/soul	View SOUL.md personality
/config	View config.yaml settings
/skills	List installed skills (26)
/learnings [query]	List or search past learnings
/sessions [N]	List recent sessions
/cost	Usage and spend
/status	System overview
/heartbeat	Quick health check

Cron

Command	Function
/loop <interval> <prompt>	Schedule recurring execution (e.g. /loop 6h /evolve)
/loops	List active cron jobs
/unloop <id>	Remove a cron job
/pause <id|--all>	Pause a cron job
/unpause <id|--all>	Resume a cron job
/notify <delay> <msg>	One-shot delayed notification

Admin

Command	Function
/model [name]	Show or switch model
/autonomy [level]	Show or set autonomy level (full/supervised/locked)
/new	Start a new session
/queue	Show skills pending approval
/approve <name>	Approve a queued skill
/reject <name>	Reject a queued skill

Telegram-only

Command	Function
/speak <text>	Text-to-speech (edge-tts, auto-detects language)
/do <instruction>	Natural language → structured command
/subscribe	Select channels to monitor for digests
/serve	Select channels/contacts where the agent responds
/lang [code]	Set persistent output language
/restart	Restart gateway remotely

All commands →

---

Architecture

User (CLI REPL / Telegram / Discord / WhatsApp / Voice)
  → Gateway (asyncio) or CLI (standalone) → Hook Dispatcher → Session + Cost Gate
  → Auto-Recall (6 layers) → claude -p → SQLite → Git Sync

No custom agent loop. Claude Code is the runtime — 25+ built-in tools, MCP servers, skills. The gateway adds memory, routing, recall, cron, voice, browser, multi-platform, and safety around it. The CLI REPL (ae) bypasses the gateway entirely, calling claude -p directly with the same memory, recall, and session infrastructure.

Key Design Decisions

• No tool/toolset system — Claude Code already has tools. We build skills and infrastructure, not abstractions.
• Bounded memory — MEMORY.md (2200 chars) + USER.md (1375 chars) + SQLite FTS5. No unbounded growth.
• Closed-loop — auto_approve_skills: true. Evolve → build → review → install → sync to git. No human gate.
• Drain-on-shutdown — In-flight requests complete before restart. No lost work.
• Modular commands — 39 Telegram commands split into 9 mixins (admin, pipelines, signals, cron, approval, search, media, misc, subscribe). 32 commands re-implemented for CLI REPL. Adapter core is 630 lines.
• Dual-layer recall — FTS5 keyword search + TF-IDF semantic search. Auto-recall injects relevant context before every Claude invocation.
• Instinct pipeline — Behavioural patterns observed across sessions are scored, deduplicated, and auto-promoted to MEMORY.md when confidence is high enough.

---

Voice Pipeline

Direction	Technology	Latency	Cost
Voice → Text	Local whisper.cpp (ggml-small multilingual)	~500ms on Apple Silicon	Free
Text → Voice	edge-tts (300+ neural voices)	~1s	Free
Language detection	CJK heuristic (嘅係唔 → Cantonese, ひらがな → Japanese)	Instant	Free

Auto-TTS modes: off (only /speak), always (every reply), inbound (reply with voice when user sends voice).

---

Browser Automation

Browser	When	How
ABP (default)	All agent browsing	Bundled Chromium, JS freeze between actions, 90.5% Mind2Web
Brave	User asks / Cloudflare blocks ABP	CDP on port 9222, isolated profile
Chrome	User asks / Cloudflare blocks ABP	CDP on port 9223, isolated profile

Agent profiles are sandboxed at ~/.agenticEvolve/browser-profiles/ — never touches user's real browser data.

---

Security

Layer	Tool	When	On Critical
L1	gateway/security.py	Pre-install: scans raw files	Block + abort pipeline
L2	AgentShield (1282 tests)	Post-install: scans ~/.claude/ config	Auto-rollback installed skills

Scans for: credential exfiltration, reverse shells, obfuscated payloads, crypto miners, macOS persistence, prompt injection, npm hook exploits.

---

Scheduled Cron Jobs

4 autonomous jobs run daily — no human trigger needed.

Job	Schedule (HKT)	What it does
evolve-daily	6:00 AM	Collects signals from 11 sources: GitHub Trending + HN + X/Twitter + Reddit + Product Hunt + Lobste.rs + ArXiv + HuggingFace + BestOfJS + WeChat groups, scores candidates, builds up to 3 new skills, security-reviews, auto-installs, pushes to git
daily-digest	8:00 AM	Morning briefing — top signals, skills built, session count, cost summary. Delivered to Telegram
wechat-digest	9:00 AM	Daily WeChat group chat digest — summarizes discussions, tools mentioned, key insights from tech groups. Delivered to Telegram
daily-ux-review	1:00 AM	Reads the day's conversations, finds friction points, identifies top 3 UX improvements, implements them directly

Managed via /loop, /loops, /unloop, /pause, /unpause. Config in cron/jobs.json.

---

Signal Sources (12)

Source	Collector	API	What it captures
GitHub Search	github.sh	GitHub API (gh CLI)	Trending repos by keyword, starred repo activity, release watch
GitHub Trending	github-trending.py	GitHub API (gh CLI)	Hot new repos created in the last 7 days
Hacker News	hackernews.sh	Algolia API	Keyword search + front page + Show HN
X / Twitter	x-search.sh	Brave Search API	Viral tweets about open source, dev tools, AI
Reddit	reddit.py	Pullpush.io API	13 subreddits: LocalLLaMA, programming, ClaudeAI, etc.
Product Hunt	producthunt.py	RSS/Atom feed	Dev tools + AI product launches
Lobste.rs	lobsters.py	JSON API	Curated tech news (high signal-to-noise)
ArXiv	arxiv.py	ArXiv API	Papers from cs.AI, cs.CL, cs.SE, cs.LG
HuggingFace	huggingface.py	HF API	Trending models and spaces
BestOfJS	bestofjs.py	Static JSON API	Trending JavaScript/TypeScript projects by daily star growth
WeChat	wechat.py	Local DB	Group chat messages (reads local data)
Discord	discord.py	Chromium cache	Cached API responses from desktop app (zero network calls)

---

Skills (26 installed)

Skill	Purpose
agent-browser-protocol	ABP browser automation via MCP
browser-switch	Multi-browser CDP switching (Brave/Chrome)
brave-search	Web search via Brave API
firecrawl	Web scraping, crawling, search, structured extraction
cloudflare-crawl	Free web crawling via Cloudflare Browser Rendering API
jshook-messenger	Discord/WeChat/Telegram/Slack interception via jshookmcp MCP
wechat-decrypt	Read local WeChat databases and export messages, contacts, groups on macOS
session-search	FTS5 session history search
cron-manager	Cron job management
skill-creator	Official Anthropic skill creation
deep-research	Multi-source research pipeline
market-research	Market/competitor analysis
article-writing	Long-form content creation
video-editing	FFmpeg video editing guide
security-review	Code security checklist
security-scan	AgentShield config scanner
autonomous-loops	Self-directed agent loops
continuous-learning-v2	Pattern extraction pipeline
eval-harness	Skill evaluation framework
claude-agent-sdk-v0.2.74	Claude Agent SDK patterns
nah	Quick rejection/undo
unf	Unfold/expand compressed content
next-ai-draw-io	Architecture diagrams from natural language
mcp-elicitation	Intercept mid-task MCP dialogs for unattended pipelines
skill-gap-scan	Diff local skills against community catalog, surface adoption gaps
context-optimizer	Auto-compact stale memory files based on /context hints

---

Recent Changes

v2.7 — IronClaw Adoption (Phases 1-6)

Phase 1: Smart Model Routing

• 13-dimension regex complexity scorer (code patterns, conversation depth, multi-step reasoning, etc.)
• Automatic Sonnet/Opus routing per message — saves $10-20/day on API costs
• Cascade detection: re-invoke with reasoning model if Sonnet shows uncertainty

Phase 2: Provider Chain

• Retry → CircuitBreaker → Cache decorator pattern (IronClaw's provider chain architecture)
• Automatic retries with exponential backoff
• Circuit breaker prevents cascade failures
• Response cache for repeated queries

Phase 3: Security Hardening

• credential_guard.py: LeakDetector scans .env secrets (raw, base64, URL-encoded)
• Two-layer output redaction (credential_guard + redact.py)
• Content sanitizer wired for all platforms (was only WhatsApp before)
• Sandbox deny patterns injected into prompt

Phase 4: Availability Upgrades

• Parallel WhatsApp message handling (was serial, now Semaphore(5))
• Event bus (pub/sub) with default triggers (cost alerts, error streaks, reconnect)
• Heartbeat health monitoring with auto-disable notifications
• All 19 hooks now wired (5 were previously dead)

Phase 5: Memory Upgrades

• Vector embeddings via sentence-transformers (all-MiniLM-L6-v2, local, no API calls)
• Hybrid search: FTS5 + embedding + RRF fusion ranking
• LLM summarization for context compaction (replaces truncation)
• Memory consolidation: auto-prune MEMORY.md via Sonnet when over limit
• Memory dashboard page with search, stats, embedding status

Phase 6: Self-Expanding Enhancement

• SubagentOrchestrator hooks into evolve BUILD stage (observability)
• skill_metrics table: track usage, ratings, stale skills
• Background /learn: non-blocking execution via BackgroundTaskManager

Stats: 822 tests passing (up from ~700). 10 new modules. Estimated savings: $10-20/day from smart routing.

v2.6 — Security, Observability & Platform Parity

• Content Sanitizer: Prompt injection defense with randomized boundary markers, Unicode homoglyph folding (adapted from OpenClaw)
• Log Redaction: 17 regex patterns auto-strip API keys, tokens, PEM blocks from all log output
• Retry Utility: Exponential backoff + jitter for transient failures, Telegram retry helper
• Rolling Logs: RotatingFileHandler (50MB, 5 backups) replaces unbounded log files
• Tool Loop Detector: 4-mode detection (generic repeat, poll no-progress, ping-pong, global circuit breaker) prevents runaway sessions
• Security Self-Audit: /doctor command checks env permissions, config secrets, cost caps, dependencies, sandbox health, DB integrity
• Diagnostic Event Bus: Typed events (message, usage, session, loop, heartbeat) with JSONL sink and status summary
• WhatsApp Commands: /cost, /status, /doctor, /help now work on WhatsApp
• Voice Pipeline Fixes: Long audio chunking (48-min → 10-min chunks), OGG→WAV conversion, triple-layer dedup, force reply for voice, 2-part response (transcript + summary)
• Security Fixes: Env sanitization in CLI/TUI, dynamic owner paths, content wrapping for served groups

v2.5 — Security + Intelligence + Plugin System

Phase 1: Security + Cost Protection

• Environment variable sanitization — strips 30+ secret patterns from claude -p to prevent prompt leaking credentials.
• Per-user sliding-window rate limiting (5/min, 30/hr configurable).
• [NO_REPLY] token — agent can skip irrelevant group messages without responding.
• Message debouncing — batches rapid messages (2.5s window, 8s max wait) for served channels.
• @agent <prompt> trigger — works for anyone in groups + DMs, with reply-to context and browser MCP.
• Docker sandbox for served chats — isolated Python execution (--network=none, --cap-drop=ALL, 512MB memory).

Phase 2: UX + Efficiency

• Telegram streaming — edit-in-place with 1.5s throttled edits and "..." placeholder.
• Context window management — token estimation, auto-compaction at 60%/85% thresholds.
• Identity linking — /link and /whoami commands for cross-platform user resolution.

Phase 3: Intelligence

• 19-hook plugin system with priority ordering, merge functions, and has_hooks() O(1) check. Hook points: message_received, before_invoke, llm_output, before_model_resolve, before_tool_call, after_tool_call, session_start, session_end, gateway_start, gateway_stop, background_task_*, before_pipeline_stage, after_pipeline_stage, subagent_spawned, subagent_ended, message_sending, message_sent.
• Plugin loader — discovers and loads from ~/.agenticEvolve/plugins/. Each plugin exports register(hooks, config).
• BackgroundTaskManager — detached long-running tasks with progress tracking. /tasks and /cancel commands.
• SubagentOrchestrator — generalized multi-Claude execution: run_parallel, run_pipeline, run_dag (dependency graph).

Phase 4: Polish

• Gateway exec mode — host execution with 3-tier security (deny/allowlist/full) and configurable approval (off/on-miss/always). 60+ safe bins auto-approved, 13 denylist patterns block dangerous commands (rm -rf, curl|sh, fork bombs, base64 eval).
• Config validation — semantic checks before applying reloaded config. /reload command with validation + config_reload hook.
• /allowlist command to manage exec allowlist (add/rm/clear/list).
• /hooks command to inspect registered hook listeners.
• 556 tests total (all passing).

v2.4 — TUI Commands + Discord Local Cache + WhatsApp File Support

WhatsApp File & Document Support

• Users can now send PDF, TXT, CSV, JSON, and other text-based files via WhatsApp. The bridge downloads the file via Baileys downloadMediaMessage, saves to /tmp/agenticEvolve-wa-files/, and passes the path to Claude Code's Read tool for analysis.
• Also supports audio/voice messages (saved to /tmp/agenticEvolve-wa-audio/).
• Supported file types: .txt, .md, .csv, .json, .xml, .yaml, .html, .pdf, .py, .js, .ts, .java, .c, .cpp, .go, .rs, .rb, .sh, .sql, .log, and any plaintext format. Images: .png, .jpg, .jpeg, .gif, .webp.
• Files and images auto-escalate to opus model for better analysis.

WhatsApp Self-Reply Fix

• Fixed bot replying to its own messages in DMs and served groups (was causing infinite response loops, 18+ replies to one message).
• Bridge.js now checks sender JID against sock.user.id — fromMe flag is unreliable on Baileys linked devices.
• Python adds 3-second cooldown per chat after sending a reply, plus message ID deduplication.

Discord Fully Disabled (Zero Network Calls)

• Discord adapter start() is now a complete no-op — no CDP connections, no REST API calls, no token extraction. Account received second community guidelines warning.
• Discord data is read ONLY from local Chromium disk cache (tools/discord-local/read_cache.py). Zero network calls to Discord servers.

TUI Slash Commands

• 6 new commands in the Textual TUI: /lang, /do, /restart, /speak, /subscribe, /serve. Previously Telegram-only, now fully functional in the CLI.
• Browsable /subscribe and /serve modal with click-to-toggle targets. Discovers channels from DB subscriptions, platform_messages history, WhatsApp auth files, and Discord Chromium cache.
• /speak uses edge-tts CLI directly (avoids asyncio conflicts in Textual worker threads).
• Ctrl+C copies last assistant response to clipboard via pbcopy.
• HelpScreen uses VerticalScroll with reordered categories.

Discord Local Cache Reader

• Built tools/discord-local/read_cache.py — reads Discord messages from the desktop app's Chromium HTTP response cache. Parses gzip-compressed JSON from ~/Library/Application Support/discord/Cache/Cache_Data/. Zero network calls.
• 5,835 messages across 353 channels recovered from cache, dating back to 2022.
• Added collectors/discord.py signal collector for the /evolve pipeline, matching the WeChat collector pattern.
• Discord is now a 12th signal source for /evolve.

/lang Works Across All Platforms

• Language preference injection moved from TUI-only to build_system_prompt() in agent.py. Now works for Telegram, CLI, and any future platform.
• user_id parameter threaded through invoke_claude() → build_system_prompt() → language lookup from user_prefs DB.

Bug Fixes

• Fixed /scan-skills → /scanskills — hyphens are invalid in Telegram bot command names, was crashing Telegram adapter startup entirely.
• Fixed Discord poll loop spam — 401 errors now trigger _auth_failed flag and stop the poll instead of retrying infinitely.

v2.3 — CLI REPL + WhatsApp LID Resolution

Interactive CLI REPL (ae)

• ae launches a standalone Rich TUI REPL — no gateway process required. Streaming output with markdown rendering, tool use spinners, and cost tracking.
• 32 slash commands with Tab autocomplete and descriptions. All pipeline commands (/produce, /evolve, /learn, /absorb, /reflect, /digest, /gc), info commands (/memory, /soul, /config, /skills, /learnings, /recall, /search, /sessions), cron management (/loop, /loops, /unloop, /pause, /unpause, /notify), and admin (/model, /autonomy, /queue, /approve, /reject).
• Session persistence — all messages saved to SQLite with auto-titles. Resume previous sessions with ae --resume <session_id>.
• Auto-recall from all 6 memory layers before every invocation. Cost cap enforcement.
• prompt-toolkit input with file-backed history and auto-suggest.

WhatsApp LID JID Resolution

• Fixed a critical bug where Baileys v7 delivers DM messages under LID JIDs (@lid) instead of phone JIDs (@s.whatsapp.net). The bridge now loads lid-mapping-*_reverse.json files on startup and resolves LID→phone for incoming messages, outbound sends, and history sync.
• Previously, served contacts like 85254083858 were silently dropped because Python couldn't match the LID JID against phone-based serve targets.

WhatsApp Serve for DM Contacts

• /serve now supports individual WhatsApp contacts (not just groups). Added _serve_contacts set alongside _serve_groups. DM routing updated to bypass allowed_users for served contacts.

WhatsApp Media Support

• Incoming WhatsApp images, documents (PDF, TXT, CSV, etc.), and audio messages are downloaded via Baileys downloadMediaMessage, saved to /tmp/, and passed to Claude Code for analysis. Messages with media auto-escalate to opus model.

Auto-Model Escalation

• Messages containing math, coding, or logic questions are auto-detected via regex and routed to serve_reasoning_model (opus) instead of the default serve_model (sonnet). Image and file messages also trigger escalation.

Channel-Specific Knowledge

• _CHANNEL_KNOWLEDGE dict in run.py maps channel/group IDs to expert knowledge prompts. Injected after personality prompt for both Discord and WhatsApp served channels. Used for DAMM v2 expertise in degen-damm Discord channel.

v2.2 — Multi-Platform + Subscribe/Serve

Multi-Platform Support

• Discord desktop adapter (gateway/platforms/discord_client.py) — hooks into the running Discord desktop app via Chrome DevTools Protocol (CDP). Extracts auth token from network requests, then uses Discord REST API for messaging. Supports guild listing, channel listing (with category grouping), DM channels, and message polling.
• WhatsApp bridge (whatsapp-bridge/bridge.js) — Baileys v7 Node.js bridge communicating via JSON over stdin/stdout. QR code delivery to Telegram for easy linking. LID-to-phone resolution for outbound messages. Group prefix filtering (/ask, @agent). Contact discovery from auth store lid-mapping files + live message tracking.
• WeChat — read-only access via decrypted local SQLCipher databases. Groups and contacts from contact.db, messages from message_0.db.

Subscribe & Serve Commands

• /subscribe — Telegram inline keyboard UI to select Discord channels, WhatsApp groups/contacts, or WeChat groups to monitor for digests. Paginated lists (40 per page) with category headers for Discord. WhatsApp split into Groups/Contacts sub-views.
• /serve — Same UI to select where the agent actively responds. WhatsApp served groups accept all messages (no prefix required, no allowed_users gate). Serve targets loaded from DB on gateway startup. Dynamic adapter updates when toggling.
• Subscriptions DB — subscriptions table in session_db with user_id, platform, target_id, target_name, target_type, mode. CRUD functions: add_subscription, remove_subscription, get_subscriptions, get_serve_targets, is_subscribed.
• Short ID registry — Telegram limits callback_data to 64 bytes. Long WhatsApp JIDs (120363427198529523@g.us) and WeChat chatroom IDs would exceed this. Solution: in-memory numeric ID map (sub:t:3 instead of sub:toggle:whatsapp:group:120363...).

v2.1 — Modular Architecture + Semantic Recall + Test Harness

Architecture

• Split telegram.py from 3870 lines into 8 command mixins (gateway/commands/): admin, pipelines, signals, cron, approval, search, media, misc. Adapter core reduced to 630 lines.
• Per-user language preferences (/lang) persisted to SQLite user_prefs table.
• Cross-source signal deduplication in evolve pipeline (URL + title matching).
• Collector retry with 5-second exponential backoff.

Semantic Search + Instinct Engine

• TF-IDF cosine similarity search layer augments FTS5 keyword search. unified_search() now queries both layers.
• Instinct auto-promotion: high-confidence behavioural patterns (confidence >= 0.8, seen across 2+ projects or 5+ times) auto-promote to MEMORY.md on session cleanup.
• Semantic corpus rebuilt from sessions, learnings, instincts, and memory files. Cached as pickle for fast reload.

Test Harness — 379 tests (379 pass, 1 xfail)

Test File	Tests	Coverage
test_commands.py	81	All 35+ command handlers: admin, pipelines, signals, cron, approval, search, media, misc + 30 authorization denial tests
test_session_db.py	25	Sessions, messages, FTS5 search, learnings, user prefs, instincts, stats
test_security.py	28	Critical patterns (reverse shell, fork bomb, miners), warnings, prompt injection, safe content, directory scan
test_evolve.py	72	Signal loading, ranking, URL/title dedup, edge cases, collectors, skill approval/rejection, hash verification, queue, reporting
test_agent.py	27	Stderr classification, history compaction (3-pass cascade), title generation, loop detector
test_semantic.py	11	Corpus build (sessions, learnings, instincts), search relevance, caching, score filtering
test_instincts.py	8	Context bug regression, auto-promotion (promote, dedup, char limit)
test_gateway.py	10	Cron parser (every/specific/step/wrap), session key, cost cap
test_voice.py	57	TTS config, language detection, audio format conversion, STT transcription, TTS directives
test_absorb.py	49	Constructor, reporting, scan prompts, security prescan, wechat hours parsing, dry run, AgentShield
test_telegram.py	13	Flag parsing (bool/value/alias/cast), user allowlist

Bug Fixes

• Fixed fork bomb regex in gateway/security.py — unescaped (){} metacharacters caused the pattern to never match.
• Fixed upsert_instinct in gateway/session_db.py — SELECT query was missing the context column, causing IndexError on repeat upserts with empty context.
• Fixed _handle_newsession in gateway/commands/admin.py — set_session_title (nonexistent function) → set_title, and missing generate_session_id() caused UNIQUE constraint violations.
• Fixed _extract_urls in gateway/commands/misc.py — _URL_RE class attribute was never defined, causing AttributeError on every plain text message.
• Fixed /restart spawning duplicate gateway instances — now uses os.getpid() to kill only current process.
• Fixed _extract_urls crash — _URL_RE class attribute was never defined on MiscMixin.

---

Documentation

Doc	Description
Interface	Usage examples and interaction patterns
Memory	6-layer memory architecture, auto-recall, instinct scoring
Commands	All 35 commands with flags and examples
Pipelines	Evolve, absorb, learn, do, gc pipelines
Skills	Full skill catalog
Security	Scanner, autonomy levels, safety gates
Architecture	Message flow, project structure, design decisions
Roadmap	Integration plan — Firecrawl, vision, sandboxing

---

Lineage

Project	Patterns Adopted
Claude Code	Agent runtime — 25+ tools, MCP, skills, subagents
hermes-agent	Bounded memory, session persistence, messaging gateway, growing status messages
ZeroClaw	Autonomy levels, deny-by-default, hot config reload, risk-tier classification
everything-claude-code	9 skills adapted, AgentShield security, eval-driven development, hook profiles
openclaw	Voice pipeline (TTS/STT), browser automation patterns, auto-TTS modes
ABP	Browser MCP — freeze-between-actions Chromium, 90.5% Mind2Web
deer-flow	Parallel subagent BUILD stage, isolated workspaces per candidate, loop detection, memory queue debounced writes

---

MIT