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. 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-5.5-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	openai/gpt-5.5	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
6.52.0	2026-06-29	feat: GAIA-derived capability hardening + TB2.1 verification hygiene. P1 first-class attachment access — every task's attachments (CLI/API, GAIA, desktop chat) stage into one agent-readable root (artifact_store/attachments/) surfaced as a ready read_file(root='artifact_store', path='attachments/...') manifest + auto native-image blocks; staged INPUTS are excluded from deliverables; tool_access relaxes user_files so benign project dotdirs (.github/.vscode/.idea) are readable while all other in-home dotfiles/dirs (credentials, shell history, VCS internals) stay blocked by default (DEFAULT-DENY dotted components + a small benign allowlist), and workspace_task may READ user_files. P4a fixes a general image↔tool_result ordering bug at the single send-time chokepoint (_normalize_system_message_placement, all providers). P4b adds ocr_pdf (text-layer) + youtube_transcript media tools. P2 strengthens the finalize nudge so a done-but-unmarked answer is reliably marked (FINAL ANSWER:); no extractor change. C adds an after-only artifact-transparency FLAG to verify_and_record (a check that built then deleted a declared deliverable is surfaced to the ADVISORY reviewer; status stays pass). GAIA adapter: agent-visible deadline (strictly < the hard kill), shared format prompt, attachment staging + METHODOLOGY honesty notes. D: run_tb 3-tier low-k report_grade.
6.51.1	2026-06-28	fix(v6.51.1): CI fix-forward of v6.51.0 — align 3 stale tests + defensive volume_suffix. v6.51.0's full-test matrix failed only on 3 pre-existing tests not updated for the v6.51.0 changes (no runtime regression): test_review_substrate_v2's two acceptance-diff tests now patch collect_turn_diff with the include_recent_commit kwarg and assert the demoted agent diff under the (now-redacted) agent_supplied block (relocated from a top-level key in v6.51.0); test_swe_pro_e1v2 passthrough no longer breaks on the consolidated run_pro reading args.volume_suffix — run_instance now reads it defensively via getattr.
6.51.0	2026-06-28	feat: SWE-Pro forensics-driven verification & finalize-grounding hardening. Runtime: verify_and_record recovers a stringified-argv check through a shared SSOT (shell_parse.recover_stringified_argv / normalize_check_argv) used by BOTH run_command and the shell guard — so the guard inspects EXACTLY what executes — and runs a string check via a non-login sh -c that inherits the bootstrapped PATH (ending the [go,: not found exit-127 and the login-shell go: not found divergences). A one-shot ADVISORY finalize nudge fires when the agent's latest host-attested verification is RED and unreconciled (outcomes.latest_unreconciled_failed_receipt — reconciled only by a later pass/observed, never a declared escape hatch): finalizing over your own red is a self-contradiction (P3/P12). The task-acceptance reviewer is now process-aware (review_evidence.build_task_acceptance_evidence): full task contract + a first-class verification_summary (RED surfaced) + a bounded, redacted tool-call trajectory + a leak-safe artifact manifest + provenance tags, under a disclosed-truncation budget — so it critiques HOW the task was solved (wrong tool / wrong direction / ignored-own-red), staying ADVISORY with NO new blocking semantics. Bench/devtools: prompt_baseline.txt discloses that local green is not the acceptance oracle, demands interface-exactness, and bans grep-proxy grounding; a host zstd image cache (OBO_SWEPRO_IMG_CACHE, atomic + fail-soft) avoids re-pulling images on re-runs; METHODOLOGY.md gains a Benchmark Legitimacy & Anti-Cheat section (scaffold allowed; git-history-mining and gold-test-seeding forbidden, with sources). Adversarial false-completion tests included.
6.50.4	2026-06-28	fix(llm): stop sending Anthropic empty text blocks / cache_control-on-empty. An empty tool output was sealed/sent as {"type":"text","text":"","cache_control":{...}}, causing Anthropic HTTP 400 cache_control cannot be set for empty text blocks (classified non-retryable bad_request) → the reroute replayed the same malformed transcript → the whole task died as provider_unavailable with an empty answer. Fixes at the source (seal_task_transcript placeholders an empty cache anchor), at the Anthropic tool-result construction (_sanitize_anthropic_tool_result_content drops empty text blocks, preserves image/document blocks, and substitutes a placeholder only when the result is otherwise empty — covering scalar "" and list []), and a content-preserving guard in the shared cache-policy chokepoint (never attach cache_control to an empty text block). The direct-lane normalizer now surfaces Anthropic stop_reason so a successful response is not misread as a finish_reason=null glitch. Regression tests cover all four surfaces; smoke-verified against the live API (control reproduces the 400, the converted empty-tool-result transcript returns 200).
6.50.3	2026-06-28	feat: TB2.1 capability + swarm-affordance hardening. Verification: verify_and_record gains an expected_match mode (substring default · exact · exact_line · json_equals) recorded in the host-attested receipt, and the task-acceptance reviewer now demands metric-grounded evidence — existence/substring-only is insufficient, with a public-information-only anti-cheat boundary. Telemetry honesty: an unrecovered access-policy block (resource_policy_blocked / resource_constraint_blocked) on a READ-ONLY exploratory tool is demoted to an ignored_tool_errors axis instead of degrading the turn. Swarm affordances: a per-task swarm_efficiency rollup (subagent_count / wave_count / Σ inter-wave latency / lanes_used) for fan-out tasks; schedule_subagent surfaces effective lane(s); wait_task emits a burst/absorb advisory when other children are still in flight; the descriptions steer burst+absorb and cooperative-multi-builder (external_workspace, omit write_root) versus genesis. UX: a one-shot no-op-attempt finalization nudge (declared expected_output + zero effects + no FINAL ANSWER) and a preserve-your-work rule.
6.50.2	2026-06-27	fix(scope-review): preserve blocking scope-review authority while handling false-window overflow. The scope reviewer window probe now uses the actual scope-review slot route instead of the active main route, recognizes direct-provider and explicit OpenRouter spellings of the shipped openai/gpt-5.5 reviewer, and only downgrades opaque gateway 400s to budget_exceeded when independent size evidence points to real oversize. Off-default reviewers without Capability Evidence now surface advisory-only under the 1M floor until owner-acked, instead of inheriting false 1M trust. Regression tests cover direct/default reviewer identity, local-main plus remote-scope routing, and large non-oversize provider errors staying fail-closed.
6.50.1	2026-06-27	docs(prompt): sync resident SYSTEM.md with the v6.50.0 reconciliation layer. The system prompt now points agents at the live schedule_subagent schema for capability declarations, names structured delegation_constraint back-pressure and parent-only override_delegation_constraint, and calls out the runtime filesystem affordance map without duplicating schema-only directive details. No release tag, build, or stable promotion was produced for this patch-only branch update.
6.50.0	2026-06-27	feat: reconciliation layer for reliable subagent swarms plus SWE-bench Pro adapter hardening. Runtime: the agent now reasons forward from a live filesystem affordance map, reconciles declared child capabilities and delegation constraints before scheduling, queues excess children instead of dropping them at the active-subagent cap, prevents clean finalization while delegated children are still running, and fixes a light-mode shell false positive by classifying interpreter writes by resolved write targets rather than repo-path mentions. Bench/devtools: the SWE-bench Pro adapter hardens Alpine/musl install-in-image transport (no Python ABI drift, pyexpat/pip/import fail-fast, unsupported Playwright dropped, no read-only editable install), captures patches without pre-existing untracked task-image fixtures, and distinguishes permanent infra non-runs from retryable transient failures. Historical note: v6.48.0/v6.49.0 were pushed without annotated tags; v6.50.0 resumes the release tag stream by operator decision.
6.49.0	2026-06-27	feat: generalize same-model provider failover to all verified-portable reasoning families (OpenRouter rate-limit resilience). Extends v6.48.0 beyond Anthropic: the allow_fallbacks=false reasoning-continuity pin is now lifted for every family whose replayed reasoning signature is cross-provider portable on OpenRouter — Anthropic (Anthropic/Bedrock/Vertex/Azure), Gemini (Vertex/AI-Studio), and OpenAI encrypted reasoning (OpenAI/Azure), each verified live by a same-model replay probe forcing each sibling provider. The prod MAIN (gpt-5.5) and FALLBACK (gemini) now fail over to a healthy sibling endpoint under an upstream rate-limit instead of surfacing it, while OpenRouter's sticky routing keeps the prompt cache warm on the primary (no cache thrash, no explicit provider ordering needed). Unverified/exotic families (e.g. z-ai/glm, deepseek) keep the conservative pin; the reactive signature-rejection 400 strip-and-retry remains the safety net for every family. New SSOT predicate _reasoning_signature_portable_across_or_providers.
6.48.0	2026-06-27	feat: Anthropic same-model provider failover keeps reasoning continuity (OpenRouter resilience). A replayed-reasoning OpenRouter request no longer pins provider.allow_fallbacks=false for Anthropic-family models: Anthropic thinking-block signatures are cross-platform compatible (Anthropic API / Bedrock / Vertex / Azure — confirmed against Anthropic's extended-thinking docs and a live same-model OpenRouter replay probe), so a same-model failover preserves reasoning continuity and no longer surfaces one upstream's rate-limit when a healthy sibling endpoint could serve the turn. Non-Anthropic families (e.g. OpenAI / Gemini encrypted reasoning) keep the conservative pin, and the reactive signature-rejection 400 strip-and-retry stays the safety net for every family.
Older releases are preserved in Git tags and GitHub releases. Older 6.x rows, 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