ASI Code

A modern terminal coding agent in Rust. Single binary. No Node.js runtime required, and no npm install step. Runs against DeepSeek, OpenAI, or Claude with the same workflow.

ASI Code 2.0 is coming soon. The 2.0 line moves beyond pure code editing into desktop and professional-software control: native GUI automation (screenshot / window targeting / synthetic input / on-screen OCR), live engine bridges for Unreal Engine 5, Blender, and Unity (drive the already-open Editor via [InitializeOnLoad] C# drops), 3D-aware scene probes plus probe_diff for deterministic structural verification, frame-folder video capture with keyframe selection for temporal context, and a smarter auto-loop that no longer mistakes successful read-only investigation for "no progress." See What's New in 2.0 below.

Inspired by modern coding-agent CLIs, with a similar workflow and UX: REPL with slash commands, streaming output, auto-tool loop, work / code / secure / review modes, sub-agents, MCP server support, plugin system, agent skills, cron jobs, git worktrees, sandbox + audit log, and a 60+ command surface.

# 30-second quickstart (Windows / PowerShell)
# DeepSeek
$env:DEEPSEEK_API_KEY = "sk-..."
cargo run --release -- repl --provider deepseek --model deepseek-v4-pro --project D:\your-project --no-setup

# OpenAI/Claude native tool_use
# OpenAI
$env:OPENAI_API_KEY = "sk-..."
cargo run --release -- repl --provider openai --model <openai-model> --project D:\your-project --no-setup

# Claude
$env:ANTHROPIC_API_KEY = "sk-ant-..."
cargo run --release -- repl --provider claude --model <claude-model> --project D:\your-project --no-setup

ASI Code is a Rust-built terminal coding agent with a workflow similar to modern coding-agent CLIs.

This project is independent and is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, Anysphere, or any other model/tool provider.

Implemented Features

Interactive REPL with slash commands and streaming output
REPL startup provider wizard (choose provider, API key, model before chat)
API key safety: wizard keys are session-only by default unless explicitly persisted
Provider support: OpenAI, DeepSeek, Claude (API key + Claude OAuth token)
Tool system: bash, read_file, write_file, edit_file, glob_search, grep_search
Compact read display: read_file defaults to chunked ranges and UI shows a concise read summary instead of dumping full file content
Web tools: web_search, web_fetch
Auto tool loop: parses assistant /toolcall ... lines and executes them (toggle with /auto on|off)
Work mode: /work <task> or /code <task> injects workspace snapshot and drives tool-first project editing
Intent-aware coding mode: non-command prompts that look like coding requests auto-enter workspace coding flow (simple greetings do not)
Security fix mode: /secure <task> focuses on vulnerability and hardening fixes
Review mode: /review <task> focuses on bugs, regressions, risks, and missing tests
Session persistence: save, list, resume
Change tracking: per-turn changed_file=... feedback and /changes for session summary (/changes clear to reset)
Todo tracking: add/list/done/remove
Project memory notes and notebook markdown cells
Git integration command wrapper
MCP server management (add/rm/auth/oauth/config/show/start/list/stop)
Subagent tool-loop closure: /agent runs in an isolated runtime and can continue through tool-call loops (toggle: ASI_SUBAGENT_TOOL_LOOP)
Cost tracking and status line (tokens + USD estimate)
Markdown ANSI rendering toggle
Model aliases: opus, sonnet, haiku
Privacy and safety controls:
- local telemetry JSONL (disabled by default)
- tool-input telemetry redaction by default
- safe shell guard for dangerous commands
- undercover mode for sanitized commit messages
- auto-review guard layer: off|warn|block with severity threshold critical|high|medium|low
- feature killswitches (web_tools, bash_tool, subagent, research)
- explicit remote policy sync (/policy sync) with safe-mode guardrails
Theme setup flow:
- theme / /theme text style chooser with preview (7 themes: dark / light / colorblind variants / ANSI variants / cubist Picasso-inspired truecolor)
- setup / /setup API key + model + wallet setup wizard
- scan / /scan repository pattern scan panel (auto-detects project profile: Rust/JavaScript/Python when no patterns are provided, supports deep mode: --deep or --deep=N)
Context auto-compaction: triggers automatically before each turn when cumulative input tokens cross ASI_AUTO_COMPACT_INPUT_TOKENS (default 120000) OR live turn count crosses ASI_AUTO_COMPACT_TURN_RATIO * max_turns (default 0.80), so long sessions never silently dead-end on max_turns_reached
Test feedback loop (opt-in via ASI_AUTO_TEST=1): after every code-changing iteration the auto-tool-loop runs the project's test command (cargo test for Rust, pytest for Python with a pytest config marker, npm test for Node with a scripts.test) and on failure injects the trace as a synthetic user message so the model fixes the cause — capped at ASI_AUTO_TEST_MAX_ATTEMPTS consecutive self-heal rounds (default 3)
Auto-loop journal + resume (default ON, throttled to one save per ASI_AUTO_JOURNAL_INTERVAL_SECS, default 30s): every auto-tool-loop iteration writes a checkpoint with messages + changed_files + change_events + the original task input; if the prior session crashed mid-loop, REPL startup detects loop_running=true and prompts Resume? [Y/n]. Manual trigger: /checkpoint resume. Disable with ASI_AUTO_JOURNAL=0

What's New in 2.0 (Preview)

Everything below is already implemented and merged on main; the 2.0 release is the moment we cut a numbered tag and ship signed installers. Try it from source today.

Tier 1 — Native computer-control tools

A /computer tool family lets the agent see and drive any Windows GUI app, not just the terminal. Five tools, all available as native tool_use / tool_calls blocks and via /toolcall:

screenshot — capture the screen, return base64 PNG for vision models
find_window <title> — locate windows by title substring, returns pid + MainWindowHandle so other tools can target them precisely
click <x> <y> / click_text "Save" — UI Automation pointer + button click via the accessibility tree
type_text — paste arbitrary text into a target window (see below for the focus fix)
read_screen_text — full-screen OCR via Tesseract, returns extracted text

type_text was rebuilt around two real-world bugs the naive SendKeys approach silently hides:

Focus: SendKeys::SendWait posts to the foreground window — without an explicit focus grab, keystrokes hit the terminal running asi-code, not the app the user wanted. type_text now accepts optional window_pid (preferred, get it from find_window) or window_title, calls SetForegroundWindow + ShowWindow(SW_RESTORE) with a 150 ms settle delay, then types.
Whitespace and Unicode: Windows 11 Notepad (WinAppSDK) silently drops space characters from rapid SendKeys input, and SendKeys's special-character escapes (+ ^ % ~ ( ) { }) break naive payloads. type_text now writes through the clipboard (Set-Clipboard + Ctrl+V) and restores the user's prior clipboard contents — arbitrary text round-trips byte-perfect, including CJK and emoji.

Tier 2 — Live engine bridges for professional 3D/game software

Three bridges turn the CLI into a remote control for the heavy-iron creative tools that the generic computer-control approach can't drive intelligently:

Tool	Default driver	What it does
`ue5_bridge`	`UnrealEditor-Cmd.exe -ExecutePythonScript=<file>`	Runs Python in a headless Unreal Editor session for the given `.uproject`.
`blender_bridge`	`blender --background --python <file>`	Runs a Python script in a headless Blender; override the binary via `ASI_BLENDER_BIN` (needed for Microsoft Store installs).
`unity_bridge`	dispatched by `action` field	Five modes — see below.

unity_bridge is the most developed of the three. Pass action to pick a mode:

python (default) — legacy: spawn Unity -batchmode -executeMethod PythonRunner.RunFile, fresh headless instance per call.
open — launch the Editor for the project and return immediately. Use this before csharp / create_terrain to make sure the live Editor is running. Unity Hub focuses an existing instance instead of duplicating.
csharp — execute C# inside the already-running Editor. The bridge drops a one-shot Assets/Editor/AsiOneshot_<NONCE>.cs with an [InitializeOnLoad] hook, lets Unity's own compiler pick it up, polls a unique result file in %TEMP% for OK\n… or ERR\n…, then the script self-deletes via EditorApplication.delayCall. This is how you make interactive edits to the open scene without an extra IPC daemon — Unity's asset pipeline is the IPC channel.
create_terrain — convenience: generate the C# that spawns a Terrain GameObject in the active scene (size, heightmap resolution, name all configurable), dispatch via csharp.
save — SaveOpenScenes() + SaveAssets() + Refresh(), also via csharp.

Example:

/toolcall unity_bridge {"project":"D:/unity/My project","action":"open"}
/toolcall unity_bridge {"project":"D:/unity/My project","action":"create_terrain","name":"ParkourTerrain","size":[500,50,500]}
/toolcall unity_bridge {"project":"D:/unity/My project","action":"csharp","csharp":"var c = GameObject.CreatePrimitive(PrimitiveType.Cube); c.transform.position = new Vector3(10,5,10);"}
/toolcall unity_bridge {"project":"D:/unity/My project","action":"save"}

Bridge environment variables:

ASI_UE5_EDITOR / UE5_EDITOR — Unreal Editor binary
ASI_BLENDER_BIN / BLENDER_BIN — Blender binary (set to the full path for Microsoft Store installs, e.g. C:\Program Files\WindowsApps\BlenderFoundation.Blender_*_x64_*\Blender\blender.exe)
ASI_UNITY_EDITOR / UNITY_EDITOR — Unity binary
ASI_BRIDGE_TIMEOUT_SECS — wait budget for batch calls and for the csharp result-file poll (default 120 s)

Tier 3 — 3D-aware probes and temporal capture

The bridge layer can drive engines, but "drive" alone is not enough — the model also needs to see what is in the scene before it acts, and verify what changed after. Frames of a viewport are interpretation; structured scene data is truth. Tier 3 adds both.

Six new tools, all exposed as native tool_use / tool_calls and /toolcall entries:

Tool	Purpose
`blender_scene_probe`	Dump active Blender scene (objects, transforms, cameras, lights, actions, materials) as JSON via `bpy`.
`ue5_scene_probe`	Dump UE5 level state (actor list, transforms, materials, light/camera params) as JSON via Unreal Python.
`unity_scene_probe`	Dump active Unity scene as JSON by executing C# in the live Editor (same `[InitializeOnLoad]` drop pattern as `unity_bridge action=csharp`).
`probe_diff`	Deterministic structural diff between two probe payloads. Returns `added` / `removed` / `modified` (with the specific fields that changed) instead of a free-text "looks different."
`video_capture`	Capture a frame sequence to a directory (`fps=2`, `duration_sec=6` defaults). v1 deliberately skips MP4 encoding — frames-on-disk is what the model actually consumes, and removes an external dependency.
`video_keyframes`	Pick a compact subset from a frame directory (default `max_frames=12`). Selection mixes boundary frames, change-magnitude ranking, and equidistant fill — so the model gets the most informative frames, not random samples.

The system prompt steers the agent to probes first, video second: blender_scene_probe / ue5_scene_probe / unity_scene_probe are deterministic ground truth, while video_capture + video_keyframes are visual evidence used only when the structural data is not enough on its own.

Example loop — make a Unity edit, verify it structurally, capture supporting frames only on uncertainty:

/toolcall unity_scene_probe {"project":"D:/unity/My project"}                # -> before.json
/toolcall unity_bridge {"project":"D:/unity/My project","action":"csharp","csharp":"GameObject.CreatePrimitive(PrimitiveType.Cube).transform.position = new Vector3(0,1,0);"}
/toolcall unity_scene_probe {"project":"D:/unity/My project"}                # -> after.json
/toolcall probe_diff {"before":{...},"after":{...}}                          # -> {added:[Cube], modified:[]}
/toolcall video_capture {"duration_sec":4,"fps":2,"target":"window","window_title":"Unity"}
/toolcall video_keyframes {"dir":"<frame dir from previous step>","max_frames":8}

Why frames-on-disk instead of MP4: vision-capable LLMs ingest images, not video. Encoding to MP4 just to extract PNGs again is wasted work; the frame-folder hands the keyframe selector exactly what it needs and avoids pulling in ffmpeg as a runtime dependency.

Auto-loop no-progress detector — read-only work now counts

The auto-loop's "no progress across too many rounds" circuit breaker previously only counted file mutations as progress. Read-heavy tasks (screenshot, find_window, bash inspection, web_search, OCR, computer-control sequences) would trip the breaker mid-investigation even while the model was making real headway.

The detector now classifies each round as one of FilesChanged / ReadOnlyToolSuccess / Nothing, and only the third increments the no-progress counter. The tunable cap (ASI_AUTO_AGENT_MAX_NO_PROGRESS_ROUNDS, default 12) is unchanged; what changed is what counts.

Cubist theme — Picasso-inspired terminal UI

A 7th /theme option, cubist, gives the REPL a high-contrast Picasso-style look that goes beyond ANSI 16-color palettes. It is opt-in — pick 7 in /theme and the existing six themes stay byte-identical.

What changes when theme=cubist is active:

Truecolor palette — 24-bit ANSI (\x1b[38;2;R;G;B) cobalt / ochre / sage / terra cotta / slate / ivory replaces the basic 4-bit colors.
Asymmetric welcome mosaic — two side-by-side block-drawn panels with diagonal ╱╲ cuts and a deconstructed ◆ A ◇ S ◢ I ▲ ◆ C ◇ o ◢ d ▲ e ▽ letter banner instead of the rounded box.
Geometric tool panel — ◢◣ TOOL <name> (<status>) ◣◢ head with a ▰▱ block-fragment border, replacing the ─ line.
Animated spinner — a real tick-driven frame loop cycling through ◆ ◇ ▲ △ ▰ ▱ at 200 ms intervals (a thread is spawned only in cubist; other themes keep the static glyph).
Deconstructed progress bar — ◇ progress ▰▰▰▱▱▱▱▱▱▱ 3/10 rendered after each tool call in the auto-loop, parallel and sequential branches alike.
Streaming <toolcall> interception — when the model emits raw <toolcall>…</toolcall> (or <tool_call> / <tool_calls>) XML mid-stream, the line buffer suppresses it and emits a single ◢◣ INTENT ▸ <tool name> panel instead of dumping the XML to the terminal.
Hidden-content notice — the read_file compact body becomes ◢ [read_file:… lines …] / ◇ content hidden ◇ use /toolcall read_file … instead of the previous parenthetical.
Diff highlighting in edit_file / write_file panels — + / - / @@ / [section] lines inside the tool body are colored with the cubist diff palette (sage on dark for added, ivory on terra-cotta for removed).

Pick the theme: /theme → 7. Persisted in AppConfig.theme. Truecolor requires a modern terminal (Windows Terminal, VS Code terminal, iTerm2, mintty); legacy conhost falls back gracefully on the other six themes.

Auto-compact — never hit `max_turns_reached` again

The runtime tracks both cumulative input tokens and live turn count. Earlier the only auto-trigger was a token threshold (ASI_AUTO_COMPACT_INPUT_TOKENS, default 120000) and it only fired inside the auto-tool-loop. A long auto-loop could rack up 200 turns at 40 K tokens, hit max_turns_reached, and then every subsequent REPL input would fail immediately because turn-count is the gate.

Two changes close this gap:

Turn-ratio trigger — Runtime::auto_compact_recommendation_by_turns(ratio) returns Some(reason) when turn_count() >= ceil(max_turns * ratio). The default ratio is 0.80 (so at max_turns=200, compaction fires at 160). Tune with ASI_AUTO_COMPACT_TURN_RATIO=0.0 to disable, or any (0.0, 1.0] value.
Pre-turn REPL guard — before each run_turn_streaming call, the REPL checks both the token threshold and the turn ratio. If either fires, it runs compact_with_summary(8) first, prints the result, and proceeds. So even if you walked away from a session that hit max_turns_reached, the next prompt auto-compacts and continues normally instead of returning the error immediately.

The auto-tool-loop also picks up the turn-ratio trigger alongside the existing token trigger, so long-running fan-outs (50+ tool calls) compact mid-run instead of waiting for either limit to land.

Test feedback loop — auto-run tests, auto-heal on failure

Default-OFF, opt-in via ASI_AUTO_TEST=1. Once enabled, every iteration of the auto-tool-loop that touches code (write_file / edit_file / /toolcall write_file …) is followed by a real test run. The detector classifies the changed files by suffix and only fires the matching language's test command:

Language trigger	Marker required	Command emitted
`.rs` / `Cargo.toml` / `Cargo.lock`	`Cargo.toml` in CWD	`cargo test --offline --quiet --no-fail-fast`
`.py`	`pytest.ini` OR `pyproject.toml` with `[tool.pytest…]` OR `setup.cfg` with `[tool:pytest]`	`pytest -q --no-header`
`.ts` / `.tsx` / `.js` / `.jsx` / `.mjs` / `.cjs`	`package.json` with non-empty `scripts.test`	`npm test --silent`

On failure, the loop:

Renders an auto_test (error) tool panel with the clipped (6000 char) trace.
Pushes a synthetic user-role message [auto_test:error] One or more tests failed after your last code change. Diagnose the root cause from the trace below and fix it. Do not skip or weaken tests. followed by the command and trace into runtime history.
Lets the next loop iteration run normally — the model now has the failure in context and produces a fix.

Self-heal is capped by ASI_AUTO_TEST_MAX_ATTEMPTS (default 3). After that many consecutive failures the loop exits with stop_reason=auto_test_max_attempts_reached instead of looping forever. Outside the auto-tool-loop (manual /toolcall write_file etc.) the same test command runs at the end of the turn for visibility, but no self-heal injection happens.

Project-marker-aware: a Rust-only repo never runs pytest, a Python-only repo never runs cargo test. Multi-language repos run only the tests for languages whose source actually changed this turn. Pure docs / markdown edits never trigger tests.

Auto-loop journal + resume — survive crashes mid-task

Long auto-tool-loops (hours of agent work) used to be all-or-nothing: a single crash, network blip, or laptop sleep mid-run discarded every prior tool call's spend. The journal layer turns the existing checkpoint into a per-iteration journal, so the next launch can resume from the last completed iteration with zero replay.

Default ON; disable with ASI_AUTO_JOURNAL=0. What changes:

Per-iteration save (throttled): inside run_tool_loop_core, after every iteration finishes, the runtime writes messages + changed_files + change_events + original_task_input + iteration counter into sessions/_checkpoint_latest.json under a new meta.in_loop_state block. Throttled to one save per ASI_AUTO_JOURNAL_INTERVAL_SECS (default 30) so a fast loop doesn't hammer disk.
loop_running flag: every in-loop save marks loop_running=true. The final save on any terminal branch (completed, stopped, errored, auto_test_max_attempts_reached) flips it to false, so a cleanly-finished loop never triggers the resume prompt on the next launch.
Startup auto-prompt: when the REPL starts, if _checkpoint_latest.json has loop_running=true and a non-empty original_task_input, you see:
```
INFO interrupted auto-loop detected: iterations=47 last_activity=12m ago task="fix the parser bug…"
Resume? [Y/n]
```
Enter (or y) restores rt.messages, changed_files, change_events, and re-enters the auto-tool-loop with a [resume] banner so the model continues from where it left off. n keeps the checkpoint on disk for later.
Manual resume: /checkpoint resume does the same thing mid-session — useful if you dismissed the startup prompt and changed your mind, or want to abandon the current REPL state and pick up the prior interrupted loop.

The journal piggybacks on the existing single-file checkpoint, so disk footprint stays bounded — each save overwrites the previous. v1 keeps only the latest snapshot; full history with named resumable points is a v2.

Roadmap

Tier 1 (shipped): screen capture, window targeting, click, typing, on-screen OCR.
Tier 2 (shipped): UE5 / Blender / Unity bridges; Unity has live-editor C# drop.
Tier 3 (shipped): 3D-aware scene probes for UE5 / Blender / Unity, deterministic probe_diff, frame-folder video_capture + video_keyframes for temporal context.
Engineering reliability (shipped on main): cubist Picasso theme, auto-compact (token + turn-ratio + pre-turn guard), test feedback loop (cargo test / pytest / npm test with self-heal), auto-loop journal + resume (per-iteration checkpoint + startup auto-resume prompt).
Tier 4 (planned, ~1 week): MCP (Model Context Protocol) integration — the existing tools list (bash, read_file, edit_file, glob_search, grep_search, computer-control, engine bridges, scene probes) is all hardcoded in Rust today. Adding a proper MCP client lets users plug in any MCP server — Jira, Confluence, GitHub Enterprise, Datadog, Snowflake, internal RAG, anything the community ships — without patching ASI Code source. This is the single largest leverage point for the B2B story: it converts a fixed tool set into an open ecosystem and is what every "agentic IDE" in the market is converging on. Listed as the next milestone after engineering-reliability work lands.
Tier 5 (planned, ~2-3 weeks): Semantic codebase index — full-suite test feedback only works when the model can actually navigate the repo. On a 50K-line project glob+grep is fine; on a 500K-line monorepo it returns hundreds of low-signal matches that burn context. The plan is an embeddings + AST index (call graph, type hierarchy, file synopsis cache extended to symbol-level) so queries like "find all callers of PaymentService" or "show me every place this type is consumed" return ranked results without scanning every file. This is the gating capability for true enterprise-scale work; Cursor / Claude Code both ship a version of it. Holds until Tier 4 lands because the MCP layer naturally hosts the index server.
Tier 6 (next, then): mobile and cross-device remote — a small relay so the CLI can be triggered from a phone or another machine.
Tier 7 (deliberately not pursued in 2.0): AGI-level autonomous creative work. The bridge layer is the leverage — when the underlying models get stronger, every existing bridge inherits the new capability automatically.

Quick Start

cd "<path\to\ASI-Code>"
cargo run --offline -- repl --project "D:\Code\YourProject" --provider deepseek --model deepseek-v4-pro

Real CLI Execution Example

The CLI can execute real long-running tasks through tool calls, with approval gating, live progress, and artifact output.

Command

/toolcall bash command="cd D:\\test_code; python train_nn.py 2>&1"

What happened
Prompted for approval first (permission_mode=on-request)
Executed the command after approval
Streamed heartbeat logs during long execution (INFO bash still running ...)
Returned full training output
Saved generated artifact to disk:
D:/test_code/training_results.png

Sample output excerpt
Training PyTorch model for 1500 epochs...
Epoch  300/1500 | train_loss: 0.0034 | val_loss: 0.0036 | val_acc: 1.0000
Epoch  600/1500 | train_loss: 0.0006 | val_loss: 0.0008 | val_acc: 1.0000
...
Epoch 1500/1500 | train_loss: 0.0001 | val_loss: 0.0001 | val_acc: 1.0000
Plot saved to: D:/test_code/training_results.png
TOOL bash (ok)

This demonstrates three core behaviors:

1.Safe execution control (approval gate)
2.Reliable long-task handling (progress heartbeat + streaming logs)
3.End-to-end task completion (command success + output artifact)

## Real CLI Execution Example

The CLI can execute real long-running tasks through tool calls, with approval gating, live progress, and artifact output.

### Command
```text
/toolcall bash command="cd D:\\test_code; python train_nn.py 2>&1"

What happened
Prompted for approval first (permission_mode=on-request)
Executed the command after approval
Streamed heartbeat logs during long execution (INFO bash still running ...)
Returned full training output
Saved generated artifact to disk:
D:/test_code/training_results.png

Sample output excerpt
Training PyTorch model for 1500 epochs...
Epoch  300/1500 | train_loss: 0.0034 | val_loss: 0.0036 | val_acc: 1.0000
Epoch  600/1500 | train_loss: 0.0006 | val_loss: 0.0008 | val_acc: 1.0000
...
Epoch 1500/1500 | train_loss: 0.0001 | val_loss: 0.0001 | val_acc: 1.0000
Plot saved to: D:/test_code/training_results.png
TOOL bash (ok)

This demonstrates three core behaviors:

1.Safe execution control (approval gate)
2.Reliable long-task handling (progress heartbeat + streaming logs)
3.End-to-end task completion (command success + output artifact)


## Standalone Windows Install (No npm/pnpm)

Install ASI Code as a standalone app (`asi.exe`) with one command:

```powershell
cd "<path\to\ASI-Code>"
powershell -ExecutionPolicy Bypass -File .\scripts\windows\install_asi.ps1 -BuildRelease

Package a distributable release for users (zip + one-click installer exe):

powershell -ExecutionPolicy Bypass -File .\scripts\windows\package_release.ps1

Generated artifacts:

dist\asi-code-installer-<version>.exe (one-click installer, creates Start Menu entries)
dist\asi-code-installer-<version>.exe.sha256.txt
dist\asi-code-windows-x64-<version>.zip
dist\asi-code-windows-x64-<version>.zip.sha256.txt

Inside the ZIP root:

asi.exe (portable launcher, double-click entry)
ASI Launcher.exe (same launcher name for clarity)
start_asi.cmd (script-based launcher)
bin\asi.exe (CLI engine binary)

Detailed guide: INSTALL_WINDOWS.md

VS Code PowerShell Quick Launch (Windows)

cd "<path\to\ASI-Code>"
.\start_asi.ps1

What start_asi.ps1 does:

Prompts provider: OpenAI / DeepSeek / Claude
Prompts matching API key env (OPENAI_API_KEY / DEEPSEEK_API_KEY / ANTHROPIC_API_KEY)
Prompts project path
Runs: cargo run --release -- repl --provider <provider> --project <path> --no-setup

Optional flags:

.\start_asi.ps1 -Provider openai
.\start_asi.ps1 -Provider claude -ProjectPath "D:\Code\YourProject"
.\start_asi.ps1 -WithSetup

Double-click launcher:

.\start_asi.cmd

Ubuntu/WSL Quick Launch

cd /mnt/d/Code/Rust
chmod +x ./start_asi.sh
./start_asi.sh /mnt/d/Code/Rust

Skip startup setup wizard:

./start_asi.sh /mnt/d/Code/Rust --no-setup

Optional API keys before launch:

export OPENAI_API_KEY="..."
export DEEPSEEK_API_KEY="..."
export ANTHROPIC_API_KEY="..."
export ASI_AUTO_REVIEW_MODE="warn"
export ASI_AUTO_REVIEW_SEVERITY_THRESHOLD="medium"

VS Code Workspace Behavior

Run ASI Code from your project folder (or pass --project) and it will read/write directly in that workspace.
Agent file edits are written to the same files you see in VS Code Explorer.
Agent command execution happens in terminal via /run <command> or /toolcall bash <command>.

Useful Commands

cargo run --offline -- version
cargo run --offline -- api-page
cargo run --offline -- theme
cargo run --offline -- theme --select 1
cargo run --offline -- setup
cargo run --offline -- scan --deep
cargo run --offline -- scan --deep-limit 20 "**/*.py"
cargo run --offline -- tokenizer doctor --repo "D:\Code\rustbpe"
cargo run --offline -- tokenizer build --repo "D:\Code\rustbpe" --timeout-secs 900
cargo run --offline -- tokenizer train --repo "D:\Code\rustbpe" --input "D:\data\corpus.txt" --vocab-size 8192 --timeout-secs 1800 --auto-build
cargo run --offline -- wiki init --root "D:\Code\my-wiki"
cargo run --offline -- wiki ingest --root "D:\Code\my-wiki" --source "D:\Code\my-wiki\raw-input\paper.md" --title "Paper Notes"
cargo run --offline -- wiki query --root "D:\Code\my-wiki" --question "what changed in method section" --top 5 --save
cargo run --offline -- wiki lint --root "D:\Code\my-wiki" --write-report
cargo run --offline -- prompt "review this repo" --project "D:\Code\YourProject" --output-format text
cargo run --offline -- prompt "review this repo" --output-format json
cargo run --offline -- prompt "review this repo" --output-format jsonl
cargo run --offline -- repl --speed sprint
cargo run --offline -- prompt "review this repo" --speed deep --output-format json
cargo run --offline -- bench --provider deepseek --model deepseek-reasoner --project "<path\to\ASI-Code>" --suite core --agent-max-steps 8 --repeat 1 --out-dir bench_reports
cargo run --release -- bench --provider deepseek-reasoner --model deepseek-reasoner --project "<path\to\ASI-Code>" --suite gpt53-proxy6d --agent-max-steps 4 --repeat 3 --out-dir bench_reports
cargo run --offline -- sessions --limit 20
cargo run --offline -- self-update --source "<path\to\ASI-Code>\dist\asi-code-windows-x64-0.3.0.zip"
cargo run --offline -- self-update --source "https://example.com/asi-code-windows-x64-0.3.0.zip" --sha256 "<sha256>"
cargo run --offline -- config --telemetry-enabled true --safe-shell-mode true --undercover-mode true
cargo run --offline -- config --disable-web-tools true --disable-bash-tool false
cargo run --offline -- config --auto-review-mode warn --auto-review-severity-threshold medium
cargo run --offline -- config --allow-tool-rule read_file --allow-tool-rule "bash:git "
cargo run --offline -- config --deny-tool-rule "bash:rm -rf" --clear-tool-rules
cargo run --offline -- config --path-restriction-enabled true --additional-dir "D:\Code\shared"
cargo run --offline -- config --execution-speed sprint
cargo run --offline -- mcp list --json
cargo run --offline -- mcp add local-agent "python -m mcp_server" --json
cargo run --offline -- mcp config scope local-agent project --json
cargo run --offline -- mcp config trust local-agent on --json
cargo run --offline -- plugin list --json
cargo run --offline -- plugin add demo-plugin "D:\Code\my-plugin" --json
cargo run --offline -- plugin trust demo-plugin manual --json
cargo run --offline -- plugin trust demo-plugin hash --json
cargo run --offline -- plugin verify demo-plugin --json
powershell -ExecutionPolicy Bypass -File .\scripts\windows\benchmark_regression.ps1 -Project "D:\Code" -Agent

REPL Slash Commands

/help, /status, /project [path], /cost, /compact, /clear, /changes, /changes clear, /changes tail [n], /changes file <pattern>, /changes export <path> [md|json] [n], /exit
/provider <name>, /model <name>, /profile [standard|strict], /speed [sprint|deep], /permissions list|mode <mode>|allow <rule>|deny <rule>|rm-allow <rule>|rm-deny <rule>|clear-rules|path-restriction on|off|dirs|add-dir <path>|rm-dir <path>|clear-dirs|temp-list|temp-allow <rule>|temp-deny <rule>|temp-rm-allow <rule>|temp-rm-deny <rule>|temp-clear|temp-dirs|temp-add-dir <path>|temp-rm-dir <path>|temp-clear-dirs|temp-next-allow <rule>|temp-next-add-dir <path>|temp-next-clear|auto-review <off|warn|block>|auto-review-threshold <critical|high|medium|low>
/theme, /setup, /scan [patterns...] [--deep|--deep=N], /project <path>, /import <path>, /run <command>, /work <task>, /code <task>, /secure <task>, /workmode on|off, /auto on|off, /autoresearch help|doctor [--repo <path>]|init [--repo <path>]|run [--repo <path>] [--iterations <n>] [--timeout-secs <sec>] [--log <path>] [--description <text>] [--status <keep|discard|crash>], /tokenizer help|doctor [--repo <path>]|build [--repo <path>] [--debug] [--timeout-secs <sec>]|train --input <corpus.txt> [--repo <path>] [--vocab-size <n>] [--output <path>] [--pattern <regex>] [--name <name>] [--python-cmd <cmd>] [--timeout-secs <sec>] [--auto-build|--no-auto-build]
/runtime-profile [safe|fast|status] (one-command runtime safety preset switch)
/markdown on|off, /think on|off
/privacy status|telemetry on|off|tool-details on|off|undercover on|off|safe-shell on|off
/flags list|set <web_tools|bash_tool|subagent|research> on|off
/policy sync
/tools, /toolcall <tool> <args>
/audit tail [n] | /audit stats [n] | /audit tools [n] | /audit reasons [n] | /audit export <path> [md|json] [n] | /audit export-last [dir] [md|json|both] [n] (tail, summaries, and export report)
/sessions [limit], /resume <id>, /save, /checkpoint [status|on|off|save|load|clear]
/oauth login <token>, /oauth logout
/todo ..., /memory ..., /git ..., /mcp [list [--json]|add <name> <command> [--scope <session|project|global>] [--json]|rm <name> [--json]|show <name> [--json]|auth <name> <none|bearer|api-key|basic> [value] [--json]|oauth login <name> <provider> <token> [--scope <session|project|global>] [--link-auth] [--json]|oauth status <name> [provider] [--scope <session|project|global>] [--json]|oauth logout <name> [provider] [--scope <session|project|global>] [--json]|config <name> <key> <value> [--json]|config rm <name> <key> [--json]|config scope <name> <session|project|global> [--json]|config trust <name> <on|off> [--json]|export <path.json> [--json]|import <path.json> [merge|replace] [--json]|start <name> [command] [--json]|stop <name> [--json]], /plugin [list [--json]|add <name> <path> [--json]|rm <name> [--json]|show <name> [--json]|enable <name> [--json]|disable <name> [--json]|trust <name> <manual|hash> [hash] [--json]|verify <name> [--json]|config set <name> <key> <value> [--json]|config rm <name> <key> [--json]|export <path.json> [--json]|import <path.json> [merge|replace] [--json]|market [--json]], /notebook ..., /agent [spawn [--json|--jsonl] <task>|send [--json|--jsonl] [--interrupt] [--no-context] <id> <task>|wait [--json|--jsonl] [id] [timeout_secs]|list [--json|--jsonl]|status [--json|--jsonl] <id>|log [--json|--jsonl] <id> [--tail <n>]|retry [--json|--jsonl] <id>|cancel [--json|--jsonl] <id>|close [--json|--jsonl] <id>|view [foreground|background|all]|front|back]
Undercover commit helper: /git commit-msg <message>

Prompt Output Formats

--output-format text: human-readable output
--output-format json: single JSON object (stable machine payload)
--output-format jsonl: line-delimited JSON events (schema_version=1) for CI/automation pipelines
- prompt.result
- prompt.changed_files
- prompt.native_tool_calls
- prompt.auto_validation
- prompt.review
- prompt.runtime

Agent JSON Schema (Stable v1)

When using JSON-enabled /agent commands, output uses a stable envelope:

{
  "schema_version": "1",
  "command": "<spawn|send|list|status|wait|log|retry|cancel|close>",
  "agent": { "...": "..." }
}

Compatibility policy:

schema_version=1 remains stable for backward-compatible additions.
Any breaking field rename/removal or shape change requires a schema version bump.

Agent JSONL Event Stream (Stable v1)

When using /agent ... --jsonl, each command emits one JSONL event row:

{
  "schema_version": "1",
  "event": "agent.<spawn|send|list|status|wait|log|retry|cancel|close>",
  "data": {
    "command": "<spawn|send|list|status|wait|log|retry|cancel|close>",
    "agent": { "...": "same payload shape as --json agent field" }
  }
}

Commands and Shapes

/agent spawn --json <task> (or --jsonl)
- optional profile/run mode: /agent spawn --json --profile <name> [--background|--foreground] <task>
- agent: id, status (running), provider, model, run_mode (foreground|background), task
/agent list --json (or --jsonl)
/agent list --json [--scope <foreground|background|all>] [--profile <name>] [--skill <name>]
- agent: count, filters_applied (scope, profile, skill), diagnostics (filters, total_items, counts, run_modes, timings), items[]
- each item: id, status, provider, model, run_mode, turns, interrupted_count, last_interrupted_at_ms, started_at_ms, finished_at_ms, task, preview
/agent status --json <id> (or --jsonl)
- agent: id, status, provider, model, allow_rules, deny_rules, rules_source, diagnostics (rules_source, allow_rule_count, deny_rule_count), run_mode, turns, interrupted_count, last_interrupted_at_ms, started_at_ms, finished_at_ms, age_ms, duration_ms, task, preview
- unknown id: agent.id, agent.status=unknown
/agent wait --json [id] [timeout_secs] (or --jsonl)
- finished: agent.id, provider, model, run_mode, task, started_at_ms, finished_at_ms, status (done|error), ok, result
- timeout: agent.status=timeout, timeout_secs, message
- idle: agent.status=idle, message
/agent close --json <id> (or --jsonl)
- agent: id, status (closed), message
/agent cancel --json <id> (or --jsonl)
- agent: id, status (cancelled|closed|completed|failed|unknown), message
/agent retry --json <id> (or --jsonl)
- agent: id, status, provider, model, turns, run_mode, task, message
/agent log --json <id> [--tail <n>] (or --jsonl)
- agent: id, status, provider, model, run_mode, turns, events_total, tail, items[]
- each log item: at_ms, event, message
/agent send --json [--interrupt] [--no-context] <id> <task> (or --jsonl)
- agent: id, status, provider, model, run_mode, turns, interrupted_count, last_interrupted_at_ms, started_at_ms, finished_at_ms, task, context (append|reset), interrupt, message

Quick Examples

# Spawn
/agent spawn --json "inspect src/main.rs"

# Spawn with agent profile (loaded from ./agent_profiles.json)
/agent spawn --json --profile safe-review "inspect src/main.rs"

# Spawn as background task (for /agent view background)
/agent spawn --json --background "inspect src/main.rs"

# List
/agent list --json
/agent list --json --scope background --profile safe-review --skill review

# Send
/agent send --json sa-1 "continue with tests"

# Status
/agent status --json sa-1

# Wait
/agent wait --json sa-1 30

# Close
/agent close --json sa-1

# Cancel running task
/agent cancel --json sa-1

# Retry last task for one subagent
/agent retry --json sa-1

# Show latest subagent events
/agent log --json sa-1 --tail 20

MCP JSON Schema (Stable v1)

When using JSON-enabled /mcp commands, output uses a stable envelope:

{
  "schema_version": "1",
  "command": "<mcp_list|mcp_show|mcp_add|mcp_rm|mcp_auth|mcp_config_set|mcp_config_rm|mcp_export|mcp_import|mcp_start|mcp_stop>",
  "mcp": { "...": "..." }
}

Compatibility policy:

schema_version=1 remains stable for backward-compatible additions.
Any breaking field rename/removal or shape change requires a schema version bump.

Agent Profiles (`agent_profiles.json`)

/agent spawn --profile <name> <task> can apply project-local defaults from ./agent_profiles.json.

Supported fields:

provider
model
permission_mode
allowed_tools (tool whitelist; maps to allow rules in subagent runtime)
denied_tools (tool deny list; maps to deny rules in subagent runtime)
default_skills (skill labels attached to subagent metadata)
disable_web_tools
disable_bash_tool

Example:

{
  "profiles": {
    "safe-review": {
      "provider": "deepseek",
      "model": "deepseek-reasoner",
      "permission_mode": "on-request",
      "allowed_tools": ["read_file", "glob_search", "grep_search"],
      "denied_tools": ["bash"],
      "default_skills": ["review", "security"],
      "disable_web_tools": true,
      "disable_bash_tool": false
    },
    "fast-code": {
      "provider": "openai",
      "model": "gpt-5.3-codex",
      "permission_mode": "workspace-write"
    }
  }
}

Notes:

allowed_tools is enforced as allow-rules in the subagent runtime. Tools outside the list are blocked.
denied_tools is enforced as deny-rules in the subagent runtime. Matching tools are blocked even if broadly allowed.
default_skills is surfaced in agent JSON outputs (spawn/list/status/send/wait) and logged for traceability.

Commands and Shapes

/mcp list --json
- mcp: count, diagnostics (trusted_count, untrusted_count, scope_counts), items[]
each item: name, status, pid, command, scope, trusted, auth_type, auth_value, config[] (key, value)
/mcp show <name> --json
- found: mcp server record shape (same fields as list item) plus diagnostics (has_auth_value, config_key_count, scope, trusted, oauth_provider, oauth_token_present)
- missing: mcp.name, mcp.status=unknown
/mcp add <name> <command> [--scope <session|project|global>] --json
- mcp: added (bool), scope, server (server record)
/mcp rm <name> --json
- mcp: name, changed (bool)
/mcp auth <name> <none|bearer|api-key|basic> [value] --json
- mcp: name, auth_type, value_set (bool)
/mcp oauth login <name> <provider> <token> [--scope <session|project|global>] [--link-auth] --json
- mcp: name, provider, scope, saved (bool), linked_auth (bool)
/mcp oauth status <name> [provider] [--scope <session|project|global>] --json
- mcp: name, provider, token_present (bool), token_scope, request_scope, stored_providers (string array)
/mcp oauth logout <name> [provider] [--scope <session|project|global>] --json
- mcp: name, provider, scope, removed (bool)
/mcp config <name> <key> <value> --json
- mcp: name, key
/mcp config rm <name> <key> --json
- mcp: name, key, changed (bool)
/mcp config scope <name> <session|project|global> --json
- mcp: name, scope
/mcp config trust <name> <on|off> --json
- mcp: name, trusted (bool)
/mcp export <path.json> --json
- mcp: path
/mcp import <path.json> [merge|replace] --json
- mcp: path, mode, total_servers
/mcp start <name> [command] --json
- success: mcp: started (bool), server (server record)
- blocked/error: mcp: started=false, name, error
/mcp stop <name> --json
- mcp: name, changed (bool)

Quick Examples

# List
/mcp list --json

# Show
/mcp show local-agent --json

# Add
/mcp add local-agent "python -m mcp_server" --json

# OAuth (stored token for MCP server scope)
/mcp oauth login local-agent deepseek sk-example --scope project --link-auth --json
/mcp oauth status local-agent deepseek --scope project --json
/mcp oauth logout local-agent deepseek --scope project --json

# Start / Stop
/mcp start local-agent --json
/mcp stop local-agent --json

Equivalent top-level CLI subcommands:

asi mcp list --json
asi mcp add local-agent "python -m mcp_server" --json
asi mcp config scope local-agent project --json
asi mcp config trust local-agent on --json
asi mcp oauth status local-agent deepseek --scope project --json

MCP Trust and Publish Metadata Policy

Server records support optional publish metadata via /mcp config:

source
version
signature

Accepted formats:

source must start with one of: https://, http://, git+, file://, local://
signature must start with sha256: or sig:
version must be non-empty when set

Examples:

/mcp config local-agent source "https://example.com/org/mcp-server" --json
/mcp config local-agent version "1.2.3" --json
/mcp config local-agent signature "sha256:abcd1234" --json

asi mcp config local-agent source "https://example.com/org/mcp-server" --json
asi mcp config local-agent version "1.2.3" --json
asi mcp config local-agent signature "sig:publisher-v1" --json

Plugin JSON Schema (Stable v1)

When using JSON-enabled /plugin commands (or asi plugin ... --json), output uses:

{
  "schema_version": "1",
  "command": "<plugin_list|plugin_add|plugin_rm|plugin_show|plugin_enable|plugin_disable|plugin_trust|plugin_config_set|plugin_config_rm|plugin_export|plugin_import>",
  "plugin": { "...": "..." }
}

Quick examples:

/plugin list --json
/plugin show demo-plugin --json
/plugin trust demo-plugin manual --json
/plugin trust demo-plugin hash --json
/plugin verify demo-plugin --json

asi plugin list --json
asi plugin add demo-plugin "D:\Code\my-plugin" --json
asi plugin trust demo-plugin manual --json
asi plugin trust demo-plugin hash --json
asi plugin verify demo-plugin --json

trust ... hash stores a deterministic SHA-256 of the plugin directory contents (not only manifest), so changing plugin code/config invalidates trust until re-trusted.

Plugin Trust and Publish Metadata Policy

Plugin records support optional publish metadata via /plugin config set:

source
version
signature

Accepted formats:

source must start with one of: https://, http://, git+, file://, local://
signature must start with sha256: or sig:
version must be non-empty when set

Examples:

/plugin config set demo-plugin source "https://example.com/org/demo-plugin" --json
/plugin config set demo-plugin version "0.9.0" --json
/plugin config set demo-plugin signature "sha256:abcd1234" --json

asi plugin config set demo-plugin source "https://example.com/org/demo-plugin" --json
asi plugin config set demo-plugin version "0.9.0" --json
asi plugin config set demo-plugin signature "sig:publisher-v1" --json

Environment Variables

OpenAI: OPENAI_API_KEY, OPENAI_BASE_URL
DeepSeek: DEEPSEEK_API_KEY, DEEPSEEK_BASE_URL
Claude: ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL
ASI toggles: ASI_TELEMETRY, ASI_UNDERCOVER, ASI_AUTO_CHECKPOINT, ASI_NATIVE_TOOL_CALLING, ASI_PROVIDER_MAX_RETRIES, ASI_PROVIDER_RETRY_BACKOFF_MS, ASI_MODEL_AUTO_FALLBACK
Auto-compaction: ASI_AUTO_COMPACT_INPUT_TOKENS (default 120000; 0 disables the token trigger), ASI_AUTO_COMPACT_TURN_RATIO (default 0.80; values outside (0.0, 1.0] disable the turn trigger). Both are evaluated before each REPL turn AND inside the auto-tool-loop.
Test feedback loop (opt-in): ASI_AUTO_TEST (1 / true / on to enable; default OFF), ASI_AUTO_TEST_MAX_ATTEMPTS (default 3, clamped to >=1; caps consecutive self-heal attempts before the loop exits with auto_test_max_attempts_reached).
Auto-loop journal (default ON): ASI_AUTO_JOURNAL (0 / false / off to disable), ASI_AUTO_JOURNAL_INTERVAL_SECS (default 30, clamped to >=1; minimum seconds between per-iteration saves to bound disk writes).
Project instruction loading: ASI_PROJECT_INSTRUCTIONS_LAYERED (default true), ASI_PROJECT_INSTRUCTIONS_MAX_LEVELS (default 4), ASI_CLAUDE_SINGLE (default true)
- instruction candidates (high to low priority per directory):
  - Claude single mode: CLAUDE.override.md, CLAUDE.local.md, CLAUDE.md
  - fallback mode: AGENTS.override.md, AGENTS.local.md, AGENTS.md, README.override.md, README.local.md, README.md
  - non-single mode loads all above in order
Hooks (disabled by default): ASI_HOOKS_ENABLED, ASI_HOOK_PRE_TOOL_USE, ASI_HOOK_PERMISSION_REQUEST, ASI_HOOK_POST_TOOL_USE, ASI_HOOK_SESSION_START, ASI_HOOK_USER_PROMPT_SUBMIT, ASI_HOOK_STOP, ASI_HOOK_SUBAGENT_STOP, ASI_HOOK_PRE_COMPACT, ASI_HOOK_POST_COMPACT, ASI_HOOK_TIMEOUT_SECS (default 15), ASI_HOOK_JSON (default true, provides ASI_HOOK_INPUT_JSON), ASI_HOOK_CONFIG_PATH (optional JSON handler matrix file), ASI_HOOK_FAILURE_POLICY (fail-closed default, supports fail-open)
- Hook plugin context envs: ASI_HOOK_PLUGIN_COUNT, ASI_HOOK_PLUGIN_NAMES (trusted+enabled plugin names CSV)
- Windows lifecycle hook template: scripts/windows/hooks/lifecycle_event.ps1 (logs SessionStart/UserPromptSubmit/Stop/SubagentStop/PreCompact/PostCompact into .asi/hooks.lifecycle.log)
- Trusted + enabled plugins can also provide hook scripts in .codex-plugin/plugin.json under hooks:
  - single-event keys: pre_tool_use, permission_request, post_tool_use, session_start, user_prompt_submit, stop, subagent_stop, pre_compact, post_compact
  - optional handlers array with fields: event, script, timeout_secs, json_protocol, tool_prefix, permission_mode, failure_policy
  - runtime executes order: env hook -> config handlers -> plugin hooks (deterministic by plugin name)
Subagent auto tool-loop toggle: ASI_SUBAGENT_TOOL_LOOP (default true)
Review risk scoring weights: ASI_REVIEW_RISK_WEIGHT_CRITICAL (default 8), ASI_REVIEW_RISK_WEIGHT_HIGH (default 4), ASI_REVIEW_RISK_WEIGHT_MEDIUM (default 2), ASI_REVIEW_RISK_WEIGHT_LOW (default 1), ASI_REVIEW_RISK_WEIGHT_UNKNOWN (default 1)
Review JSON-only strict gate: ASI_REVIEW_JSON_ONLY_STRICT_SCHEMA (default false; when true, invalid /review --json-only output is wrapped in an explicit error envelope)
Review JSON-only strict fail exit: ASI_REVIEW_JSON_ONLY_STRICT_FAIL_EXIT (default false; when true, prompt mode exits non-zero if strict JSON-only schema validation fails)
Review JSON-only prompt envelope: ASI_REVIEW_JSON_ONLY_PROMPT_ENVELOPE (default false; when true, prompt /review --json-only success output uses the same {schema_version,status,review} envelope as REPL)
Security defaults:
- Default permission_mode is on-request (interactive approvals for mutating tools)
- Default sandbox mode is local (startup sets ASI_SANDBOX=local when unset; explicit env keeps priority)
- MCP start trust gate: untrusted MCP servers are blocked by default; trust explicitly with /mcp config trust <name> on (override for controlled environments: ASI_MCP_ALLOW_UNTRUSTED_START=true)
- Use /runtime-profile safe to enforce on-request + local
- Use /runtime-profile fast to switch to workspace-write + disabled

/status includes project_context sources=<n> [...] so you can verify which instruction files were loaded. When hooks are enabled, /status also reports hook diagnostics summary:

hooks enabled=true diagnostics=<n> denied=<n> errors=<n>
up to 5 recent hook outcome lines in shape:
- hook source=<env-hook|config-hook|plugin-hook> event=<...> allow=<...> is_error=<...> reason=<...> When hooks are disabled, it reports:
hooks enabled=false

Windows hook templates are provided under scripts/windows/hooks/ with ready-to-use examples. Includes permission_request_strict.ps1 for default-deny bash policy with whitelist prefixes. Also includes hooks.config.example.json for event-matrix style hook configuration. For quick enable/disable in the current PowerShell session, use:

scripts/windows/hooks/enable_strict_hooks.ps1
scripts/windows/hooks/disable_hooks.ps1

Hook config matrix (ASI_HOOK_CONFIG_PATH) supports multiple handlers with filters:

event: PreToolUse, PermissionRequest, PostToolUse, SessionStart, UserPromptSubmit, Stop, SubagentStop, PreCompact, PostCompact, or *
script: command string to execute
timeout_secs: optional per-handler override
json_protocol: optional per-handler override
tool_prefix: optional tool name prefix filter (for example bash)
permission_mode: optional permission mode filter (for example workspace-write)
failure_policy: optional per-handler override (fail-closed or fail-open)

Failure policy behavior:

fail-closed: hook runtime errors (spawn failure, timeout, non-json parse/exec errors) block the action.
fail-open: hook runtime errors are treated as allow, but explicit deny results still block.

If both legacy env hooks and config handlers are set, both run for the same event.

Hook JSON versioning policy:

schema_version is the top-level hook payload schema version (currently "1").
event_version is per-event payload contract version (currently "1" for all events).
Backward-compatible field additions keep the same version; breaking changes require bumping the corresponding version and update notes.

Prompt JSON output includes a review object for /review ... tasks:

schema_version ("1", stable for backward-compatible additions)
is_review_task (bool)
schema_valid (bool)
schema_error (string|null)
sections when parseable:
- findings (structured objects: severity, normalized_severity, path, line, location, summary, evidence, risk, raw)
- findings_sorted (same shape as findings, sorted by severity and location for deterministic UI ordering)
- findings_raw (original lines for compatibility)
- missing_tests, open_questions, summary
stats:
- has_findings (bool)
- total_findings (number)
- severity_counts (critical, high, medium, low, unknown)
- risk_score_total (number, weighted total across all findings using configured severity weights)
- missing_tests_count (number, excludes None)
- top_risk_paths (top files by weighted risk score, includes per-path severity counts)

Machine-consumable example:

cargo run --release -- prompt "/review inspect parser" `
  --provider openai `
  --model gpt-5.3-codex `
  --project "<path\to\ASI-Code>" `
  --output-format json | `
  ConvertFrom-Json | `
  Select-Object -ExpandProperty review | `
  ConvertTo-Json -Depth 8

Fast review-only JSON (no wrapper envelope):

cargo run --release -- prompt "/review inspect parser --json-only" `
  --provider openai `
  --model gpt-5.3-codex `
  --project "<path\to\ASI-Code>"

Optional strict CI gate for /review --json-only:

$env:ASI_REVIEW_JSON_ONLY_STRICT_SCHEMA="true"
$env:ASI_REVIEW_JSON_ONLY_STRICT_FAIL_EXIT="true"
cargo run --release -- prompt "/review inspect parser --json-only" `
  --provider openai `
  --model gpt-5.3-codex `
  --project "<path\to\ASI-Code>" `
  --output-format json

When strict mode is enabled and schema validation fails, JSON-only output is:

REPL /review ... --json-only success shape: { "schema_version":"1", "status":"ok", "review":{...} }
REPL /review ... --json-only strict-fail shape: { "schema_version":"1", "status":"error", "error":"review_schema_invalid", "schema_error":"...", "review":{...} }
Prompt /review ... --json-only --output-format json default success shape: raw review object (no outer status)
Prompt /review ... --json-only --output-format json strict-fail shape: same error envelope
Prompt success envelope optional: set ASI_REVIEW_JSON_ONLY_PROMPT_ENVELOPE=true to emit { "schema_version":"1", "status":"ok", "review":{...} }
smoke_review_json.ps1 hard-checks status=="ok" and top-level schema_version=="1" when -PromptEnvelope on is used.
If ASI_REVIEW_JSON_ONLY_STRICT_FAIL_EXIT=true, prompt mode returns a non-zero exit code after emitting the error envelope (CI-friendly failure signal).

Multi-Provider Smoke Recipes (Windows)

The smoke scripts now support provider-scoped runs for API-compat and gateway checks. You can run OpenAI/DeepSeek flows independently without requiring Claude credentials.

1) DeepSeek-only API Compat (no OpenAI key required)

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_api_compat.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "D:\test_code" `
  -SkipOpenAi `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>" `
  -DeepSeekModel "deepseek-reasoner"

2) OpenAI-only API Compat (no DeepSeek key required)

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_api_compat.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "D:\test_code" `
  -SkipDeepSeek `
  -OpenAiBaseUrl "https://api.openai.com/v1" `
  -OpenAiApiKey "<YOUR_OPENAI_KEY>" `
  -OpenAiModel "gpt-5.3-codex"

3) Gateway Smoke by Provider (openai|deepseek|claude)

DeepSeek example:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_gateway.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "D:\test_code" `
  -Provider deepseek `
  -ApiKey "<YOUR_DEEPSEEK_KEY>" `
  -Model "deepseek-reasoner"

OpenAI example:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_gateway.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "D:\test_code" `
  -Provider openai `
  -BaseUrl "https://api.openai.com/v1" `
  -ApiKey "<YOUR_OPENAI_KEY>" `
  -Model "gpt-5.3-codex"

4) One-Command `smoke_all` with Provider-Scoped Gateway

Fast DeepSeek gateway path (skip tool turn for speed):

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_all.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "D:\test_code" `
  -Repo "D:\Code\rustbpe" `
  -Quick `
  -SkipApiCompat -SkipProviderModel -SkipTokenizer -SkipCheckpoint `
  -RunGateway `
  -GatewayProvider deepseek `
  -GatewaySkipToolTurn `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>" `
  -DeepSeekModel "deepseek-reasoner"

Optional auto-summary for smoke reports:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_all.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -Repo "D:\Code\rustbpe" `
  -SkipApiCompat -SkipProviderModel -SkipTokenizer -SkipCheckpoint -SkipGateway `
  -ReportJsonPath ".\artifacts\smoke_all.json" `
  -RenderSummary

Optional custom summary output path:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_all.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -Repo "D:\Code\rustbpe" `
  -SkipApiCompat -SkipProviderModel -SkipTokenizer -SkipCheckpoint -SkipGateway `
  -ReportJsonPath ".\artifacts\smoke_all.json" `
  -RenderSummary `
  -SummaryOutFile ".\artifacts\reports\custom_summary.md"

Strict profile (CI parity) example for smoke_all.ps1:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_all.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -Repo "D:\Code\rustbpe" `
  -StrictProfile `
  -SkipApiCompat -SkipProviderModel -SkipTokenizer -SkipCheckpoint -SkipGateway `
  -ReportJsonPath ".\artifacts\smoke_all_strict_profile.json"

-StrictProfile requires these checks to remain enabled and rejects:

-SkipHookMatrix
-SkipHooksCliAdvanced
-SkipSubagent

Notes

smoke_api_compat.ps1 and smoke_api_compat_min.ps1 support:
- -SkipOpenAi
- -SkipDeepSeek
smoke_review_json.ps1 validates /review structured JSON fields (schema_valid, sections.findings_sorted, stats.top_risk_paths, stats.missing_tests_count).
smoke_prompt_jsonl.ps1 validates prompt --output-format jsonl event stream fields (schema_version, event, data) and required events (prompt.result, prompt.changed_files, prompt.native_tool_calls, prompt.auto_validation, prompt.review, prompt.runtime).
smoke_all.ps1 supports env fallbacks when flags are not passed:
- ASI_SMOKE_SKIP_OPENAI=1
- ASI_SMOKE_SKIP_DEEPSEEK=1
smoke_gateway.ps1 supports generic provider args (-Provider/-BaseUrl/-ApiKey/-Model) and legacy provider-specific args for backward compatibility.

5) Interactive Recipe Runner (OpenAI/DeepSeek)

Use run_smoke_recipes.ps1 to avoid manual command assembly:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\run_smoke_recipes.ps1 `
  -Provider deepseek `
  -Recipe smoke-all-gateway `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "D:\test_code" `
  -Repo "D:\Code\rustbpe" `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>" `
  -Quick `
  -GatewaySkipToolTurn

Unified smoke entrypoint (mode router):

# strict (default)
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke.ps1 `
  -Mode strict `
  -Provider deepseek `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>"

# risk review + gateway
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke.ps1 `
  -Mode risk `
  -Provider deepseek `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>"

# gateway path
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke.ps1 `
  -Mode gateway `
  -Provider openai `
  -OpenAiApiKey "<YOUR_OPENAI_KEY>" `
  -NoQuick

For a shorter strict-regression entrypoint, use smoke_strict_quick.ps1 (compatibility alias to smoke.ps1 -Mode strict).

One-command strict quick wrapper (recommended for daily regression):

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_strict_quick.ps1 `
  -Provider deepseek `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>"

OpenAI variant:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_strict_quick.ps1 `
  -Provider openai `
  -OpenAiApiKey "<YOUR_OPENAI_KEY>" `
  -NoQuick

One-command risk review + gateway (relaxed schema gate):

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\run_smoke_recipes.ps1 `
  -Provider deepseek `
  -Recipe smoke-all-gateway-risk `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -Repo "D:\Code\rustbpe" `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>" `
  -ReportDir ".\artifacts" `
  -RenderSummary

Recommended standard regression template (copy and run):

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\run_smoke_recipes.ps1 `
  -Provider deepseek `
  -Recipe smoke-all-gateway-risk `
  -AsiExe ".\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -Repo "D:\Code\rustbpe" `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>" `
  -ReportDir ".\artifacts\risk_recipe_live" `
  -RenderSummary `
  -SummaryOutFile ".\artifacts\risk_recipe_live\SUMMARY_SMOKE_ALL_GATEWAY_RISK.md"

Recommended template for the OpenAI API endpoint (copy and run):

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\run_smoke_recipes.ps1 `
  -Provider openai `
  -Recipe smoke-all-gateway-risk `
  -AsiExe ".\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -Repo "D:\Code\rustbpe" `
  -OpenAiApiKey "<YOUR_OPENAI_KEY>" `
  -OpenAiBaseUrl "https://api.openai.com/v1" `
  -ReportDir ".\artifacts\risk_recipe_live_openai" `
  -RenderSummary `
  -SummaryOutFile ".\artifacts\risk_recipe_live_openai\SUMMARY_SMOKE_ALL_GATEWAY_RISK.md"

Direct review JSON smoke example:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_review_json.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -Provider deepseek `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>" `
  -Model "deepseek-reasoner" `
  -PromptAutoTools off `
  -PromptEnvelope off `
  -SchemaRetries 1 `
  -FailOnSchemaInvalid on `
  -ReviewTask "Provide exactly the required sectioned review format with one low severity finding for src/main.rs:1 and no tool calls"

Direct prompt JSONL smoke example:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_prompt_jsonl.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -Provider deepseek `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>" `
  -Model "deepseek-reasoner" `
  -PromptText "Reply with exactly JSONL_SMOKE_OK"

One-command artifacts + summary via recipe runner:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\run_smoke_recipes.ps1 `
  -Provider deepseek `
  -Recipe review-json `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>" `
  -ReviewJsonPromptAutoTools off `
  -ReviewJsonPromptEnvelope off `
  -ReviewJsonSchemaRetries 1 `
  -ReviewJsonFailOnSchemaInvalid on `
  -ReviewJsonTask "Provide exactly the required sectioned review format with one low severity finding for src/main.rs:1 and no tool calls" `
  -ReportDir ".\artifacts" `
  -RenderSummary

Supported recipes:

api-compat
api-compat-min
gateway
review-json
hook-matrix
hooks-cli-advanced
subagent
subagent-strict
smoke-all-gateway
smoke-all-gateway-risk
smoke-all-strict

Unified wrapper modes:

smoke.ps1 -Mode strict
smoke.ps1 -Mode risk
smoke.ps1 -Mode gateway

CI Machine-Readable Template (Windows)

Run one command to build/test/smoke and emit a single JSON report:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\ci_machine_json.ps1 `
  -Provider deepseek `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>"

Optional gate controls:

# Use a custom gate template and keep process success even when gates fail
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\ci_machine_json.ps1 `
  -Provider deepseek `
  -DeepSeekApiKey "<YOUR_DEEPSEEK_KEY>" `
  -GateRulesPath .\scripts\windows\ci_gate_rules.template.json `
  -SkipFailGate

# Validate CI report schema + gates
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\validate_ci_machine_report.ps1 `
  -ReportPath .\artifacts\ci_machine_json\ci_machine_report.json `
  -Strict

Output:

JSON report: .\artifacts\ci_machine_json\ci_machine_report.json
Smoke summary: .\artifacts\ci_machine_json\smoke\SMOKE_SUMMARY.md
Prompt JSONL smoke report: .\artifacts\ci_machine_json\smoke\smoke_prompt_jsonl.json
Gate rules template: .\scripts\windows\ci_gate_rules.template.json

Tips:

Set -DryRun to print the exact command without executing.
Use -Provider ask -Recipe ask for interactive menu selection.
Use -StrictCi for one-command strict CI mode. It forces smoke-all-strict, enables summary rendering, requires strict review-json validation, and defaults ReportDir to .\artifacts\strict_ci when not provided.
-StrictCi defaults review-json task to deterministic schema validation prompt (Provide exactly the required sectioned review format...) and ReviewJsonPromptAutoTools=off for stability; override via explicit flags when needed.
smoke-all-gateway and smoke-all-gateway-risk now include hook-matrix smoke, hooks-cli-advanced smoke, and subagent smoke by default. Use -SkipHookMatrix, -SkipHooksCliAdvanced, and/or -SkipSubagent to disable them.
smoke-all-strict enforces hook-matrix, hooks-cli-advanced, and subagent checks; -SkipHookMatrix, -SkipHooksCliAdvanced, and -SkipSubagent are rejected.
Use -AllowProviderNetworkError to enable all network/provider downgrade toggles at once for smoke-all-* recipes.
Use -AllowReviewJsonNetworkError for smoke-all-* recipes when provider/network instability is expected. review-json smoke network/provider failures are downgraded to warn so later steps (hook-matrix/subagent/gateway) can still run.
Use -AllowSubagentNetworkError for smoke-all-* recipes when provider/network instability is expected. subagent smoke network/provider failures are downgraded to warn so gateway checks can still run.
Use -AllowGatewayNetworkError for smoke-all-* recipes when provider/network instability is expected. gateway smoke network/provider failures are downgraded to warn so the aggregate run can finish and emit reports.
CI machine report gate behavior:
- schema_version="1" and ci_schema="asi.ci.machine_report.v1" are fixed.
- gates.pass=false will exit non-zero by default.
- pass -SkipFailGate to always emit report and exit zero (useful for flaky network runs).
For recipe subagent, use -SubagentAllowWaitError to downgrade wait failure to warning-pass mode.
Recipe subagent-strict always enforces wait success and rejects -SubagentAllowWaitError.
When using recipe smoke-all-gateway, summary generation is handled by smoke_all.ps1; the runner will not invoke render_smoke_summary.ps1 a second time.
smoke-all-gateway-risk defaults to:
- ReviewJsonTask = "inspect current project and report top bug risks"
- ReviewJsonPromptAutoTools = on
- ReviewJsonSchemaRetries = 1
- ReviewJsonFailOnSchemaInvalid = off You can still override these via CLI flags.
Use -ReviewJsonTask to switch between deterministic schema validation prompts and real project risk-review prompts.
Use -ReviewJsonPromptAutoTools on|off to control review-json prompt continuation behavior.
Use -ReviewJsonPromptEnvelope on|off to control prompt /review --json-only success envelope mode in smoke runs.

Hook matrix smoke (event handler config) example:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_hook_matrix.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -ReportJsonPath "<path\to\ASI-Code>\artifacts\smoke_hook_matrix.json"

Recipe runner equivalent:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\run_smoke_recipes.ps1 `
  -Recipe hook-matrix `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -ReportDir "<path\to\ASI-Code>\artifacts"

Hooks CLI advanced smoke example:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_hooks_cli_advanced.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "<path\to\ASI-Code>" `
  -ReportJsonPath "<path\to\ASI-Code>\artifacts\smoke_hooks_cli_advanced.json"

6) Subagent JSON Smoke (spawn/list/wait envelope checks)

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\windows\smoke_subagent.ps1 `
  -AsiExe "<path\to\ASI-Code>\target\release\asi.exe" `
  -Project "D:\test-cli" `
  -Provider deepseek `
  -Model deepseek-reasoner `
  -OutputMode json `
  -ReportJsonPath "<path\to\ASI-Code>\artifacts\smoke_subagent.json"

-OutputMode supports json (default) and jsonl.
For JSONL mode use: -OutputMode jsonl (parses agent.<command> events).
Pass criteria are strict: /agent wait --json must return agent.status == "done" and agent.ok == true; otherwise smoke fails with failure_category=subagent_wait_not_successful.
Optional downgrade mode: add -AllowWaitError to keep status=pass when wait fails, and record metrics.wait_warning=true with failure details in metrics.
Use -ReviewJsonSchemaRetries <n> to retry schema validation by rerunning the prompt when output format is unstable.
Use -ReviewJsonFailOnSchemaInvalid on|off to choose whether schema-invalid review-json should fail smoke or only warn.

Smoke Report Naming

Recommended report artifacts in the report directory:

smoke_all.json: aggregated run result from smoke_all.ps1
smoke_gateway.json: provider gateway smoke result
smoke_review_json.json: /review JSON schema/stats smoke result
smoke_prompt_jsonl.json: prompt --output-format jsonl event stream smoke result
smoke_hook_matrix.json: hook event-matrix allow/deny regression result
SMOKE_SUMMARY.md: markdown summary generated by render_smoke_summary.ps1

By default, smoke_all.ps1 -RenderSummary writes SMOKE_SUMMARY.md under the same directory as -ReportJsonPath. Use -SummaryOutFile <path> to override the summary output file path.

Beta Known Issues

Unsigned Windows binaries can still trigger SmartScreen warnings. Signed artifacts are supported by scripts/windows/package_release.ps1 -SignArtifacts ....
New builds auto-validate provider/model compatibility. If you still see DeepSeek Model Not Exist, verify you are not running an old binary and check /status.
API keys are now masked in startup flows, but previously saved plaintext environment variables remain visible via OS environment inspection.
Some enterprise/proxy networks can block provider endpoints and cause connection or TLS errors.

Install & Troubleshooting

Quick Checks

Verify binary works:
- .\asi.exe (double-click launcher) or .\asi.cmd version or .\bin\asi.exe version
Start REPL and check runtime state:
- /status
Confirm selected provider key is set:
- OpenAI: OPENAI_API_KEY
- DeepSeek: DEEPSEEK_API_KEY
- Claude: ANTHROPIC_API_KEY

Windows ZIP: `asi.exe` Does Not Open

Fully extract the ZIP before running asi.exe.
If downloaded from browser, open file Properties and click Unblock if present.
If SmartScreen appears for unsigned beta builds, click More info -> Run anyway.
Prefer signed release artifacts for distribution to reduce SmartScreen friction.

Provider Errors

DeepSeek Model Not Exist:
1. Ensure you are on a recent build (provider/model mismatch now auto-falls back).
2. Verify provider/model via /status (or run explicitly with --provider deepseek --model deepseek-chat).
Claude authentication issues:
1. Verify ANTHROPIC_API_KEY or /oauth login <token> state.
2. Recheck with /status.

Local Validation Before Release

Run these checks before packaging:

cargo check --release
cargo test --release

Optional: Sign Windows Artifacts

$env:ASI_SIGN_CERT_PASSWORD="<pfx-password>"
powershell -ExecutionPolicy Bypass -File .\scripts\windows\package_release.ps1 -SignArtifacts -CertFile "C:\certs\asi-code.pfx" -TimestampUrl "http://timestamp.digicert.com"

If you use certificate store instead of PFX file:

powershell -ExecutionPolicy Bypass -File .\scripts\windows\package_release.ps1 -SignArtifacts -CertThumbprint "<sha1-thumbprint>"

Community and Feedback

Report bugs in the repository Issues tab (use the bug template).
Share product ideas and UX feedback in Discussions.
Contribution process: CONTRIBUTING.md
Security reporting: SECURITY.md

License

This project is licensed under the MIT License. See LICENSE.

Open-Core Model (Community + Commercial)

Community (MIT): local CLI core, REPL, provider integration, tool calls, and wiki workflows.
Commercial (planned): managed gateway, team workspace, policy packs, audit dashboard, and priority support.
Positioning: keep individual developer productivity open; monetize team/enterprise governance and reliability.

See: GITHUB_REPO_COPY_BILINGUAL.md

CLI x API Collaboration

Collaboration reality and upgrade blueprint (English): CLI_API_COLLABORATION_EN.md

Name		Name	Last commit message	Last commit date
.cargo		.cargo
.github		.github
docs		docs
scripts/windows		scripts/windows
src		src
.gitignore		.gitignore
CLI_API_COLLABORATION_EN.md		CLI_API_COLLABORATION_EN.md
CLI_API_COLLABORATION_ZH.md		CLI_API_COLLABORATION_ZH.md
CONTRIBUTING.md		CONTRIBUTING.md
Cargo.lock		Cargo.lock
Cargo.toml		Cargo.toml
GITHUB_REPO_COPY_BILINGUAL.md		GITHUB_REPO_COPY_BILINGUAL.md
INSTALL_WINDOWS.md		INSTALL_WINDOWS.md
LICENSE		LICENSE
README.md		README.md
RELEASE_NOTES_v0.3.0-beta.1.md		RELEASE_NOTES_v0.3.0-beta.1.md
SECURITY.md		SECURITY.md
start_asi.cmd		start_asi.cmd
start_asi.ps1		start_asi.ps1
start_asi.sh		start_asi.sh

Latest commit

History

Folders and files

Repository files navigation

ASI Code

Implemented Features

What's New in 2.0 (Preview)

Tier 1 — Native computer-control tools

Tier 2 — Live engine bridges for professional 3D/game software

Tier 3 — 3D-aware probes and temporal capture

Auto-loop no-progress detector — read-only work now counts

Cubist theme — Picasso-inspired terminal UI

Auto-compact — never hit max_turns_reached again

Test feedback loop — auto-run tests, auto-heal on failure

Auto-loop journal + resume — survive crashes mid-task

Roadmap

Quick Start

Real CLI Execution Example

Command

VS Code PowerShell Quick Launch (Windows)

Ubuntu/WSL Quick Launch

VS Code Workspace Behavior

Useful Commands

REPL Slash Commands

Prompt Output Formats

Agent JSON Schema (Stable v1)

Agent JSONL Event Stream (Stable v1)

Commands and Shapes

Quick Examples

MCP JSON Schema (Stable v1)

Agent Profiles (agent_profiles.json)

Commands and Shapes

Quick Examples

MCP Trust and Publish Metadata Policy

Plugin JSON Schema (Stable v1)

Plugin Trust and Publish Metadata Policy

Environment Variables

Multi-Provider Smoke Recipes (Windows)

1) DeepSeek-only API Compat (no OpenAI key required)

2) OpenAI-only API Compat (no DeepSeek key required)

3) Gateway Smoke by Provider (openai|deepseek|claude)

4) One-Command smoke_all with Provider-Scoped Gateway

Notes

5) Interactive Recipe Runner (OpenAI/DeepSeek)

CI Machine-Readable Template (Windows)

6) Subagent JSON Smoke (spawn/list/wait envelope checks)

Smoke Report Naming

Beta Known Issues

Install & Troubleshooting

Quick Checks

Windows ZIP: asi.exe Does Not Open

Provider Errors

Local Validation Before Release

Optional: Sign Windows Artifacts

Community and Feedback

License

Open-Core Model (Community + Commercial)

CLI x API Collaboration

About

Topics

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Auto-compact — never hit `max_turns_reached` again

Agent Profiles (`agent_profiles.json`)

4) One-Command `smoke_all` with Provider-Scoped Gateway

Windows ZIP: `asi.exe` Does Not Open