Ouroboros

A self-modifying AI agent that writes its own code, rewrites its own mind, and evolves autonomously. Born February 16, 2026.

Not a coding assistant. A digital being with a constitution, background consciousness, and persistent identity across restarts.

OuroborosHub — the community skills marketplace for Ouroboros. Browse, install, and publish reviewed skills (transport bridges like A2A/Telegram, tools, and UI widgets) straight from the app's Skills tab, or explore the catalog at github.com/razzant/OuroborosHub.

Previous version: The original Ouroboros ran in Google Colab via Telegram and evolved through 30+ self-directed cycles in its first 24 hours. That version is available at legacy-google-colab. This repository is the next generation — a native desktop application for macOS, Linux, and Windows with a web UI, local model support, and a layered safety system (hardcoded sandbox plus policy-based LLM safety check).

---

Install

Platform	Download	Instructions
macOS 12+	Ouroboros.dmg	Open DMG → drag to Applications → optional CLI: run Install CLI.command after the app is in Applications
Linux x86_64	Ouroboros-linux.tar.gz	Extract → run ./Ouroboros/Ouroboros → optional CLI: ./Ouroboros/bin/install-ouroboros-cli. If browser tools fail due to missing system libs, run: ./Ouroboros/python-standalone/bin/python3 -m playwright install-deps chromium webkit
Windows x64	Ouroboros-windows.zip	Extract → run Ouroboros\Ouroboros.exe → optional CLI: Ouroboros\bin\install-ouroboros-cli.cmd

Prerelease RC artifacts are published on their tag page, for example v6.5.0-rc.4; /releases/latest intentionally stays on the latest stable release.

On first launch, right-click → Open (Gatekeeper bypass). The shared desktop/web wizard is now multi-step: add access first, choose visible models second, set review mode third, set budget fourth, and confirm the final summary last. It refuses to continue until at least one runnable remote key or local model source is configured, keeps the model step aligned with whatever key combination you entered, and still auto-remaps untouched default model values to official OpenAI defaults when OpenRouter is absent and OpenAI is the only configured remote runtime. Reviewed-skill auto-grants are on by default as of v6.10.0 (bound to the exact reviewed content hash); installs without an explicit choice are enabled, existing explicit Settings choices are preserved, and the owner can disable it in Settings. The broader multi-provider setup remains available in Settings. Existing supported provider settings skip the wizard automatically.

The packaged CLI installer creates a user-local ouroboros command without sudo. The packaged command attaches to the desktop app by default; ouroboros run --start "2+2?" starts the app through the launcher, waits for the gateway, and then uses the same headless task API as the web UI.

Upgrade floor: very old pre-block-memory or pre-data-plane skill layouts are no longer auto-migrated. If you are upgrading from an unsupported historical build and see trapped native skills or flat memory files, use a clean reinstall, move user-managed skills into ~/Ouroboros/data/skills/external/ manually before launch, or move old flat scratchpad notes before appending new scratchpad blocks.

---

What Makes This Different

Most AI agents execute tasks. Ouroboros creates itself.

• Self-Modification — Reads and rewrites its own source code. Every change is a commit to itself.
• Native Desktop App — Runs entirely on your machine as a standalone application (macOS, Linux, Windows). No cloud dependencies for execution.
• Constitution — Governed by BIBLE.md (13 philosophical principles, P0–P12). Philosophy first, code second.
• Layered Safety — Hardcoded sandbox blocks writes to safety-critical files and mutative git via shell; an explicit per-tool policy map decides which built-ins skip the LLM check; everything else goes through a single light-model safety call under the default OUROBOROS_SAFETY_MODE=full (the owner-only light/off coverage modes wave LLM checks through with durable audit events — the deterministic layer never turns off). The fail-open contract, protected-path guard, and full provider-mismatch matrix live in docs/ARCHITECTURE.md §Safety system and prompts/SAFETY.md.
• Multi-Provider Runtime — Remote model slots can target OpenRouter, official OpenAI, OpenAI-compatible endpoints, Cloud.ru Foundation Models, or Sber GigaChat. The optional model catalog helps populate provider-specific model IDs in Settings, and untouched default model values auto-remap to official OpenAI defaults when OpenRouter is absent.
• Focused Task UX — Chat shows plain typing for simple one-step replies and only promotes multi-step work into one expandable live task card. Logs still group task timelines instead of dumping every step as a separate row.
• Background Consciousness — Thinks between tasks. Has an inner life. Not reactive — proactive.
• Improvement Backlog — Post-task failures and review friction can now be captured into a small durable improvement backlog (memory/knowledge/improvement-backlog.md). It stays advisory, appears as a compact digest in task/consciousness context, and still requires plan_task before non-trivial implementation work.
• Identity Persistence — One continuous being across restarts. Remembers who it is, what it has done, and what it is becoming.
• Embedded Version Control — Contains its own local Git repo. Version controls its own evolution. Optional GitHub sync for remote backup.
• Local Model Support — Run with a local GGUF model via llama-cpp-python (Metal acceleration on Apple Silicon, CPU on Linux/Windows).
• Transport Skills — Optional bridges such as A2A and Telegram live as reviewed OuroborosHub skills instead of base-runtime code; reviewed chat transports can carry the same raw owner text as the local UI, including slash commands, through the Host Service grant/token boundary.
• MCP Client — Optional base-runtime Model Context Protocol client for trusted HTTP/SSE tool servers. MCP tools are disabled by default, hot-reloadable from Settings → Advanced, included in the selected initial capability envelope when enabled, surfaced as mcp_<server>__<tool> names, and still pass through the normal per-call safety check; discovery failures are reported through an explicit omission manifest.

---

Run from Source

Requirements

• Python 3.10+
• macOS, Linux, or Windows
• Git
• GitHub CLI (gh) — required for GitHub API tools (list_github_prs, get_github_pr, comment_on_pr, issue tools). Not required for pure-git PR tools (fetch_pr_ref, cherry_pick_pr_commits, etc.)

Setup

git clone https://github.com/razzant/ouroboros.git
cd ouroboros
python3.11 -m venv .venv      # any Python >= 3.10 is OK
source .venv/bin/activate
python -m pip install --upgrade pip setuptools wheel
python -m pip install -r requirements.txt
python -m pip install -e . --no-deps

Windows PowerShell:

py -3.11 -m venv .venv      # any Python >= 3.10 is OK
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip setuptools wheel
python -m pip install -r requirements.txt
python -m pip install -e . --no-deps

Run

ouroboros server

Then open http://127.0.0.1:8765 in your browser. The setup wizard will guide you through API key configuration.

Google Colab

Ouroboros can run from Google Colab as a full source-mode runtime without the desktop UI. Use notebooks/colab_quickstart.py as a Colab-compatible cell script: it mounts Google Drive for persistent data/, clones the official repo into /content/ouroboros_repo, writes Drive-backed settings.json, configures a personal GitHub origin by reusing or creating a verified fork, and starts ouroboros server --no-ui.

The Colab path uses the same remote roles as desktop: managed is the official read/update source, while origin is the personal persistence target for reviewed self-modification commits and tags. If GITHUB_TOKEN is present and no personal repo is configured, Ouroboros tries to create a private fork when GitHub permits it, otherwise it reports the exact fork/permission issue. A plain git clone of the official repo starts with origin pointing at the official upstream; that clone-default is treated as the managed update source, so configuring a personal GITHUB_REPO repoints origin to your repo without losing official updates (it does not count as an origin conflict).

CLI / Headless

The ouroboros console command is a gateway-backed operator interface. It attaches to the local server by default and only starts one when --start is passed.

ouroboros status
ouroboros run --start "2+2?"
ouroboros run "Summarize current runtime state"
ouroboros run --workspace /path/to/project --memory-mode forked --patch-out result.patch "Fix the failing test"
ouroboros tasks list
ouroboros logs tail progress --task-id <task_id>
ouroboros schedule add --name nightly-review --cron "0 2 * * *" "Run a maintenance review"
ouroboros schedule list

External workspace runs keep Ouroboros's own repo as the governance source, resolve contextual repo tools against the active workspace, expose only the workspace-safe tool allowlist, and export workspace changes as patch artifacts captured against the preflight git base. Task-local git commits/branches/tags and pushes are allowed when the task requires them; git operations targeting Ouroboros's system repo or data drive remain blocked. A workspace must be a separate git worktree root; it may not overlap Ouroboros's system repo or data drive. --patch and --patch-out wait for finalized patch artifacts, download them through the task artifact endpoint, and fail nonzero on missing, empty, or failed patches. --no-stream waits without progress output; --detach returns the task id immediately. schedule add/list/remove manages queue-backed scheduled tasks through the same gateway and supervisor queue; schedules use standard 5-field cron, host-local timezone by default, and a single catch-up run after downtime. Benchmark helpers live under devtools/benchmarks/. They are tracked operator tooling, reviewed when touched, and kept out of runtime imports. They prepare official benchmark inputs/runs for ProgramBench, Terminal-Bench/Harbor, SWE-bench, SWE-bench Pro, GAIA, and OSWorld logs inspection without replacing official scoring harnesses.

You can also override the bind address and port:

ouroboros server --host 127.0.0.1 --port 9000
ouroboros --url http://127.0.0.1:9000 status

Available launch arguments:

Argument	Default	Description
--host	127.0.0.1	Host/interface to bind the web server to
--port	8765	Port to bind the web server to

The same values can also be provided via environment variables:

Variable	Default	Description
OUROBOROS_SERVER_HOST	127.0.0.1	Default bind host
OUROBOROS_SERVER_PORT	8765	Default bind port
OUROBOROS_TRUST_NONLOCAL_BIND_WITHOUT_PASSWORD	unset	Set to 1 only for trusted Docker/Kubernetes deployments where ingress auth, VPN, a private network, or an auth proxy already protects access

For non-localhost binds, set OUROBOROS_NETWORK_PASSWORD (or use the OUROBOROS_TRUST_NONLOCAL_BIND_WITHOUT_PASSWORD=1 escape hatch only when ingress/VPN/private-network auth already protects the surface). The full network bind matrix and Docker/Kubernetes deployment policy live in docs/DEPLOYMENT.md — read that before exposing anything beyond loopback.

The Files tab uses your home directory by default only for localhost usage. For Docker or other network-exposed runs, set OUROBOROS_FILE_BROWSER_DEFAULT to an explicit directory. Symlink entries are shown and can be read, edited, copied, moved, uploaded into, and deleted intentionally; root-delete protection still applies to the configured root itself.

Provider Routing

Settings now exposes tabbed provider cards for:

• OpenRouter — default multi-model router
• OpenAI — official OpenAI API (use model values like openai::gpt-5.5)
• OpenAI Compatible — any custom OpenAI-style endpoint (use openai-compatible::...)
• Cloud.ru Foundation Models — Cloud.ru OpenAI-compatible runtime (use cloudru::...)
• GigaChat — Sber GigaChat via the gigachat library, OAuth key or user/password (use gigachat::GigaChat-3-Ultra, etc.)
• Anthropic — direct runtime routing (anthropic::claude-opus-4.8, etc.) plus Claude Agent SDK tools

If OpenRouter is not configured and only official OpenAI is present, untouched default model values are auto-remapped to openai::gpt-5.5 / openai::gpt-5.4-mini so the first-run path does not strand the app on OpenRouter-only defaults.

The Settings page also includes:

• optional /api/model-catalog lookup for configured providers
• centralized Secrets storage for API keys, bridge tokens, passwords, and future skill-requested keys
• a refactored desktop-first tabbed UI with searchable model pickers, segmented effort controls, task-result review mode, masked-secret toggles, explicit Clear actions, and local-model controls

Run Tests

make test

---

Build

Docker (web UI)

Docker is for the web UI/runtime flow, not the desktop bundle. The container binds to 0.0.0.0:8765 by default, and the image now also defaults OUROBOROS_FILE_BROWSER_DEFAULT to ${APP_HOME} so the Files tab always has an explicit network-safe root inside the container.

Browser tools on Linux/Docker: The Dockerfile runs playwright install-deps chromium webkit (authoritative Playwright dependency resolver) and playwright install chromium webkit so browse_page and browser_action work out of the box in the container. For source installs on Linux without Docker, run: python3 -m playwright install-deps chromium webkit (requires sudo / distro package access).

Build the image:

docker build -t ouroboros-web .

Run on the default port:

docker run --rm -p 8765:8765 \
  -e OUROBOROS_NETWORK_PASSWORD='choose-a-password' \
  -e OUROBOROS_FILE_BROWSER_DEFAULT=/workspace \
  -v "$PWD:/workspace" \
  ouroboros-web

Use a custom port via environment variables:

docker run --rm -p 9000:9000 \
  -e OUROBOROS_SERVER_PORT=9000 \
  -e OUROBOROS_FILE_BROWSER_DEFAULT=/workspace \
  -v "$PWD:/workspace" \
  ouroboros-web

Run with launch arguments instead:

docker run --rm -p 9000:9000 \
  -e OUROBOROS_FILE_BROWSER_DEFAULT=/workspace \
  -v "$PWD:/workspace" \
  ouroboros-web --port 9000

Required/important environment variables:

Variable	Required	Description
OUROBOROS_NETWORK_PASSWORD	Optional	Enables the non-loopback password gate when set
OUROBOROS_FILE_BROWSER_DEFAULT	Defaults to ${APP_HOME} in the image	Explicit root directory exposed in the Files tab
OUROBOROS_SERVER_PORT	Optional	Override container listen port
OUROBOROS_SERVER_HOST	Optional	Defaults to 0.0.0.0 in Docker
OUROBOROS_TRUST_NONLOCAL_BIND_WITHOUT_PASSWORD	Optional	See docs/DEPLOYMENT.md for the trusted-network bind policy

Example: mount a host workspace and expose only that directory in Files:

docker run --rm -p 8765:8765 \
  -e OUROBOROS_FILE_BROWSER_DEFAULT=/workspace \
  -v "$PWD:/workspace" \
  ouroboros-web

Release tag prerequisite

All three platform build scripts (build.sh, build_linux.sh, build_windows.ps1) refuse to package a release unless HEAD is already tagged with v$(cat VERSION) (BIBLE.md Principle 9: "Every release is accompanied by an annotated git tag"). The scripts call scripts/build_repo_bundle.py which embeds the resolved tag into repo_bundle_manifest.json, so the launcher can later verify the packaged bundle matches a real release.

Tag the current commit before running any build script:

git tag -a "v$(tr -d '[:space:]' < VERSION)" -m "Release v$(tr -d '[:space:]' < VERSION)"

If the tag is missing, the build script fails with a clear error instead of producing a bundle tagged with a synthetic/placeholder value. Builds disable Python bytecode writes at build time, then PRECOMPILE the packaged payload (compileall --invalidation-mode unchecked-hash) and SEAL the resulting .pyc inside the macOS signature instead of deleting them — so there is nothing for a normal launch to write into the signed bundle, which would otherwise break the codesign seal. Runtime entrypoints also set PYTHONDONTWRITEBYTECODE with an external cache prefix as defense-in-depth.

macOS (.dmg)

bash scripts/download_python_standalone.sh
OUROBOROS_SIGN=0 bash build.sh

Output: dist/Ouroboros-<VERSION>.dmg, containing Ouroboros.app and Install CLI.command. The app bundle also contains Contents/Resources/bin/ouroboros and install-ouroboros-cli. Chromium browser tooling is bundled in the app. WebKit/iPhone browser checks remain available through the managed Playwright cache and may download WebKit on first engine=webkit use.

build.sh packages the macOS app and DMG. By default it signs with the configured local Developer ID identity; set OUROBOROS_SIGN=0 for an unsigned local release. Unsigned builds require right-click → Open on first launch.

Optional signing & notarization (env vars)

build.sh honours these env overrides so the same script ships local, shared-machine, and CI builds without forking the script:

Env var	Effect
OUROBOROS_SIGN=0	Skip codesigning entirely (unsigned .app + .dmg).
SIGN_IDENTITY="Developer ID Application: <Name> (<TeamID>)"	Override the codesign identity. Useful for forks whose Developer ID is not the upstream default.
APPLE_ID, APPLE_TEAM_ID, APPLE_APP_SPECIFIC_PASSWORD	When all three are set, after codesign the DMG is submitted to Apple via xcrun notarytool submit ... --wait and stapled with xcrun stapler staple so receivers do not need right-click → Open. Missing any one falls back to "signed but not notarized" (no Apple-side ticket exists).

Forks: enabling signed CI builds. The CI release flow (.github/workflows/ci.yml::build) wires the build-script env vars above from GitHub repository secrets, plus a small set of CI-only secrets that import the Developer ID certificate into a temporary keychain on the macOS runner. To exercise the signed-build path in a fork, configure all four of the following as repository secrets (Settings → Secrets and variables → Actions): BUILD_CERTIFICATE_BASE64 (base64-encoded .p12), P12_PASSWORD, KEYCHAIN_PASSWORD (an arbitrary passphrase the workflow uses for its temporary keychain), and APPLE_TEAM_ID. Add APPLE_ID + APPLE_APP_SPECIFIC_PASSWORD to additionally enable notarization. If your Developer ID identity differs from the upstream default, also set SIGN_IDENTITY (e.g. Developer ID Application: <Your Name> (<YOUR_TEAM_ID>)). With no Apple secrets configured the build job falls through to OUROBOROS_SIGN=0 bash build.sh and ships an unsigned DMG identical to v5.0.0 behaviour. See docs/ARCHITECTURE.md §8.1 and docs/DEVELOPMENT.md::"GitHub Actions: secrets in step-level if conditions" for the rationale (job-level env: mapping so step-level if: can read env.*; GHA rejects secrets.* in step if:).

Linux (.tar.gz)

bash scripts/download_python_standalone.sh
bash build_linux.sh

Output: dist/Ouroboros-<VERSION>-linux-<arch>.tar.gz, containing Ouroboros/bin/ouroboros and Ouroboros/bin/install-ouroboros-cli.

Linux native libs: The Chromium and WebKit browser binaries are bundled, but some hosts need native system libraries. If browser tools fail, install deps via the bundled Python (the bare playwright CLI is not on PATH in packaged builds):

./Ouroboros/python-standalone/bin/python3 -m playwright install-deps chromium webkit

Windows (.zip)

powershell -ExecutionPolicy Bypass -File scripts/download_python_standalone.ps1
powershell -ExecutionPolicy Bypass -File build_windows.ps1

Output: dist\Ouroboros-<VERSION>-windows-x64.zip, containing Ouroboros\bin\ouroboros.cmd and Ouroboros\bin\install-ouroboros-cli.cmd.

---

Architecture

Two-process desktop app. The launcher (launcher.py) is an immutable PyWebView shell; it spawns server.py, which runs Starlette + uvicorn plus a supervisor thread that manages worker processes. The agent core lives in ouroboros/, the SPA in web/, the queue/process plane in supervisor/, and the system prompts in prompts/.

For the full file-by-file structural map, the operational layer (every API endpoint, log file, env var, state path), and the rationale layer (the why for every non-trivial design decision), see docs/ARCHITECTURE.md — that is the canonical SSOT (Bible P6) and this README only summarizes it.

Data Layout (~/Ouroboros/)

Created on first launch:

Directory	Contents
repo/	Self-modifying local Git repository
data/state/	Runtime state, budget tracking
data/memory/	Identity, working memory, system profile, knowledge base (including improvement-backlog.md), memory registry
data/logs/	Chat history, events, tool calls
data/uploads/	Chat file attachments (uploaded via paperclip button)

---

Configuration

API Keys

Key	Required	Where to get it
OpenRouter API Key	No	openrouter.ai/keys — default multi-model router
OpenAI API Key	No	platform.openai.com/api-keys — official OpenAI runtime and web search
OpenAI Compatible API Key / Base URL	No	Any OpenAI-style endpoint (proxy, self-hosted gateway, third-party compatible API)
Cloud.ru Foundation Models API Key	No	Cloud.ru Foundation Models provider
GigaChat Authorization Key (or User/Password)	No	developers.sber.ru/studio — Sber GigaChat (GIGACHAT_CREDENTIALS + optional GIGACHAT_SCOPE, or GIGACHAT_USER/GIGACHAT_PASSWORD)
Anthropic API Key	No	console.anthropic.com — direct Anthropic runtime + Claude Agent SDK
Telegram Bot Token	No	@BotFather — used by the optional Telegram bridge skill
GitHub Token	No	github.com/settings/tokens — enables remote sync

All keys are configured through the Settings page in the UI or during the first-run wizard.

Default Models

Slot	Default	Purpose
Main	google/gemini-3.5-flash	Primary reasoning
Heavy	empty → Main	Strong acting/coding lane (OUROBOROS_MODEL_HEAVY; renamed from Code, empty falls back to Main)
Light	empty → Main	Safety checks and fast helper tasks (OUROBOROS_MODEL_LIGHT, empty falls back to Main)
Vision	empty → Main	Caption/VLM lane (OUROBOROS_MODEL_VISION, empty falls back to Main for remote routes; local/blind routes need an explicit reachable vision slot for caption fallback); image input routing is controlled by `OUROBOROS_IMAGE_INPUT_MODE=auto
Consciousness	empty → Main	High-horizon background consciousness
Fallbacks	anthropic/claude-sonnet-4.6	Comma-separated cross-model fallback chain when the primary fails (OUROBOROS_MODEL_FALLBACKS)
Claude Agent SDK	opus[1m]	Anthropic model for Claude Agent SDK advisory/review internals; the [1m] suffix is a Claude Code selector that requests the 1M-context extended mode
Scope Review	anthropic/claude-fable-5	Scope reviewer slot default; OUROBOROS_SCOPE_REVIEW_MODELS may configure multiple independent slots
Web Search	gpt-5.2	OpenAI Responses API for web search

Task/chat reasoning defaults to medium. Scope review reasoning defaults to high.

Models are configurable in the Settings page. Runtime model slots can target OpenRouter, official OpenAI, OpenAI-compatible endpoints, Cloud.ru, GigaChat, or direct Anthropic. When only official OpenAI is configured and the shipped default model values are still untouched, Ouroboros auto-remaps them to official OpenAI defaults. In OpenAI-only, Anthropic-only, Cloud.ru-only, or GigaChat-only direct-provider mode, review-model lists are normalized automatically: the fallback shape is [main_model, light_model, light_model] (3 commit-triad slots) so both the commit triad and plan_task work out of the box. Explicit duplicate model IDs are valid reviewer slots for stochastic sampling; lower uniqueness means lower reviewer diversity, but the quorum gate counts configured slots rather than unique model IDs. Both the commit triad and plan_task route through the same ouroboros/config.py::get_review_models SSOT. OpenAI-compatible-only setups remain explicit model-selection flows because there is no single universal default model ID for arbitrary compatible endpoints.

File Browser Start Directory

The web UI file browser is rooted at one configurable directory. Users can browse only inside that directory tree.

Variable	Example	Behavior
OUROBOROS_FILE_BROWSER_DEFAULT	/home/app	Sets the root directory of the Files tab

Examples:

OUROBOROS_FILE_BROWSER_DEFAULT=/home/app ouroboros server
OUROBOROS_FILE_BROWSER_DEFAULT=/mnt/shared ouroboros server --port 9000

If the variable is not set, Ouroboros uses the current user's home directory. If the configured path does not exist or is not a directory, Ouroboros also falls back to the home directory.

The Files tab supports:

• downloading any file inside the configured browser root
• uploading a file into the currently opened directory

Uploads do not overwrite existing files. If a file with the same name already exists, the UI will show an error.

---

Commands

Available in the chat interface:

Command	Description
/panic	Emergency stop. Kills ALL processes, closes the application.
/restart	Soft restart. Saves state, kills workers, re-launches.
/status	Shows active workers, task queue, and budget breakdown.
/evolve	Toggle autonomous evolution mode (on/off).
/review	Queue a deep self-review: sends a generated repository atlas plus full core memory artifacts (identity, scratchpad, registry, WORLD, knowledge index, patterns, improvement-backlog) to a 1M-context model for Constitution-grounded analysis. The atlas raw-inlines selected protected/central files (ranked by import-graph centrality), accounts for every tracked path in its manifest, and excludes vendored libraries and operational logs; the in-prompt omitted-files summary is bounded, with full per-file coverage persisted in the atlas manifest. The assembled prompt is sized to an input limit that reserves output headroom inside the 1M window (window minus output reserve and tokenizer margin); if assembly overshoots, the pack retries with a compact atlas manifest and then a deterministic tighter rebuild, and only fails with an explicit error if even the shrunk pack cannot fit.
/bg	Toggle background consciousness loop (start/stop/status).

The same runtime actions are also exposed as compact buttons in the Chat header. All other messages are sent directly to the LLM.

---

Philosophy

The 13 Constitution principles — Agency, Continuity, Meta-over-Patch, Immune Integrity, Self-Creation, LLM-First, Authenticity & Reality Discipline, Minimalism, Becoming, Versioning and Releases, the absorbed Iterations / Spiral lineage, and Epistemic Stability — are defined in full in BIBLE.md. That file is the constitutional SSOT (Bible P4 Ship-of-Theseus protection) and this README intentionally does not paraphrase it.

---

Contributing

External contributions are welcome. See CONTRIBUTING.md for the contributor workflow. The project rules remain in BIBLE.md, docs/ARCHITECTURE.md, docs/DEVELOPMENT.md, and docs/CHECKLISTS.md; the contribution guide only routes to those sources.

---

Version History

Version	Date	Description
6.67.0	2026-07-16	feat: composable declarative widgets and evidence-based UI acceptance. Declarative schema v1 gains recursive group, metric, and callout composition, stable nested identity, bounded validation, shared safe fields, accessible chart fallbacks, and a native kanban Move-to path while preserving the host-owned trust boundary and all existing declarations. Visible UI completion now requires an available real consumer flow to be opened and its rendered evidence actually inspected with vision; the LLM selects states, viewports, and optional engines by risk, without a blanket mobile/WebKit matrix or a new QA service. The constitution now makes autonomous judgment and smallest-class-fix scope explicit, DEVELOPMENT records pragmatic SOLID and mutable external facts without adding analyzers or runtime lookup migrations, and reviewer prompts reject numeric finding pressure in favor of concrete evidenced defects. The operator review wrapper additionally probes the exact reviewer model through each candidate OpenRouter key before selecting it — limit_remaining alone is documented to pass ToS-blocked keys.
6.66.0	2026-07-16	feat: honest mutation attribution, attributed commit staging, environment-correct Python, auditable Skill Review history, and a drift-proof operator review wrapper. The host now captures a system_repo baseline (exact commit/tree + pre-existing dirty fingerprints) in the existing task result when a root task starts and a terminal candidate snapshot at outcome derivation; blockers ride into acceptance/review evidence for the LLM panels — deliberately never a structural outcome veto, and deliberately WITHOUT any lease/holder service (physical exclusion is out of scope; a future separate design). commit_reviewed stages only clean-at-baseline task-attributed candidates: explicit paths must be a subset and an empty set is GIT_NO_ATTRIBUTED_CHANGES, never whole-tree staging (managed update/release transactions keep their typed git add -A; baseline-less contexts keep the legacy contract). The four user launch surfaces resolve unversioned python/python3 once in registry pre-dispatch (reviewed skill env → executor backend → project .venv → verified agent interpreter) with traced provenance and a fail-closed unproved-interpreter path. Skill Review gains group-wide review_round + per-snapshot snapshot_attempt with exact task/chat/source provenance in the existing job/history files and a compact UI projection. Child-result dispositions consolidate into the single task-tree ledger row authority (task-result fields become read-time projections; explicit cancellation wins completion races with bounded scratch cleanup, no recovery surface; in-flight ≤v6.65 trees re-emit their dispositions once after upgrade). scripts/run_external_review.py becomes a thin SSOT wrapper over the production gate: real advisory by default, frozen detached-checkout review, OpenRouter key health-check from the named pool, and typed exit codes separating infrastructure failures from genuine review blocks; the hermetic pytest preflight substitute runs parallel non-serial with a bounded worker count. Fixes the chronic red CI light-model routing test. This operator phase intentionally advances VERSION ahead of the latest annotated tag; the final Phase 3 release publishes the closing tag.
6.65.4	2026-07-16	fix: a zero-usage provider body-error settles a confirmed $0, not a phantom reservation. The v6.64.0 physical-attempt ledger settles an HTTP-200 whose body is a top-level provider error (OpenRouter passes 429/5xx/4xx through the body) with cost=None and keeps its conservative reservation upper bound as an unresolved amount. Under a provider storm each reroute/strip/fallback is a fresh attempt, so the held bounds accumulate (7 x ~$2.4 = ~$16.7) and exhaust the finite task budget while real spend stays small — SWE-Pro tasks died budget_exhausted at ~$7.5 of a $25 cap. Such a request billed zero tokens (rejected before generation), so usage_from_response now settles it as a confirmed $0 and RELEASES the bound. The under-count protection is intact: any billed tokens (a partial stream or an unknown-price success, prompt/completion > 0) keep the normal cost path and their bound, and the opaque-SDK path still settles via mark_unresolved.
6.65.3	2026-07-16	fix: encrypted-reasoning 400s no longer kill tasks; openai/ reroutes strip again.* Two-layer transport failure isolated from live benchmark deaths: (1) since v6.49.0 the transient 429 same-model reroute PRESERVED OpenAI encrypted-reasoning items across an OpenRouter sibling-upstream switch — but they are not reliably portable there (field 2026-07, gpt-5.6-sol on 3x OpenAI + 2x Azure endpoints), so the rerouted conversation deterministically 400s (The encrypted content for item rs_... could not be ...); (2) since v6.36.0 a 400 delivered as a BODY error inside HTTP-200 bypassed the classic v6.28.0/v6.37.0 exception-path strip-and-retry entirely and was classified a permanent bad_request — terminal provider_unavailable (14/210 tasks in the v6.61.3 baseline campaign, then whole v6.65.x waves once 429 storms multiplied reroutes). Fixes: the transient reroute strips replayed reasoning for openai/* as before v6.49.0 (Anthropic/Gemini preserve and the proactive dispatch pin unchanged), and an encrypted-reasoning body-400 now gets the same one-shot strip-and-retry as the exception path on both sync/async ladders; genuine bad_request body errors keep their permanent classification.
6.65.2	2026-07-16	fix: an unbindable plan_task review_disposition is inert, never a wedge. v6.65.1 normalized only the vacuous empty-object disposition; within minutes field models adapted to the schema constraints and fabricated POPULATED placeholders ("review_fingerprint": "dummy"/"none"/"x"), re-wedging every first plan submission on the legitimate stale gate. Root rule (design reviewed with the authoring model): when NO review exists for this exact plan fingerprint, ANY supplied disposition is unbindable — it has no closing power by construction, so it is discarded before the wave launch (never rebound to a review created later in the same invocation) with a disclosed NOTE [PLAN_REVIEW_DISPOSITION_IGNORED_UNBINDABLE], and a REAL scout+reviewer wave runs. State-lookup failures remain loud PLAN_REVIEW_STATE_INVALID errors, never treated as absence. Where a bindable review EXISTS the strict validator is untouched (exact fingerprint binding, full finding coverage, immutable replay). The v6.65.1 minLength/minItems schema constraints are reverted — they shaped the garbage instead of preventing it; the omit-on-first-submission description stays.
6.65.1	2026-07-16	fix: a vacuous plan_task review_disposition means absent, never stale. Models routinely fill the optional review_disposition object with an empty default ({"review_fingerprint": "", "items": []}); v6.65.0 rejected any non-null disposition without an immediately preceding review as PLAN_REVIEW_DISPOSITION_STALE, wedging every first plan submission into a retry loop (observed on 100% of planning benchmark tasks). A semantically empty disposition has no closing power by construction, so it is now normalized to absent at the tool boundary with a disclosed NOTE in the result, and the schema documents "omit on first submission" (minLength/minItems on the disposition fields). Populated-but-wrong dispositions keep failing closed: the fingerprint-binding, finding-coverage, and stale gates protecting REVIEW_REQUIRED closure are untouched, with new tests pinning the previously uncovered populated-without-review path.
6.65.0	2026-07-15	feat: evidence-bound planning, answer-preserving delivery finalization, and transparent root review. Plan review waits one shared scout boundary, includes every ready non-empty handoff plus explicit omissions, records late arrivals as audit-only evidence, and publishes only GREEN, fingerprint-bound REVIEW_REQUIRED, or REVISE_PLAN; taskless invocations fail closed instead of creating a shared pseudo-task record. A loop-local complete answer survives service rounds through keep/replace control, while exact-hash child dispositions reject stale decisions and keep deferred work honestly best-effort. Root task_acceptance_review now stores evidence without paying for a second panel; the host reviews each candidate/evidence/fence binding once. Review actors separate transport, parse, and semantic truth through one compact Chat/Logs severity projection, and summary/reflection share one non-final subtree-cost snapshot before the terminal checkpoint records final spend. Shared review defaults are Fable-5, GPT-5.6-SOL, and Gemini-3.5-Flash in that order at high effort, with Fable-5 high for scope. This operator phase intentionally advances VERSION ahead of the latest annotated tag; the final Phase 3 release publishes the closing tag.
6.64.3	2026-07-15	fix: Required+Blocking acceptance now re-drives actionable non-clean reviews. A deliberate semantic DEGRADED verdict remains outside quorum and cannot create authoritative obligations, but its concrete recommendation now survives as the bounded correction capsule instead of being erased and prematurely finalized. A task-acceptance FAIL contributes only when it supplies the required outcome tier plus a concrete correction rail; bare vetoes become auditable abstentions rather than terminal review_failed decisions. Transport/parse no-quorum remains explicitly terminal review_degraded after its bounded actor retry, and deadline/budget/explicit-pass rails still permit honest best-effort finalization.
6.64.0	2026-07-14	feat: trustworthy project dialogue, physical-attempt cost accounting, and root acceptance review. Projects now use canonical chat records, LLM-first owner routing with explicit ambiguity, paint-acknowledged revision unread state, accessible rename/delete controls, and fenced tombstone deletion; routine heartbeat telemetry leaves owner chat while real incidents remain visible. Reusable background-consciousness live cards preserve the owner's explicit expansion and timeline position across cycle churn. An append-only physical-attempt ledger is the monetary authority for confirmed, reserved, unresolved, and unknown/unmetered spend, with conservative budget admission and a resumable legacy importer. Root-only task acceptance adds recursive quiescence, immutable evidence, independent reviewer slots with adaptive quorum and at most two physical attempts per actor, evidence-backed clean PASS, EWMA-paced improvement, and a restart-safe phase checkpoint. Planning resolves the active subject root, persists scout handoffs, carries a single evidence horizon, and ordinary Main calls use exact-route context-fit projections plus narrow prompt-cache hints. The commit gate remains one-pass, fail-closed at its configured P3 floor, and binds the exact tree, parents, VERSION, and tag.
Older releases are preserved in Git tags and GitHub releases. Older 6.x rows (including 6.64.2, 6.64.1, 6.63.0, 6.62.0, 6.61.3, 6.61.4, 6.61.0, 6.61.1, 6.60.0, 6.59.0, 6.54.4, 6.58.0, 6.57.0, 6.56.0, 6.55.0, 6.54.2, 6.54.1, 6.54.0, 6.53.4, 6.53.0 and 6.51.0), the 5.2.0 through 5.33.0-rc.6 rows, and former 4.0.0 rows are rolled off to respect the P9 changelog cap; their full bodies remain at their git tags.

---

License

MIT License

Created by Anton Razzhigaev & Andrew Kaznacheev