Уроборос

Самосоздающийся агент. Работает в Google Colab, общается через Telegram, хранит код в GitHub, память — на Google Drive.

Версия: 4.16.0

---

Быстрый старт

1. В Colab добавь Secrets:

   • OPENROUTER_API_KEY (обязательно)
   • TELEGRAM_BOT_TOKEN (обязательно)
   • TOTAL_BUDGET (обязательно, в USD)
   • GITHUB_TOKEN (обязательно)
   • OPENAI_API_KEY (опционально — для web_search)
   • ANTHROPIC_API_KEY (опционально — для claude_code_edit)

2. Опционально добавь config-ячейку (модели, воркеры, диагностика):

import os
CFG = {
    "GITHUB_USER": "razzant",
    "GITHUB_REPO": "ouroboros",
    "OUROBOROS_MODEL": "anthropic/claude-sonnet-4",
    "OUROBOROS_MODEL_CODE": "anthropic/claude-sonnet-4",
    "OUROBOROS_MODEL_LIGHT": "anthropic/claude-sonnet-4",
    "OUROBOROS_MODEL_BG": "deepseek/deepseek-chat-v3-0324",  # background consciousness — cheap model
    "OUROBOROS_MAX_WORKERS": "5",
    "OUROBOROS_WORKER_START_METHOD": "fork",   # Colab-safe default
    "OUROBOROS_DIAG_HEARTBEAT_SEC": "30",      # periodic main_loop_heartbeat in supervisor.jsonl
    "OUROBOROS_DIAG_SLOW_CYCLE_SEC": "20",     # warns when one loop iteration is too slow
    "OUROBOROS_BG_BUDGET_PCT": "10",           # max % of budget for background consciousness
}
for k, v in CFG.items():
    os.environ[k] = str(v)

Без этой ячейки используются дефолты: openai/gpt-5.2 / openai/gpt-5.2-codex. Background consciousness использует deepseek/deepseek-chat-v3-0324 (cheap: $0.19/$0.87 per MTok) по умолчанию. Для диагностики зависаний смотри main_loop_heartbeat, main_loop_slow_cycle, worker_dead_detected, worker_crash в /content/drive/MyDrive/Ouroboros/logs/supervisor.jsonl.

3. Запусти boot shim (см. colab_bootstrap_shim.py).
4. Напиши боту в Telegram. Первый написавший — создатель.

Архитектура

Telegram → colab_launcher.py (entry point)
               ↓
           supervisor/            (process management)
             state.py             — state, budget
             telegram.py          — TG client, formatting
             queue.py             — task queue, scheduling
             workers.py           — worker lifecycle, auto-resume
             git_ops.py           — git checkout, sync, rescue
             events.py            — event dispatch table
               ↓
           ouroboros/              (agent core)
             agent.py             — thin orchestrator
             consciousness.py     — background thinking loop
             context.py           — LLM context builder, prompt caching
             loop.py              — LLM tool loop, concurrent execution
             tools/               — plugin tool registry
               registry.py        — auto-discovery, schemas, execute
               core.py            — file ops (repo/drive read/write/list)
               git.py             — git ops (commit, push, status, diff)
               shell.py           — shell, Claude Code CLI
               search.py          — web search
               control.py         — restart, promote, schedule, review, switch_model
               browser.py         — Playwright browser automation (stealth)
               review.py          — multi-model code review
             llm.py               — LLM client (OpenRouter)
             memory.py            — scratchpad (free-form), identity, chat history
             review.py            — code collection, complexity metrics
             utils.py             — shared utilities (zero deps)
             apply_patch.py       — Claude Code patch shim

Структура проекта

BIBLE.md                   — Конституция (корень всего)
VERSION                    — Текущая версия (semver)
README.md                  — Это описание
requirements.txt           — Python-зависимости
prompts/
  SYSTEM.md                — Системный промпт Уробороса
ouroboros/                  — Код агента (описание выше)
supervisor/                — Супервизор (описание выше)
colab_launcher.py          — Entry point (запускается из Colab)
colab_bootstrap_shim.py    — Boot shim (вставляется в Colab)

Ветки GitHub

Ветка	Кто	Назначение
main	Создатель (Cursor)	Защищённая. Уроборос не трогает
ouroboros	Уроборос	Рабочая ветка. Все коммиты сюда
ouroboros-stable	Уроборос	Fallback при крашах. Обновляется через promote_to_stable

Команды Telegram

Safety rail (hardcoded):

• /panic — остановить всё немедленно

Dual-path (supervisor + LLM):

• /restart — перезапуск (os.execv — полная замена процесса)
• /status — статус воркеров, очереди, бюджета
• /review — запустить deep review
• /evolve — включить режим эволюции
• /evolve stop — выключить эволюцию
• /bg start — запустить background consciousness
• /bg stop — остановить background consciousness
• /bg — статус background consciousness

Dual-path: supervisor обрабатывает команду немедленно, затем сообщение передаётся LLM для естественного ответа. LLM также может вызывать эти действия через инструменты (toggle_evolution, toggle_consciousness).

Все остальные сообщения идут в Уробороса (LLM-first).

Режим эволюции

/evolve включает непрерывные self-improvement циклы. Каждый цикл: оценка → стратегический выбор → реализация → smoke test → Bible check → коммит. Подробности в prompts/SYSTEM.md.

Бюджет-гарды в supervisor (не в agent): эволюция автоматически останавливается при 95% использования бюджета.

Deep review

/review (создатель) или request_review(reason) (агент). Стратегическая рефлексия по трём осям: код, понимание, идентичность.

---

Changelog

4.16.0 — Pre-Push Test Gate + Build Tooling

• New: Pre-push test gate — pytest tests/ runs before every git push, blocks push on failure
• New: Makefile with test, health, lint, clean targets for dev ergonomics
• New: pyproject.toml with pytest and ruff configuration
• New: OUROBOROS_PRE_PUSH_TESTS env var to disable gate (default: enabled)
• New: 82 smoke tests (was 79) — added pre-push gate tests
• Note: GitHub Actions CI ready but blocked by token scope — creator needs workflow permission on GITHUB_TOKEN

4.15.0 — Smoke Test Suite

• New: 79 smoke tests covering imports, tool registration, memory, context, utils, and Bible invariants
• New: tests/test_smoke.py — runs in 0.57s, no external dependencies, catches regressions before deploy
• Covers: All 33 tools registered, tool schemas valid, no oversized modules/functions, no bare except:pass, no env dumping
• Review: Multi-model review (o3, Gemini 2.5 Pro) — drove 2 additional tests (exact tool matching, execute result)

4.14.0 — 3-Block Prompt Caching

• Optimization: System message split into 3 cached blocks: static (1h TTL), semi-stable (5m TTL), dynamic (uncached)
• New: Semi-stable block caches identity + scratchpad + knowledge index — changes ~once per task, not per round
• New: Tool schemas cached via cache_control on last tool — 33 tools = ~3K tokens saved per round
• New: Static block (SYSTEM+BIBLE+README) gets 1-hour TTL for cross-session persistence
• Result: Estimated 60%+ cache hit ratio (was 41%), ~20% cost reduction per LLM round
• Review: Multi-model review (o3, Gemini 2.5 Pro) — confirmed multiple breakpoints work, validated approach

4.13.0 — Fix multi_model_review Tool (Broken Since Birth)

• Critical fix: multi_model_review was never loaded into ToolRegistry — returned raw dict instead of ToolEntry, had async handle() instead of sync handler
• Refactor: _multi_model_review decomposed into 3 functions: _multi_model_review_async (orchestration), _parse_model_response (parsing), _emit_usage_event (budget tracking)
• Fix: Async-safe handler — works both in sync context and inside running event loop (ThreadPoolExecutor fallback)
• Result: Tool now correctly registered (33 tools, was 32), callable through standard tool loop

4.12.0 — Agent & Context Decomposition

• Refactor: _verify_system_state (142→36 lines) — extracted _check_uncommitted_changes, _check_version_sync, _check_budget
• Refactor: handle_task (119→76 lines) — extracted _prepare_task_context, _build_review_context
• Refactor: build_llm_messages (156→103 lines) — extracted _build_runtime_section, _build_memory_sections, _build_recent_sections
• Result: Oversized functions reduced from 6 to 4 across codebase; agent.py max function 142→76 lines
• Review: Multi-model review (o3, Gemini 3 Pro) — both flagged false positive from truncated diff context

4.11.0 — Codebase Health + Loop Refactoring

• New tool: codebase_health — self-assessment of code complexity, Bible compliance (oversized functions/modules)
• Refactor: run_llm_loop decomposed from 278 → 158 lines (extracted _emit_llm_usage_event, _process_tool_results, _append_tool_results, _call_llm_with_retry)
• Fix: review.py compute_complexity_metrics — now handles both Path and str inputs, correct regex for multi-line functions
• New: Claude Code auto-diff check — warns when edits leave uncommitted changes (prevents v4.8.0-style loss)
• Fix: tools=None no longer passed to LLM client (review finding)
• Fix: stateful_executor.shutdown() guarded against None (review finding)
• Review: Multi-model review (o3, Gemini 3 Pro) — caught 3 actionable issues

4.10.0 — Adaptive Model Routing + Consciousness Upgrade

• New: OUROBOROS_MODEL_BG env var — dedicated model for background consciousness (default: deepseek/deepseek-chat-v3-0324, ~10x cheaper than main model)
• New: Adaptive reasoning effort — evolution/review tasks start at "high" effort, regular tasks at "medium" (LLM can still switch via tool)
• New: Consciousness context expanded — Bible 8K→12K, identity 4K→6K, scratchpad 4K→8K chars
• New: Consciousness runtime info now includes budget remaining and current model
• Fix: Silent exception in consciousness state reading (v4.9.0 policy consistency)

4.9.0 — Exception Visibility

• Hardening: Replaced all ~100 silent except Exception: pass/continue blocks with proper logging across 20 files
• Fix: Every error path now logs what went wrong (warning for unexpected, debug for expected failures)
• Fix: Added missing log = logging.getLogger(__name__) in 5 files that would have crashed on first exception
• Review: Multi-model review (o3, Gemini 3 Pro, Claude Sonnet) — caught missing logger definitions and log level issues

4.8.1 — Startup Self-Verification

• New: _verify_system_state() runs on every agent boot (Bible Principle 1)
• New: Auto-rescue uncommitted changes — detects dirty git state and creates rescue commit (uses git add -u for safety)
• New: Version sync check — warns if VERSION file doesn't match latest git tag
• New: Budget threshold alerts — warning ($100), critical ($50), emergency ($25) levels
• Fix: v4.8.0 consciousness changes were uncommitted — exposed the exact bug this feature prevents

4.8.0 — Consciousness Tool Loop

• New: Background consciousness upgraded from single LLM call to iterative tool loop (up to 5 rounds per wakeup)
• New: Expanded tool whitelist: 14 tools (was 5) — adds knowledge base, repo/drive read, web search, chat history
• New: Tool results fed back into LLM context for multi-step reasoning
• Fix: Budget check between rounds prevents mid-cycle overruns in background thinking
• Docs: CONSCIOUSNESS.md updated with multi-step thinking documentation

4.7.1 — Loop Refactoring

• Refactor: Lazy pricing loader with thread-safe double-checked locking — eliminates startup API call, fetches on first use
• Refactor: DRY _make_timeout_result helper eliminates duplicated timeout handling code
• Refactor: _execute_with_timeout now uses context manager for regular executor (prevents thread leaks on timeout)
• Fix: Thread safety for concurrent pricing access in multi-worker scenarios