YAACLI CLI

TUI reference implementation for ya-agent-sdk.

Usage

Run with uvx:

uvx --from 'yaacli[rs]' yaacli

Install with uv:

uv tool install 'yaacli[rs]'
yaacli

Install a trusted capability plugin into the same isolated tool environment with --with:

uv tool install 'yaacli[rs]' --with acme-agent-plugin

[rs] installs the native Rust filesystem search binding. The equivalent extra-dependency form is:

uv tool install yaacli --with ya-ripgrep-core

ya-ripgrep-core is a library dependency, so --with is the matching uv form; --with-executables-from applies to companion packages that also expose CLI executables.

Update with uv:

uv tool upgrade yaacli

Subagent Configuration

YAACLI accepts both generic Markdown definitions and native versioned SubagentSpec YAML/JSON documents under ~/.yaacli/subagents/. In the interactive TUI, the visible delegate tool is processor-hosted and fully asynchronous: it immediately returns a readable <subagent-name>-bg-<short-id> handle. Resume, wait, steering, and cancellation reuse that handle; internal durable UUIDs are not model-facing. The one-shot headless frontend fixes the same processor-hosted tool to foreground mode and returns the child result before shutdown.

---
name: code-reviewer
description: Review a bounded change and report evidence-based findings.
instruction: Use this agent after meaningful implementation or refactoring.
model: inherit
model_settings: inherit
model_cfg: inherit
tools: [glob, grep, ls, view]
---

You are a senior code reviewer. Prioritize architecture, correctness, security, and
critical-path behavior.

Markdown is normalized into the current portable SubagentSpec; it does not activate a legacy runtime. The frontmatter description plus optional instruction form the parent-facing delegation roster entry; the Markdown body becomes the child's native instructions. Omitted or inherit model fields use the active root model, settings, and context configuration. YAACLI materializes its current standard child capability template for the common inherited-tool behavior, then turns any tool list into a final visibility allowlist. It does not copy live parent, MCP, plugin, delegation, or main-agent-only capabilities.

A same-basename Markdown file is the complete authoritative definition. Any native preset copied by an earlier upgrade is ignored rather than merged, so a stale capability snapshot cannot override the current standard child template. Other duplicate routes remain errors. New setup runs do not copy code-reviewer.yaml when code-reviewer.md already exists. Use native YAML/JSON by itself when exact capability grants, plugin types, nesting, or full delegation policy are required. See spec/02-configuration.md for both schemas.

Capability Plugins

YAACLI optionally loads the SDK's strict plugin manifest from the fixed global path ~/.yaacli/plugins.toml:

schema_version = 1
entry_points = ["acme.search"]

[[capabilities]]
name = "acme.search"
arguments = { result_limit = 10 }

Package installation and authorization are separate. The distribution must be installed in YAACLI's own Python environment; entry_points then selects the exact installed types YAACLI may import, and capabilities grants ordered instances to the main root agent. YAACLI never scans or loads every installed entry point.

The file is global-only because selected Python code executes with YAACLI process authority. A project .yaacli/plugins.toml is ignored. A missing global file means no external plugins; invalid TOML, unknown fields, missing or duplicate entry points, import failures, grant argument signature mismatches, and catalog collisions stop startup. Manifest arguments are durable non-secret configuration, so secret-like keys such as API keys, tokens, passwords, or credentials are rejected recursively. This name-based guard cannot detect a secret stored under a neutral key; keep every secret value outside the manifest.

Selected types are available to native subagent documents, but root grants do not implicitly enter named children or self forks. A child must declare the selected serialization name in its own agent.capabilities. TUI and headless startup each load one catalog snapshot and reuse it for current, historical, child, and restored runtime construction. Restart YAACLI after changing the manifest or installed distribution.

See the SDK file configuration contract and the installable example.

Install with pip:

pip install 'yaacli[rs]'
yaacli

Run as a module:

python -m yaacli

Headless and Saved Sessions

Run one prompt without the TUI:

yaacli -p "Fix the failing tests"
yaacli -p "Continue" --session <session-id> --profile <profile-id>
yaacli -p "Run an isolated worker task" --worker

Headless stdout is an NDJSON event stream. Human-readable diagnostics, fatal details, and resume hints are written to stderr so scripts can parse every stdout line as JSON. Every successful, failed, or cancelled logical run commits a terminal durable revision. --worker requires --prompt and disables delegate subagents for that run. --profile applies only to the current invocation; selecting a profile through /model persists it for future launches.

Inspect or delete durable sessions without starting the TUI:

yaacli sessions list
yaacli sessions show <session-id>
yaacli sessions delete <session-id>

Session IDs may be supplied by unique prefix. The YAACLI 2 product store defaults to ~/.yaacli/sessions/sessions-v2.sqlite3; the former sessions.sqlite3 is left untouched and is not migrated automatically. An explicit [session] database_path or YAACLI_DATABASE_PATH remains authoritative and is validated strictly. See spec/05-session-persistence.md.

TUI Interaction

The output viewport has priority over auxiliary UI. The task pane is hidden when empty, uses one summary row by default, and expands with F2. The model selector is an overlay and does not permanently consume output rows.

Enter submits while idle and sends guidance to the active run while an agent is running.
Ordinary text submitted during an active agent run is durable steering. YAACLI persists the original text before displaying a replayable Guidance sent to the active run. receipt and waking the owning local execution task. The native enqueue boundary applies the established model-facing steering reminder envelope rather than treating the text as a new ordinary prompt; native application later displays a distinct replayable Guidance injected projection. This uses native Pydantic AI enqueue and does not restore MessageBus. The status bar counts accepted/enqueued inputs without exposing their content. Registered slash commands and !shell remain local control syntax: safe busy commands execute and idle-only commands are rejected without clearing the draft. Other slash-prefixed text, including absolute paths such as /home/user/file, remains ordinary user input.
/cancel or Ctrl+C requests cancellation of cancellable foreground work. Once the TUI enters SAVING, persistence is allowed to finish and cannot be cancelled; Ctrl+C does not exit while the save is in progress.
/clear clears only the visible transcript; /new starts a fresh conversation and session. Tombstoning the previous session atomically closes its main and child input, persists cancellation intent for every nonterminal child, and retries process-local cancellation dispatch; late child success, steering, and completion delivery are fenced while the runtime environment remains reusable.
Background results never modify or clear the compose area. A terminal background subagent is projected as session-scoped readiness only; its committed result enters the canonical parent continuation path on the next accepting agent turn and does not automatically wake the model. Monitored-shell notifications are persisted as feature input and may start one idle turn according to shell-monitor policy. There is no /integrate command or MessageBus delivery path.
/agents shows running and recently completed background subagents; /process shows active background shell processes.
/attachments and /remove-image inspect or edit images queued for the next turn.
/tool <call-id> shows the complete retained result for a tool call.
/session <id> restores a saved session. The CLI equivalent is yaacli --session <id>.
Use /help for the complete built-in and configured command list. Slash commands and available skills complete while typing.
Prefix an idle prompt with one or more available skill names to request them explicitly: /lark-cli /agent-builder Build an agent that replies in Lark. Only the leading consecutive /skill-name tokens are selected. If the first token is also a built-in or configured command, command dispatch takes precedence.
The interactive TUI enables ask_user_question by default. When the agent needs clarification, YAACLI renders one to four structured questions and accepts an option number, comma-separated numbers for multi-select questions, or free text.
Long status text wraps to the available terminal width. Foreground elapsed time uses compact forms such as 42s, 3m 05s, and 1h 05m 09s.

Built-in Skills

YAACLI ships with building-agents from the repository canonical source skills/agent-builder/.

The YA Claw deployment skill lives in skills/ya-claw-deploy/ and is published as YA_CLAW_DEPLOY_SKILL.zip during release.

The repository sync script keeps bundled skill files under packages/yaacli/yaacli/skills/ aligned.

YAACLI refreshes and resolves /skill-name against the effective SDK skill catalog at submission time, including built-in, global, shared, and project skills after normal priority rules. The visible transcript and prompt history retain the original input; the model receives a catalog-grounded explicit-selection marker plus the remaining task. A slash prefix that matches neither a registered command nor an available skill is submitted as ordinary user input.

CodeAct

YAACLI exposes run_code and run_program by default. They execute restricted Python for tool orchestration; run_program reads reviewed programs through the runtime's FileOperator. Shell remains a separate execution surface and is not made callable from CodeAct. Disable both CodeAct tools globally in ~/.yaacli/tools.toml or per project in .yaacli/tools.toml:

[tools]
enable_codeact = false

MCP Tool Exposure

MCP tools are exposed directly to the model by default. Configure the behavior globally in ~/.yaacli/tools.toml or per project in .yaacli/tools.toml:

[tools]
mcp_mode = "direct" # "direct" (default) or "proxy"

Direct mode registers each configured MCP server as a native toolset and exposes namespaced <server>_<tool> names by default. A server's optional prefix field in mcp.json overrides <server>; set it to "" to expose the server's native tool names without a prefix. Omitting prefix or setting it to null preserves the default server-name prefix. Proxy mode exposes the fixed mcp_search_tool and mcp_call_tool pair instead, which can improve prompt-cache stability when many MCP tools are configured. Servers marked "required": false remain optional in both modes. If an optional server cannot connect or becomes unavailable, the TUI reports the degraded MCP namespace once per status change and continues without that server.

For direct host-managed MCP tools, structured output remains the tool's Python return value even when the response also contains media. Completed MCP error responses are returned for model or CodeAct inspection without consuming Pydantic AI's ModelRetry budget; failures that produce no MCP result are terminal failed tool outcomes.

{
  "servers": {
    "docs": {
      "transport": "streamable_http",
      "url": "https://example.com/mcp",
      "prefix": "reference"
    },
    "local": {
      "transport": "stdio",
      "command": "local-mcp-server",
      "prefix": ""
    }
  }
}

Structured User Input

The interactive TUI opts into the SDK's deferred ask_user_question tool. Disable it globally in ~/.yaacli/tools.toml or for one project in .yaacli/tools.toml:

[tools]
enable_user_input = true
user_input_timeout_seconds = 120

enable_user_input defaults to true. Each structured question waits up to user_input_timeout_seconds, which defaults to 120 seconds. If no answer arrives, YAACLI rejects the deferred call with an English model prompt that explains the timeout and directs the agent to continue using its best judgment rather than requesting the same input again. Set enable_user_input = false to remove the tool entirely. Headless mode does not expose this tool because it cannot collect interactive answers. The SDK also leaves it disabled unless a host explicitly registers it and implements deferred continuation. Project tools.toml replaces the global tool policy as a whole; if a project file exists, repeat any non-default user-input settings there rather than relying on global values.

Development

This package lives in the ya-mono workspace.

git clone git@github.com:YOUR_NAME/ya-mono.git
cd ya-mono
uv sync --all-packages
cp packages/yaacli/.env.example packages/yaacli/.env

YAACLI loads .env from packages/yaacli/.env and the current working directory without replacing variables already present in the process. The package file is loaded first and therefore wins duplicate keys; the working-directory file supplies only keys that remain unset. Provider API keys can live in that .env file or in ~/.yaacli/config.toml under [env]. SDK and tool variables such as YA_AGENT_* and search API keys can also live in that same .env file because YAACLI loads it into the process environment at startup. Use packages/ya-agent-sdk/.env.example as the reference list for SDK and tool variables.

The TUI detects the terminal's light or dark background at startup. For recognized local terminals such as VS Code's integrated terminal, it performs a short OSC 11 query; active queries are skipped over SSH to avoid delayed terminal responses. Detection then falls back to COLORFGBG, followed by the dark theme. Override detection in ~/.yaacli/config.toml when needed:

[display]
code_theme = "auto" # auto, dark, or light

The equivalent environment override is YAACLI_CODE_THEME=auto.

Terminal Bells

YAACLI emits a terminal bell by default after a successful interactive agent turn and whenever a turn pauses for HITL user input. Either notification can be disabled independently:

[notifications]
bell_on_turn_complete = false
bell_on_user_action_required = false

A run's elapsed timer is paused for the entire HITL wait and resumes after all requested user actions have been answered. The bell is emitted by the terminal, so it passes through SSH and is handled by the local terminal client. In VS Code Remote, configure the local VS Code User settings (not Remote or workspace settings) to show the visual bell and always play its terminal-bell accessibility signal:

{
  "terminal.integrated.enableVisualBell": true,
  "terminal.integrated.bellDuration": 1000,
  "accessibility.signals.terminalBell": {
    "sound": "on",
    "announcement": "off"
  }
}

Use "sound": "on" for unconditional playback; "auto" lets VS Code decide whether to play the accessibility signal. announcement controls screen-reader output and can remain "auto" when needed. VS Code plays its own accessibility sound through the local audio output; this is not an operating-system desktop notification.

To verify terminal support independently of YAACLI, run printf '\a' in the same integrated terminal. It should show VS Code's visual-bell indicator. Remote YAACLI processes cannot directly invoke the client machine's operating-system notification API.

Codex OAuth credentials can be created once and reused from YAACLI:

uvx ya-oauth login codex

Then set model = "oauth@codex:gpt-5.5" in a YAACLI model profile.

Model profiles are configured in ~/.yaacli/config.toml and selected with /model inside the TUI:

[general]
model = "anthropic:claude-sonnet-4-5"
model_settings = "anthropic_adaptive_high"
model_cfg = "claude_200k"
instructions = """
Prefer careful analysis and concise answers.
"""

[model_profiles.fast]
label = "Fast"
model = "openai-responses:gpt-5.6-luna"
model_settings = "openai_responses_luna"
model_cfg = "gpt5_270k"
instructions = """
Optimize for speed. Avoid broad exploration unless necessary.
"""

[model_profiles.pro]
label = "GPT-5.6 Pro"
model = "openai-responses:gpt-5.6"
model_settings = "openai_responses_pro"
model_cfg = "gpt5_270k"

[model_profiles.sol]
label = "GPT-5.6 Sol"
model = "openai-responses:gpt-5.6-sol"
model_settings = "openai_responses_max"
model_cfg = "gpt5_270k"

[model_profiles.codex_oauth]
label = "Codex OAuth"
model = "oauth@codex:gpt-5.5"
model_settings = "openai_responses_high"
model_cfg = "gpt5_350k"

instructions is an optional static model-instruction segment for the active main-agent profile. YAACLI evaluates it for every model request, alongside the built-in prompt (or general.system_prompt_file), project guidance, and user rules. Switching profiles with /model replaces the active profile instructions for later requests in the same session, including restored or compacted histories.

[general] is the startup fallback profile. The last selected profile is remembered in ~/.yaacli/state.json and restored on the next launch when that profile still exists.

Shell command review is configured in ~/.yaacli/config.toml under security.shell_review:

[security.shell_review]
enabled = true
model = "gateway@openai-responses:gpt-5.4-mini"
model_settings = "openai_responses_low"
on_needs_approval = "defer"
risk_threshold = "high"

When enabled, model is required. model_settings accepts SDK preset names or an inline TOML table. risk_threshold defaults to high and controls when the configured action triggers.

Run CLI tests from the workspace root:

make test-cli

Clipboard Image Paste

Plain terminal paste always inserts text into the input box. Use Ctrl+V or /paste-image to attach an image from the system clipboard. During an active agent run the image remains queued for the next turn and is never converted into steering text. Generated attachment chips are removed before registered commands, explicit skills, ! control syntax, or ordinary prompts are classified, so a visible chip cannot hide a command. If the user deleted the chip, its binary is removed before dispatch. On macOS terminal apps over SSH, map Command+Shift+V to send Ctrl+V if you want a native-feeling shortcut.

YAACLI reads clipboard images through Pillow first on macOS and Windows. macOS also reads Finder-copied image files through Cocoa pasteboard APIs via pyobjc-framework-Cocoa. Linux image paste still relies on wl-paste on Wayland or xclip on X11.

License

BSD 3-Clause License. See the repository license.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

yaacli-2.0.1.tar.gz (394.8 kB view details)

Uploaded Aug 21, 2026 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

yaacli-2.0.1-py3-none-any.whl (317.6 kB view details)

Uploaded Aug 21, 2026 Python 3

File details

Details for the file yaacli-2.0.1.tar.gz.

File metadata

Download URL: yaacli-2.0.1.tar.gz
Upload date: Aug 21, 2026
Size: 394.8 kB
Tags: Source
Uploaded using Trusted Publishing? No
Uploaded via: uv/0.11.14 {"installer":{"name":"uv","version":"0.11.14","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

File hashes

Hashes for yaacli-2.0.1.tar.gz
Algorithm	Hash digest
SHA256	`ad871fc3a0ed20ddabe1ead2b32f989e3bd3a7424a34ed98d357fe2be182573b`
MD5	`76752a17a468dda78849347d92ce619c`
BLAKE2b-256	`69ffa65a4153d44895988cbd9b37c95c99de568d06e211f0d6394b9b04a753a9`

See more details on using hashes here.

File details

Details for the file yaacli-2.0.1-py3-none-any.whl.

File metadata

Download URL: yaacli-2.0.1-py3-none-any.whl
Upload date: Aug 21, 2026
Size: 317.6 kB
Tags: Python 3
Uploaded using Trusted Publishing? No
Uploaded via: uv/0.11.14 {"installer":{"name":"uv","version":"0.11.14","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

File hashes

Hashes for yaacli-2.0.1-py3-none-any.whl
Algorithm	Hash digest
SHA256	`700ae92bd0cec39a3f72d7804eecc26622bf10f1ab9cd8a3f173112126abb787`
MD5	`0741db913fc09c8cf41049ae867d4ab9`
BLAKE2b-256	`4a850c5e1283227ff178d1299e8f648d6e2a3446d627256a8afbf9c03eee7bcd`