ArgusAgent

<div align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="docs/assets/brand/svg/argus-logo-horizontal-dark.svg"> <source media="(prefers-color-scheme: light)" srcset="docs/assets/brand/svg/argus-logo-horizontal.svg"> <img src="docs/assets/brand/svg/argus-logo-horizontal.svg" width="420" alt="Argus"> </picture> ### Persistent, reviewed autonomy for research and engineering Long-running agent work that can plan, execute, verify, pause, and continue beyond a single model turn. **Argus v0.1.1 · Source updates and packaged desktop previews are separate channels.** [](https://github.com/lbx154/Argus/stargazers) [](LICENSE) [](https://www.python.org/) [](https://arxiv.org/abs/2608.05144) [Website](https://argusbot.cn) · [Video Demo](https://www.youtube.com/watch?v=i8Qy9HCboQE) · [Technical Report · arXiv:2608.05144](https://arxiv.org/pdf/2608.05144) · [WeChat Community](#wechat-community) · **English** / [简体中文](README.zh-CN.md) `Manager` → `Planner` → `Engineer` ⇄ `Reviewer` </div> --- > [!IMPORTANT] > **Repository channels:** **[microsoft/ArgusAgent](https://github.com/microsoft/ArgusAgent)** > is the official source repository; > **[lbx154/Argus](https://github.com/lbx154/Argus)** is the development preview. > Changes reach the official repository through synchronization. Installing > source from `main` is not the same as installing a published Desktop release. ## The Driver–Harness Model A **model** is an engine: it burns compute and puts out tokens. A **harness** is the drivetrain that couples those tokens to files, shells, compilers, GPUs, and tests. The **Driver** is the seat — choosing what to do next, judging whether the last result was any good, and knowing when to stop and ask. In every other agent system that seat holds a human, which is why the work stops when they go to bed. **Argus takes the Driver's seat**, splitting it across four roles deliberately not allowed to do each other's jobs: | | Owns | May **not** | |---|---|---| | **Manager** | Stage transitions, and where an admitted lesson is kept | Perform the work it is admitting | | **Planner** | The next task, and the evidence it must produce | Move the campaign to the next stage | | **Engineer** | Implementation, research, experiments, artifacts | Declare its own work complete | | **Reviewer** | The verdict — correctness, evidence, limitations; may return `blocked` | Edit anything. It runs **read-only** | Credentials, payment, irreversible actions, and publication always stop for a human. It also improves without retraining: admitted Skills and source-linked Wiki findings are scoped `project` → `vertical` → `global` by how far they were shown to hold, and new domains ship as **verticals** against a core that does not change — 24 of them, with zero references to the authority boundary across 53,871 lines of domain code. Because the worker cannot grade its own work, nobody has to watch it: across 27 campaigns and 1,548 hours it needed a human research decision about **once every 310 hours**, at **95–99%** duty cycle. Everything else is in the **[technical report](https://arxiv.org/pdf/2608.05144)**. **Native backends:** `GitHub Copilot CLI` · `Pi` · `OpenAI Codex CLI` · `Claude Code` · `Cursor CLI` · `OpenCode` · `Grok Build` · `Qoder` · `DeepSeek Harness` **Harbor evaluation:** Harbor Framework can invoke the complete bounded Argus Manager/Planner/Engineer/Reviewer runtime as a custom agent. See **[Harbor integration](docs/harbor.md)**. **Coding-agent plugin:** use the packaged MCP bridge and host-specific Skills without changing the core runtime. See **[Plugin quick start](docs/plugin.md)**. **Counterexample research:** use a live Counterexample Lab, an isolated Jacobian MCP bridge, and safe in-app source updates. See **[Counterexample Lab and Jacobian setup](docs/counterexample-lab-jacobian.md)**. ## WeChat community Scan the QR code to join the Argus community. Click the image to open it at full size. If the printed expiry date has passed, open an Issue and ask the maintainers for the latest code. <p align="center"> <a href="docs/assets/argus-wechat-group-2.jpg?v=5fd55d09"> <img src="docs/assets/argus-wechat-group-2.jpg?v=5fd55d09" width="360" alt="Argus WeChat Group 2 QR code"> </a> </p> <p align="center"><strong>Community Group 1 is full. Please join Group 2.</strong></p> ## Quick Install Choose the section for your operating system. Do not mix commands between platforms. All platforms need Node.js **22.12+** from [nodejs.org](https://nodejs.org/en/download) and one authenticated Agent CLI. Reuse the CLI you already work in; Argus does not require a separate account. Docker is not required for a normal Argus installation; it is only an optional prerequisite for the separate Harbor evaluation integration. > [!TIP] > **Recommended: let the Code Agent you already use install and verify Argus.** > Copy the prompt in the Agent-assisted section below. The manual commands remain > available for users who prefer to install each step themselves. | Agent CLI | Backend | Install | Authenticate | |---|---|---|---| | GitHub Copilot CLI | `copilot` | `npm install -g @github/copilot` | `copilot login` | | OpenAI Codex CLI | `codex` | `npm install -g @openai/codex@latest` | `codex login` | | Claude Code | `claude` | `npm install -g @anthropic-ai/claude-code` | Run `claude`, then `/login` | | Cursor CLI | `cursor` | `curl https://cursor.com/install -fsS \| bash` ([Windows](https://cursor.com/install?win32=true)) | `agent login` or `CURSOR_API_KEY` | | Pi | `pi` | `npm install -g --ignore-scripts @earendil-works/pi-coding-agent` | Run `pi`, then `/login` | | OpenCode | `opencode` | [Official install](https://opencode.ai/docs/) | `opencode auth login` | | Grok Build | `grok` | [Official install](https://x.ai/cli) | `grok login` | | Qoder CLI | `qoder` | `npm install -g @qoder-ai/qodercli` | `qodercli login` | | DeepSeek Harness | `dsh` | `npm install -g @deepseek-ai/dsh` | Configure `DEEPSEEK_API_KEY` or the dsh Models page | **Choose your installation channel.** The commands below install the official source repository's `main` branch directly from GitHub, not from PyPI. To deliberately install the development preview, replace `microsoft/ArgusAgent` with `lbx154/Argus` in your platform's install and update commands. Keep the same channel when updating an existing installation. For a Windows EXE, use **[Windows Desktop](docs/windows-desktop.md)**. Official release assets, when available, are under [microsoft/ArgusAgent Releases](https://github.com/microsoft/ArgusAgent/releases); packaged previews are under [lbx154/Argus Releases](https://github.com/lbx154/Argus/releases). Source fixes do not update an already published EXE. ### Recommended: Agent-assisted installation Send this prompt to an already installed Code Agent: ```text Read https://github.com/microsoft/ArgusAgent/blob/main/docs/agent-install.md and install the official source unless I explicitly request the development preview. Keep an existing installation's channel when updating. Use the section for this operating system. Prefer the Agent CLI running this conversation as the Argus backend. Do not create a venv on Windows or macOS; keep the documented venv on Linux. Run setup through its real Agent-turn smoke test, then run `argus doctor --deep --advisor auto`. Before account login, sudo, or global configuration changes, explain why and wait for approval. Never ask me to paste a password, token, or API key into the conversation. ``` The agent follows the **[installation execution contract](docs/agent-install.md)**. ### Windows 10/11 — direct pip, no virtual environment Install Python 3.11+ from [python.org](https://www.python.org/downloads/windows/) and select **Add Python to PATH** in the installer. Then open a new PowerShell: ```powershell py --version node --version py -m pip install --upgrade pip py -m pip install --upgrade --force-reinstall "argus-skill @ https://github.com/microsoft/ArgusAgent/archive/refs/heads/main.zip" $Scripts = py -c "import sysconfig; print(sysconfig.get_path('scripts'))" $Argus = Join-Path $Scripts "argus.exe" if (-not (Test-Path $Argus)) { throw "Argus entry point not found at $Argus" } $env:Path = "$Scripts;$env:Path" & $Argus --version & $Argus --setup & $Argus doctor --deep --advisor auto & $Argus --status & $Argus ``` Calling `$Argus` proves setup is not accidentally using another stale installation. `$env:Path` also makes plain `argus` available in the current PowerShell. The troubleshooting section covers persistent PATH repair. `argus doctor` is read-only by default. To explicitly authorize an installed Agent CLI to inspect and repair Argus-scoped files, configuration, runtime state, or dependencies, use `argus doctor --advisor auto` (or name a specific advisor). Use `argus doctor --advisor none --verify` for deterministic, no-model verification. An explicitly requested active repair may take several minutes because it performs a real Agent turn; it is not a quick version check. Windows currently supports installation, Manager chat, pairing, Web/TUI, terminal-scoped daemon control, and native durable subagents. On native Windows, a detached worker owns direct or supervised long commands, persists registry and log state, and uses bounded process-tree cleanup; WSL2 remains optional rather than required for this path. The Windows Desktop installer is documented separately in **[Windows Desktop](docs/windows-desktop.md)**. ### macOS — managed command install, no manual virtual environment Install [uv](https://docs.astral.sh/uv/getting-started/installation/) if needed, then: ```bash uv --version node --version uv tool install --force --python 3.12 \ "argus-skill @ https://github.com/microsoft/ArgusAgent/archive/refs/heads/main.zip" ARGUS_BIN="$(uv tool dir --bin)/argus" test -x "$ARGUS_BIN" "$ARGUS_BIN" --version uv tool update-shell "$ARGUS_BIN" --setup "$ARGUS_BIN" doctor --deep --advisor auto "$ARGUS_BIN" --status "$ARGUS_BIN" ``` `ARGUS_BIN` works immediately even when uv's tool directory was not previously on PATH. `uv tool update-shell` makes plain `argus` available in a new terminal. `uv tool` already owns the isolated environment; do not create another venv. ### Linux — isolated source venv Linux servers keep an explicit venv so Python, CUDA tooling, and long-running process ownership remain reproducible. Install Python 3.11+, Git, Node.js 22.12+, and your distribution's `python3-venv` package first: ```bash git clone https://github.com/microsoft/ArgusAgent.git "$HOME/Argus" cd "$HOME/Argus" python3 -m venv .venv .venv/bin/python -m pip install --upgrade pip .venv/bin/python -m pip install -e . ARGUS_BIN="$HOME/Argus/.venv/bin/argus" "$ARGUS_BIN" --version "$ARGUS_BIN" --setup "$ARGUS_BIN" doctor --deep --advisor auto "$ARGUS_BIN" --status "$ARGUS_BIN" ``` Private-preview collaborators use `https://github.com/lbx154/argus-skill.git` in the Linux clone command. On Windows/macOS, install a private wheel or authenticated private archive rather than putting a GitHub token in shell history. Do not rely on a globally installed `argus` on Linux. In a new shell, use `$HOME/Argus/.venv/bin/argus` (or activate that venv explicitly). If venv creation reports that `ensurepip` is unavailable, install the distribution's `python3-venv` package and rerun the command. ### Backend notes Use `copilot`, `pi`, `codex`, `claude`, `cursor`, `opencode`, `grok`, `qoder`, or `dsh` for `--backend`. Setup adopts a model from the selected CLI's own catalog when one is available; otherwise it keeps that CLI's native default. It does not inject an OpenAI model id into Claude Code, Cursor CLI, Pi, OpenCode, Grok, Qoder, or dsh. If you have an OpenAI-compatible endpoint, setup installs Pi when needed and configures it directly: ```bash ARGUS_SETUP_API_KEY=... argus --setup --non-interactive \ --api-url https://api.example.com/v1 \ --api-model model-id ``` For Grok Build, install and authenticate the official xAI CLI first: ```bash curl -fsSL https://x.ai/cli/install.sh | bash grok login argus --setup --non-interactive --backend grok ``` `XAI_API_KEY` is also supported for headless environments. Argus uses Grok's native headless JSON stream, resumes sessions by ID, and keeps role prompts out of process arguments. In PowerShell, use a backtick instead of `\` for line continuation. #### Choosing a provider on the multi-provider CLIs Pi and OpenCode are provider-agnostic fronts: which account they bill depends on what you authenticated them against (a native DeepSeek key, Anthropic, Azure, a local vLLM, a Copilot proxy). Argus passes your configured model id straight through, so a bare id like `deepseek-chat` is resolved by the CLI itself. Name the provider when a bare id is ambiguous or when the CLI requires it: ```bash # Pi — only needed when two authenticated catalogs carry the same model id export ARGUS_SKILL_PI_PROVIDER=deepseek # OpenCode — required: `opencode run --model` only accepts provider/id export ARGUS_SKILL_OPENCODE_PROVIDER=deepseek ``` Both are also settable from the cockpit `/config` view, and persist across restarts once set there. `argus --doctor` reads the CLI's authenticated catalog and tells you when the configured provider is not one you hold a key for, or when a model id you selected is not on offer. Use `argus --config-help` to inspect each role's effective model and where it came from. Catalog commands are backend-specific, for example `pi --list-models`, `opencode auth list`, and `qodercli --list-models`. Full details, including the breaking change for Pi deployments that relied on the old implicit `github-copilot` prefix: **[backend providers](docs/backend-providers.md)**. ### Launch Windows and macOS can use `argus` after PATH setup. On Linux, replace `argus` below with `$HOME/Argus/.venv/bin/argus` unless the venv is active. ```bash argus ``` ```bash argus doctor # deterministic, read-only diagnostics argus doctor --advisor auto # explicitly request Agent-driven inspection and repair argus doctor --advisor none --verify # deterministic verification, no model call argus --status # inspect the current runtime ``` ## Interfaces ### Windows Desktop The Windows x64 source tree includes a Tauri/Rust host that supervises a frozen copy of the same Argus runtime and opens the existing Web cockpit—there is no separate Desktop fork of Manager, Workbench, or the WebAPI. It also provides signed update discovery and user-confirmed installation. Source setup, security boundaries, verification, and packaging commands are documented in **[Windows Desktop](docs/windows-desktop.md)**. ### Terminal cockpit ```bash argus ``` Use the terminal cockpit to talk to the Manager, follow live work, inspect state, and resume projects. Without an explicit `--port`, Argus reuses a compatible backend or selects the first available port starting at `8799` when another program or stale backend occupies it. On Windows, a plain `argus` launch also opens the Web UI; use `argus --no-open` for the terminal cockpit only. ### Web UI Start Argus and open the Web UI in your default browser: ```bash argus --web ``` Preferred address: [http://127.0.0.1:8799](http://127.0.0.1:8799); Argus advances to the next available port when needed. The Web UI follows the browser language on first launch and supports English and Simplified Chinese. Use the language button in the session sidebar to switch; the selection is saved in the browser. ```bash argus --web --web-port 8800 # use another port ``` #### Remote server over SSH On the server: ```bash argus --web ``` On your computer: ```bash ssh -L 8799:127.0.0.1:8799 user@server ``` Then open [http://127.0.0.1:8799](http://127.0.0.1:8799) locally. <details> <summary><strong>Direct LAN access</strong></summary> A non-loopback bind is always protected by a bearer token. If `ARGUS_SKILL_WEB_TOKEN` is set it is used; otherwise one is minted for that run: ```bash argus --web --web-host 0.0.0.0 --web-port 8799 ``` This prints the address other devices can reach, the token, and a QR code. Set the token yourself to keep one across restarts: ```bash export ARGUS_SKILL_WEB_TOKEN="$(python -c 'import secrets; print(secrets.token_urlsafe(32))')" ``` To serve without a token — only behind your own authenticating proxy — set `ARGUS_SKILL_WEB_ALLOW_INSECURE=1`. </details> ### From a phone Telegram, Feishu/Lark, and the web UI all work from a phone. The two chat bots dial out, so a daemon behind NAT needs no tunnel and no public URL: ```bash # Feishu / Lark — WebSocket long connection, no request URL to configure pip install 'argus-skill[feishu]' export ARGUS_SKILL_ENABLE_FEISHU=1 export ARGUS_SKILL_FEISHU_APP_ID=cli_xxx ARGUS_SKILL_FEISHU_APP_SECRET=xxx # Telegram export ARGUS_SKILL_ENABLE_TELEGRAM=1 export ARGUS_SKILL_TELEGRAM_BOT_TOKEN=... ARGUS_SKILL_TELEGRAM_CHAT_ID=... ``` Both bots serve the same commands (`/add`, `/status`, `/nudge`, `/backlog`, …). The web UI is installable to the home screen and pairs by scanning the QR code printed by `argus --web --web-host 0.0.0.0`. See **[docs/mobile.md](docs/mobile.md)** for the full setup. ## Advanced usage Argus is designed to be changed, not merely configured. ### Autonomy level The default `pragmatic` mode handles recoverable engineering choices—timeouts, failed tests, benchmark sizing, and technical routes—without interrupting you. It asks only for credentials, more spending, irreversible/outward-facing actions, or changes to an operator-owned acceptance boundary. ```bash export ARGUS_SKILL_AUTONOMY_MODE=cautious # ask on every explicit question export ARGUS_SKILL_AUTONOMY_MODE=pragmatic # default: recover technical issues export ARGUS_SKILL_AUTONOMY_MODE=autonomous # maximize reversible execution ``` The Web configuration view and `/config` expose the same setting. ### Adapt the runtime If you are an agent enthusiast, deploy Argus locally and make the complete loop fit the way you work. Tune role prompts, workflow boundaries, review policy, tools, and operating conventions; connect your own infrastructure; preserve the behavior you care about with tests. The compact constitution behind those choices is **[Argus Principles](docs/PRINCIPLES.md)**: agents judge meaning while the runtime guarantees mechanics; Core provides capability while Verticals provide policy; thought stays natural language; tokens buy information or action; exploration pursues upside while claims remain evidence-bound; failure changes strategy; programme outcomes and end-to-end evidence judge progress. One worked engineering case is **[exploration without local hill climbing](docs/exploration-without-local-hill-climbing.md)**: how report-only research, high-risk mechanism portfolios, single-run screening, and strict final claims were separated after an MI300X serving campaign exposed overly conservative incentives. For the wider set, **[what goes wrong and what we did about it](docs/failure-modes-and-fixes.md)** records six real failure modes — an experiment that measured its own token cap, world knowledge frozen at the training cut-off, settling for finishing rather than achieving something, ceremony applied before an idea had earned it, local hill climbing, and a reluctance to report a win. It also records a fix we got wrong: adding a gate to enforce ambition, when the honest answer was that defensive checks do not produce good work, they produce work that passes checks. A measured counterpart is **[system audit: six complaints, checked against the code](docs/system-audit.md)** — over-defensiveness, a rigorous verification bar, unnecessary operator questions, weak instruction following, redundancy, and schema abuse, each confirmed or refuted with counts from the tree. The follow-up **[architecture simplification plan](docs/architecture-simplification-plan.md)** separates a short direct-engineering lane from the full research team, proposes one Host-generated shared mission view, and outlines a compatibility-first Vertical package split. What follows from that audit is **[the simplification plan](docs/simplification-plan.md)**: an ordered set of deletions, a mechanical rule for sorting 2,277 exception handlers, an explicit list of what must not be removed, and the trap to avoid — replacing deleted machinery with a unified system that becomes the same mistake. ### Build your own Vertical A Vertical gives your field its own stages, Skills, datasets, tools, evidence expectations, evaluation methods, and completion criteria. Planning and review can then follow the real standards of your domain instead of a generic process. The `math` vertical is the worked example: three stages, a content-addressed evidence store, Lean-backed mechanical verification, and an explicit rule for which kind of check is allowed to settle which kind of question. See **[mathematical research](docs/research-mathematics.md)**. ### Use another agent as the outer layer GitHub Copilot, Pi, Codex, Claude Code, Cursor CLI, OpenCode, Grok Build, OpenClaw, or Hermes can be the environment from which you invoke Argus, inspect its state, operate its local CLI or Web/API surface, and continue improving the deployment. - **Native Argus backends:** GitHub Copilot CLI, Pi, Codex CLI, Claude Code, Cursor CLI, OpenCode, Grok Build, Qoder, DeepSeek Harness - **External agent operators:** OpenClaw, Hermes, or any agent that can use a shell or HTTP API For durable missions, install or adapt the portable [`argus-runtime-orchestration` Agent Skill](integrations/agent-skills/argus-runtime-orchestration/SKILL.md). It defines the two-party operator model, the active `Needs you` intervention loop, host-specific adapters, evidence boundaries, and closeout checks. Useful entry points: ```bash argus doctor argus --status argus --web ``` The most capable setup is often an Argus instance deliberately adapted to your own ambitious field and way of working. ## Update For source checkouts, pip ZIP installations, and uv-managed installations, use the installed Argus command: ```bash argus update argus --version argus doctor --advisor none --verify ``` `argus --update` and `argus -update` are equivalent aliases. The updater keeps the installation's existing source and channel, uses its package manager, and does not switch between the official and preview repositories. Source checkouts must be clean and on a branch; updates only fast-forward. If `argus` is not on PATH, use the executable from installation: `& $Argus update` in Windows PowerShell, `"$(uv tool dir --bin)/argus" update` for uv, or `"$HOME/Argus/.venv/bin/argus" update` for the Linux source checkout. Older versions do not yet include this updater. Bootstrap once with the original installation command below, keeping the original repository URL (replace `microsoft/ArgusAgent` with `lbx154/Argus` for an existing preview installation). Subsequent upgrades can use `argus update`. Windows bootstrap: ```powershell py -m pip install --upgrade --force-reinstall "argus-skill @ https://github.com/microsoft/ArgusAgent/archive/refs/heads/main.zip" $Argus = Join-Path (py -c "import sysconfig; print(sysconfig.get_path('scripts'))") "argus.exe" & $Argus --version & $Argus doctor --advisor none --verify ``` macOS bootstrap: ```bash uv tool install --force --python 3.12 \ "argus-skill @ https://github.com/microsoft/ArgusAgent/archive/refs/heads/main.zip" "$(uv tool dir --bin)/argus" --version "$(uv tool dir --bin)/argus" doctor --advisor none --verify ``` Source-checkout bootstrap (inspect local changes first): ```bash git -C "$HOME/Argus" status --short git -C "$HOME/Argus" pull --ff-only "$HOME/Argus/.venv/bin/python" -m pip install -e "$HOME/Argus" "$HOME/Argus/.venv/bin/argus" --version "$HOME/Argus/.venv/bin/argus" doctor --advisor none --verify ``` Only continue the source bootstrap when `git status --short` is empty and the branch tracks the intended repository. Argus detects stale local WebAPI and daemon processes and replaces them at a controlled task boundary. Verification with `--advisor none --verify` does not spend a model call. Packaged Desktop EXEs use the separate signed desktop update channel. The CLI updater does not replace a signed EXE; see [Windows Desktop](docs/windows-desktop.md). ## Uninstall ```powershell # Windows py -m pip uninstall argus-skill ``` ```bash # macOS uv tool uninstall argus-skill ``` On Linux, stop Argus, preserve any work you need, then remove the `$HOME/Argus` checkout and its `.venv`. Package removal intentionally leaves runtime state under `$HOME/.argus-skill` untouched on every platform; delete that directory only when you also want to remove projects, configuration, and logs. ## Installation troubleshooting - `unresolved provider cost` is an accounting hold, not a login or balance diagnosis. Argus rechecks late or partial Copilot billing, including pending records left behind by older reconciliation caches, before new calls. It settles pending token records when their recorded model and complete token counts can be priced. Unknown prices or missing usage remain blocked rather than being treated as free. Check the reported provider, model and reason in `cost-control.json` under the Argus data directory and the project's `usage.jsonl`; do not delete the ledger. Run diagnostic commands in a terminal, not in the Web chat box. - Confirm which executable the shell is using: `Get-Command argus -All` on PowerShell, or `type -a argus` on macOS/Linux. Its `argus --version` release id should change after an update. - On macOS, use `"$(uv tool dir --bin)/argus"` immediately. Run `uv tool update-shell` once and open a new terminal for plain `argus`. - On Windows, recover the exact Scripts directory with `$Scripts = py -c "import sysconfig; print(sysconfig.get_path('scripts'))"`. Add it to the current window with `$env:Path = "$Scripts;$env:Path"`. For new windows, use the Python installer’s **Modify** action and enable **Add Python to PATH** rather than creating a venv. - On Linux, use `$HOME/Argus/.venv/bin/argus`; a global `argus` may be an older installation. Install `python3-venv` if `python3 -m venv` lacks `ensurepip`. - Use `argus doctor --advisor none --verify` for deterministic diagnostics. Use `argus doctor --advisor auto` when you explicitly want an installed Agent to inspect and repair Argus directly. - Use `argus --config-help` to check the effective backend/model before blaming setup or authentication. ## What Argus has done so far A partial record, grouped by **who decides whether the result counts** — and none of those deciders is Argus. ### Open code | Repository | Result | |---|---| | **[ace-2](https://github.com/Argus-AiTeam/ace-2)** | A Qwen2.5-0.5B W4A8 inference accelerator with no human author of record for its spec, RTL, verification, or physical flow. **13,914/13,914** runtime commands; SKY130 at **0.614 mm²** (cap 2.0), **+0.6966 ns** slack, WNS/TNS 0.00 ns, 100 MHz. The certificate publishes its own exclusions: no DRC/LVS, no GDS, no silicon. | | **[minimax-h3-mac](https://github.com/Argus-AiTeam/minimax-h3-mac)** · **[-desktop](https://github.com/Argus-AiTeam/minimax-h3-desktop)** · **[ComfyUI nodes](https://github.com/Argus-AiTeam/ComfyUI-MiniMax-H3-MLX)** | A ~62 GiB model, not shrunk but run: **47 min 58.7 s** at **~15.8 GB** peak on a 24 GB M4 Pro. On one RTX A6000, Turbo 8-step reaches **6.159×** over a BF16 N=10 baseline. Candidates that failed the quality gate are published as *rejected*. | | **[FlashDA](https://github.com/SJTU-DENG-Lab/FlashDA/tree/feature/dllm-fa4-adaptation)** · **[Diffulex](https://github.com/SJTU-DENG-Lab/Diffulex)** | Six diffusion-LM mask families carried into a **FlashAttention-4 CuTe DSL** kernel in **21.85 hours** of model compute, under **80 CNY**, with 2 interruptions in 87.7 hours. **19/19** parity across SM80/SM90; on H200/SM90 it reaches **92–95% of native FA4** and **1.61–2.57×** over the Diffulex Triton backend, both sides CUDA-Graph captured. The early route was **4.9–29.6× slower** than native; recognising the data path was wrong and abandoning it is what produced the result, and that rejected route is retained with its evidence. | FlashDA builds on the excellent FlashAttention-4 / CuTe DSL work by [Tri Dao](https://github.com/tridao) and collaborators. Reproductions and ports beyond SM90 are very welcome; the full protocol and per-scenario latencies are in [`EXPERIMENT_RESULTS.md`](https://github.com/SJTU-DENG-Lab/FlashDA/blob/feature/dllm-fa4-adaptation/EXPERIMENT_RESULTS.md). ### Judged by outside maintainers | Submission | Outcome | |---|---| | **[sglang#35038](https://github.com/sgl-project/sglang/pull/35038)** — native SenseNova U1 multimodal generation and interleave serving | 36 files, **+11,263/−72**, 14 commits. 1,116 tensors, 0 missing; VQA exact **160/160**; concurrency-8 exact **8/8**; **5.108×** at BS8. One engineer with a turn-by-turn agent invested **60+ hours** without completing it; Argus finished inside **24.14 hours**. *Open.* | | **[fla-org#1045](https://github.com/fla-org/flash-linear-attention/pull/1045)** — TileLang RWKV6 backend | **Merged**, 1.21× fwd+bwd on H100 NVL, no inline change requests. Its description states the work was completed autonomously — and an outside maintainer accepted that sentence with the code. | | **[fla-org#1109](https://github.com/fla-org/flash-linear-attention/pull/1109)** — SM100 autotune crash | **Merged.** Two lines, no speedup: before the fix the test file could not finish; after it, **76 tests pass**. | | **[fla-org#1128](https://github.com/fla-org/flash-linear-attention/pull/1128)** · **[#1114](https://github.com/fla-org/flash-linear-attention/pull/1114)** | KDA training 1.29× over Triton, and `AttnRes` 1.102× geomean on B200. Both report their *worst* row alongside the mean, and #1128 shipped the 1.078–1.099× number it could verify rather than the 1.541× it measured. *Open.* | ### Scored by official harnesses | Arena | Result | |---|---| | SWE-Bench Pro (731 tasks) | **≈78%** vs **59%** for direct Copilot, same model both sides — and **35** tasks declared `blocked` rather than reported as unsupported successes | | SOL-ExecBench | Rank **#6 globally**; 7 kernels top-3; beat the #1 entrant on 2 | | MLE-Bench Lite | **69.2%** medal rate (9/13): 3 gold, 3 silver, 3 bronze, against Kaggle leaderboards | | AARRI-Bench | **63/82 (76.8%)** vs 68.3% paper best | | nanochat / nanoGPT speedrun | 0.9636 vs 0.9646 BPB on B200; **79.77 s** vs 80.18 s same-device human record | ### Decided by external checkers - **MOF generation** — chemical control 92.5 / 100.0 / 74.5%, AUC 0.594 → 0.833, verified by external `MOFChecker`. The admitted method is *smaller* than the one it replaces. - **Erdős–Gyárfás** — six proof-backed frontier updates, with one falsified route retained as evidence rather than deleted. - **Research writing** — six paper pipelines to submission across 254 missions, including 16 Stage rollbacks under review. - **On itself** — at maturity, **21% fewer tokens** and **15% less active time** per solved SWE-Bench Pro task than at startup, with weights unchanged. Longest single campaign: **8.1 days**. > [!NOTE] > Every number is reproduced from the [technical report](https://arxiv.org/pdf/2608.05144) > or the linked repository, and carries the scope conditions stated there.