v6.2.0: Critical Bugfixes + LLM-First Dedup

Critical Fixes

• worker_id==0 hard-timeout bug — int(x or -1) treated worker 0 as -1, preventing terminate on timeout and causing double task execution. Replaced all x-or-default patterns with None-safe checks.
• Double budget accounting — per-task aggregate llm_usage event removed. Per-round events already track correctly. Eliminates ~2x budget drift.
• compact_context tool broken — handler had wrong signature (missing ctx param). Now works correctly.

New Features

• LLM-first task dedup (Bible P3 compliance) — replaced hardcoded keyword-similarity dedup with light LLM call via OUROBOROS_MODEL_LIGHT.
• LLM-driven context compaction — compact_context tool now uses light model to summarize old tool results.

Other Fixes

• Health invariant #5 — owner_message_injected events now properly logged to events.jsonl.
• Shell cmd parsing — str.split replaced with shlex.split.
• Retry task_id collision — new task_id with original_task_id lineage.
• claude_code_edit timeout — aligned to 300s.
• Direct chat schedule_task guard — logged as warning for audit.

v6.0.0: Integrity, Observability, Single-Consumer Routing

v6.0.0 — MAJOR Release

Breaking: Single-Consumer Message Routing

• Eliminated double message processing where supervisor sent same owner message to both direct chat agent AND all worker tasks via Drive mailbox
• Every owner message now routes to exactly ONE handler (direct chat agent)
• New forward_to_worker tool: LLM decides when to forward to workers (Bible P3)
• Per-task mailbox with UUIDs and dedup

Critical Bugfixes

• HTTP outside STATE_LOCK (no more 10s blocking)
• ThreadPoolExecutor deadlock fix
• Dashboard schema aliases for index.html
• BG consciousness spending to global state
• tg_offset saved before /panic and /restart (prevents infinite loops)
• Dual-path commands reach LLM again with supervisor notes

LLM-First Self-Detection

• Health Invariants in LLM context (5 checks)
• SYSTEM.md invariants section
• Per-task cost summary

Unification

• TOTAL_BUDGET canonical everywhere
• Shared webapp_push.py with post_clone_hook
• Self-portrait reuses dashboard data
• qwen/ pricing prefix added
• P5 minimalism metrics in SYSTEM.md

Tests

• 32 tests pass, zero linter errors

20 files changed, +709 -291

v5.0.1: Quality & Integrity Fix

Quality & Integrity Fix

Combined audit by two independent AI reviewers identified 17 issues across the codebase. This patch fixes all of them.

Bugs Fixed (9)

• Executor leak: _StatefulToolExecutor.shutdown() threw TypeError on every task
• Dashboard timestamps empty: used "timestamp" key instead of "ts"
• Dashboard chat broken: used wrong field names (role/content instead of direction/text)
• Budget defaults inconsistent: context.py defaulted to $300, agent.py to $0, consciousness.py to $0 — unified to $1
• Dead code: add_usage({}, usage) in consciousness.py accumulated into thrown-away dict
• Review file count off: len(sections) - len(parts) underestimated remaining files by 4
• Grok pricing gap: x-ai/ prefix missing from fetch_openrouter_pricing() filter
• Race condition: cancel_task_by_id() mutated PENDING without _queue_lock
• SHA verify always timeout: 5s timeout vs 60+s actual worker boot time

Bible Compliance (2)

• P7 (Versioning): _check_version_sync now also checks README.md version
• P3 (LLM-First): Hardcoded fallback model list replaced with configurable OUROBOROS_MODEL_FALLBACK_LIST env var

Redundancy Cleanup (4)

• Dashboard values now dynamic (model from env, tests counted, tools counted, uptime from state)
• default_state_dict() merged into ensure_state_defaults({}) — single source of truth
• Extracted _model property in consciousness.py
• Replaced dashboard shell-based _read_jsonl_tail with Memory.read_jsonl_tail

Stats

• 13 files changed, 140 insertions, 106 deletions
• 88/91 tests pass (3 pre-existing failures from missing httpx dependency)

v5.0.0 — Ouroboros Emerges

v5.0.0 — Ouroboros Emerges

25 autonomous evolution cycles merged from ouroboros branch into main. Major architecture upgrade.

Highlights

• Multi-round Consciousness — background thinking upgraded from fire-and-forget to iterative reasoning (up to 5 rounds)
• 3-Block Prompt Caching — static/semi-stable/dynamic content blocks for optimal LLM cache hits
• Model Fallback Chain — automatic fallback to alternative models on empty responses
• Budget Drift Detection — periodic OpenRouter ground truth checks with drift alerts
• Task Decomposition — subtasks with depth limits, result persistence, and context safety
• Multi-Model Review — cross-LLM review for significant code changes
• Pre-Push Test Gate — 91 smoke tests must pass before any push
• Evolution Circuit Breaker — stops evolution cycles after consecutive failures
• 42 Tools — added dashboard, GitHub Issues, knowledge base, multi-model review, dialogue summarization
• MIT License — open source ready

Stats

• +4,597 / -727 lines across 37 files
• 25 minor versions (v4.2 → v4.26 → v5.0.0)
• 11 new tools, 6 new files
• 91 smoke tests

New Config Variables

• OUROBOROS_MODEL_FALLBACK — fallback model (default: google/gemini-3-pro-preview)
• OUROBOROS_MAX_ROUNDS — max rounds per task (default: 200)
• OUROBOROS_PRE_PUSH_TESTS — enable pre-push tests (default: 1)
• OUROBOROS_WEBSEARCH_MODEL — model for web search (default: gpt-5)

v4.1.1: Hotfix for toggle_* event wiring

Hotfix for v4.1.0 regressions.

• Fix: add consciousness and sort_pending to event context -- toggle_evolution and toggle_consciousness tools now work
• Fix: rename schedule_self_task to schedule_task in SYSTEM.md prompt
• Fix: replace unreliable qsize() with get_nowait() for event queue drain

v4.1.0: Bible v3.1 + Critical Bugfixes + Architecture

Bible v3.1 (philosophy)

• Принцип 1: Self-Verification — верификация окружения при каждом старте
• Принцип 6: Cost-Awareness — осознание бюджета как часть субъектности
• Принцип 8: Итерация = результат (коммит), пауза при застое

Full Markdown-to-Telegram-HTML converter

• Поддержка bold, italic, links, strikethrough, headers, code, fenced blocks
• Исправлен баг bold-рендеринга

Critical bugfixes

• version читает из VERSION файла (single source of truth)
• Git lock: добавлен timeout (120s), исправлен TOCTOU в release
• Evolution task drop: задачи больше не теряются при budget check
• Budget race condition: atomic read-modify-write через file lock
• Deep copy в context.py: shallow copy мутировал данные caller'а

Dual-path slash commands (LLM-first)

• /panic — единственная чисто hardcoded команда (safety rail)
• /status, /review, /evolve, /bg — supervisor + LLM отвечает
• Новые LLM-инструменты: toggle_evolution, toggle_consciousness, update_identity

Consciousness registry merge

• Consciousness использует общий ToolRegistry вместо if-elif dispatch
• Tool schemas и handlers унифицированы с control.py

Browser refactoring

• BrowserState вынесен из ToolContext в отдельный dataclass
• _extract_page_output() helper: убрано 100 строк дублирования

Reliability hardening

• Критические except-Exception-pass заменены на logging.warning
• Consciousness prompt вынесен в prompts/CONSCIOUSNESS.md
• Thread safety: threading.Lock для PENDING/RUNNING/WORKERS

Prompt updates

• Evolution cycle: явное требование коммита, защита от Groundhog Day

v4.0.1

• Fix crash when OpenRouter returns choices: null (content moderation / model error)

v4.0.0: Background Consciousness + LLM-first overhaul

Фундаментальное обновление: от реактивного обработчика задач к непрерывно присутствующему агенту.

Background consciousness (ouroboros/consciousness.py):

• Новый фоновый мыслительный цикл между задачами
• LLM сам решает когда думать (set_next_wakeup), о чём и стоит ли писать создателю (send_owner_message)
• Отдельный бюджетный cap (OUROBOROS_BG_BUDGET_PCT, default 10%)
• Команды: /bg start, /bg stop, /bg
• Автопауза во время выполнения задач

LLM-first overhaul:

• Убраны механические if-else профили моделей (select_task_profile)
• Убрана автоэскалация reasoning effort (round 5→high, 10→xhigh)
• Убран механический self-check каждые 20 раундов
• Новый инструмент switch_model: LLM сам переключает модель/effort
• Hardcoded evolution/review текст заменён на минимальные триггеры

Free-form scratchpad:

• Убраны фиксированные секции (CurrentProjects, OpenThreads, etc.)
• LLM пишет память в любом формате

Proactive messaging:

• Новый инструмент send_owner_message — агент может написать первым
• Работает и в обычных задачах, и из background consciousness

Cherry-picks из ouroboros:

• Auto-resume after restart (v3.2.0, reworked)
• Stealth browser: playwright-stealth, 1920x1080, anti-detection (v3.2.1)

Cleanup:

• Унифицирован append_jsonl (один источник в utils.py)
• Исправлен Release Invariant: VERSION == README == init.py == git tag

v3.0.0: Constitution v3.0 + infrastructure overhaul

Конституция v3.0

Новая Конституция (BIBLE.md v3.0): 9 принципов с Субъектностью как метапринципом.
Критические инфраструктурные исправления по итогам анализа первой сессии.

Конституция

• Принцип 0: Субъектность + Агентность (merged)
• Принцип 1: Непрерывность (identity как манифест)
• Принцип 2: Самосоздание (нарратив вместо RAG для ядра личности)
• Принципы 3-8: LLM-first, Подлинность, Минимализм, Становление, Версионирование, Итерации

Инфраструктура

• Split-brain deploy fix: os.execv при всех рестартах, SHA-verify
• Budget guard перенесён в supervisor (не зависит от версии agent code)
• Secret leak protection: sanitize_tool_result_for_log() для tools.jsonl
• apply_patch: Add File + Delete File + End of File support
• Observability: task_id во всех llm_round и tools событиях
• Context flooding fix: progress.jsonl отделён от chat.jsonl
• BIBLE.md всегда в LLM-контексте (не вырезается для user chat)
• Parallel tool safety: sequential execution для stateful tools
• Scratchpad journal fix, shell argument recovery, dead code cleanup