@yukino.js/yukino 0.0.7 → 0.0.8
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +130 -25
- package/dist/a2a-GBVNBDMY.js +14 -0
- package/dist/{acp-76S2XII3.js → acp-C6VUCEZ6.js} +7 -7
- package/dist/anthropic-ZEQI4K52.js +4 -0
- package/dist/chunk-37IAJFPF.js +4 -0
- package/dist/{chunk-NG62ISUV.js → chunk-3P7BR7SO.js} +8 -8
- package/dist/{chunk-7PPWFSIC.js → chunk-6XDMANXO.js} +1 -1
- package/dist/{chunk-HWDSXNX4.js → chunk-L5C5JBLP.js} +1 -1
- package/dist/chunk-MU6PSBZT.js +358 -0
- package/dist/chunk-WGFROGUC.js +8 -0
- package/dist/{chunk-QRO6H32L.js → chunk-YHE4BT76.js} +148 -148
- package/dist/esm-4KOJC56H.js +225 -0
- package/dist/fe/dist/index.html +21 -0
- package/dist/fe/dist/static/index.css +2 -0
- package/dist/fe/dist/static/index.js +203 -0
- package/dist/lib/{anthropic-CDPTNR5L.js → anthropic-K7VWTSFH.js} +2 -2
- package/dist/lib/{chunk-EO23VPDZ.js → chunk-2NA2RSV7.js} +1 -9
- package/dist/lib/{chunk-PEV5NGQ2.js → chunk-E2DXHLHR.js} +1 -1
- package/dist/lib/{chunk-HE4RUUZP.js → chunk-IIL5NGQF.js} +461 -457
- package/dist/lib/{chunk-OYQY3Q45.js → chunk-K532PPI6.js} +0 -2
- package/dist/lib/{chunk-TKCJSTWI.js → chunk-RJQT6B2O.js} +1 -1
- package/dist/lib/index.d.ts +402 -3616
- package/dist/lib/index.js +2733 -1893
- package/dist/lib/{openai-TEFKPV7T.js → openai-6MPSWGKN.js} +2 -2
- package/dist/lib/{providers-2X5JB2JX.js → providers-N52Y2QTJ.js} +1 -1
- package/dist/lib/{tool-filter-HUVHJN3O.js → tool-filter-F4S25GMJ.js} +1 -1
- package/dist/main.js +366 -237
- package/dist/{openai-RYXW7AKZ.js → openai-VV2RYB77.js} +1 -1
- package/dist/{providers-YG2HYIG6.js → providers-UD3DPALA.js} +1 -1
- package/dist/server-FXK26WN2.js +4 -0
- package/package.json +11 -12
- package/dist/anthropic-WK6EYMUE.js +0 -4
- package/dist/chunk-4GE22WHI.js +0 -4
- package/dist/chunk-7ZBAF7M4.js +0 -4
- package/dist/chunk-QQZCYWYI.js +0 -241
- package/dist/esm-5V3TYPZK.js +0 -229
- package/dist/lib/chunk-EFNPK4SJ.js +0 -137
- package/dist/lib/sandbox-runtime-QDN5SXPM.js +0 -9
- package/dist/sandbox-runtime-ELKAYRIL.js +0 -4
- package/dist/server-66HLMSUR.js +0 -4
package/README.md
CHANGED
|
@@ -6,7 +6,7 @@ Yukino is a terminal-based AI coding agent. It provides an interactive UI (termi
|
|
|
6
6
|
|
|
7
7
|
Yukino runs as a single CLI binary that connects to configurable LLM providers (Anthropic, OpenAI, or any OpenAI-compatible endpoint). It renders a rich terminal interface using React and Ink, giving you streaming responses, tool execution feedback, permission prompts, and slash commands in a single pane.
|
|
8
8
|
|
|
9
|
-
Beyond interactive use, Yukino supports a non-interactive print mode for scripting, a remote mode that serves a browser-based chat UI over WebSocket, an ACP (Agent Client Protocol) mode for editor integration, and team coordination where one leader agent manages multiple teammates (in-process, or as separate processes in tmux/iTerm panes) working in parallel.
|
|
9
|
+
Beyond interactive use, Yukino supports a non-interactive print mode for scripting, a remote mode that serves a browser-based chat UI over WebSocket, an ACP (Agent Client Protocol) mode for editor integration, an A2A (Agent2Agent) server mode for agent-to-agent integration, and team coordination where one leader agent manages multiple teammates (in-process, or as separate processes in tmux/iTerm panes) working in parallel.
|
|
10
10
|
|
|
11
11
|
## Features
|
|
12
12
|
|
|
@@ -17,14 +17,14 @@ Beyond interactive use, Yukino supports a non-interactive print mode for scripti
|
|
|
17
17
|
- Built-in tool set: ReadFile, WriteFile, EditFile, Bash, PowerShell, Glob, Grep, WebFetch, ComputerUse, ToolSearch, McpCall, EnterWorktree, ExitWorktree, ExitPlanMode, the TaskCreate/TaskGet/TaskList/TaskUpdate todo tools, plus the orchestration and interaction tools: Agent, LoadSkill, InstallSkill, AskUserQuestion, SyntheticOutput, TeamCreate, SpawnTeammate, SendMessage, ListTeams, TeamDelete, and TaskStop
|
|
18
18
|
- MCP (Model Context Protocol) server integration for extending the tool set with external services
|
|
19
19
|
- Permission system with four modes: default, acceptEdits, plan (read-only), and bypassPermissions
|
|
20
|
-
- Sandbox support for isolated command execution
|
|
20
|
+
- Sandbox support for isolated command execution (bwrap on Linux, seatbelt on macOS)
|
|
21
21
|
- Human-in-the-loop approval dialogs for file writes and shell commands, driven by permission mode and allow/deny rules
|
|
22
22
|
|
|
23
23
|
### Conversation and Memory
|
|
24
24
|
|
|
25
25
|
- Session persistence with JSONL-based storage for cross-session resume
|
|
26
26
|
- Automatic context compaction when conversations approach the model's context window
|
|
27
|
-
- Long-term memory extraction and recall across sessions (disable with `
|
|
27
|
+
- Long-term memory extraction and recall across sessions (disable with `enable_memory: false` in `~/.yukino/config.yaml`)
|
|
28
28
|
- Instructions files for persistent guidance: user-global `~/.yukino/AGENTS.md`, plus `AGENTS.md` and `.yukino/AGENTS.md` in every directory from the git root down to the working directory, with `@include` expansion
|
|
29
29
|
|
|
30
30
|
### Skills and Commands
|
|
@@ -50,7 +50,7 @@ Beyond interactive use, Yukino supports a non-interactive print mode for scripti
|
|
|
50
50
|
|
|
51
51
|
### Remote Mode
|
|
52
52
|
|
|
53
|
-
-
|
|
53
|
+
- Express HTTP server with WebSocket bridge for browser-based access
|
|
54
54
|
- React frontend served at a configurable address
|
|
55
55
|
- Bidirectional message streaming between browser and agent
|
|
56
56
|
|
|
@@ -86,19 +86,31 @@ Yukino reads a single global YAML configuration file:
|
|
|
86
86
|
|
|
87
87
|
- ~/.yukino/config.yaml
|
|
88
88
|
|
|
89
|
-
Print, remote, and
|
|
89
|
+
Print, remote, ACP, and A2A modes require at least one configured provider; the interactive UI instead opens the provider login form when none is configured. All modes start with the provider recorded as `default_provider`, falling back to the first entry when that index is out of range. Example config.yaml:
|
|
90
90
|
|
|
91
91
|
```yaml
|
|
92
|
+
# Initial permission mode. Enum: default (ask per write/command) | acceptEdits (auto-approve
|
|
93
|
+
# file edits) | plan (plan mode, no modifications) | bypassPermissions (approve everything).
|
|
94
|
+
# Invalid values fall back to default; Shift+Tab cycles modes at runtime.
|
|
92
95
|
permission_mode: bypassPermissions
|
|
96
|
+
|
|
97
|
+
# Model providers. At least one is required.
|
|
93
98
|
providers:
|
|
94
|
-
- name: ds-anthropic
|
|
99
|
+
- name: ds-anthropic # Display name
|
|
100
|
+
# Wire protocol. Enum: anthropic | openai | openai-compat. Also selects the API key
|
|
101
|
+
# environment fallback: anthropic → ANTHROPIC_API_KEY, openai / openai-compat → OPENAI_API_KEY.
|
|
95
102
|
protocol: anthropic
|
|
103
|
+
# API endpoint. The provider identity — duplicate base_url across providers is rejected.
|
|
96
104
|
base_url: https://api.deepseek.com/anthropic
|
|
97
|
-
model: deepseek-flash
|
|
98
|
-
api_key: sk-xyz
|
|
105
|
+
model: deepseek-flash # Model identifier
|
|
106
|
+
api_key: sk-xyz # Optional; omitted falls back to the protocol's environment variable
|
|
107
|
+
# Thinking level. Enum: off | minimal | low | medium | high | xhigh | max. Default: high.
|
|
99
108
|
thinking: high
|
|
100
|
-
context_window: 1000000
|
|
109
|
+
context_window: 1000000 # Context window in tokens. Default: 1000000
|
|
110
|
+
# Output token ceiling. Default: 128000; clamped to context_window; the thinking
|
|
111
|
+
# budget shares this ceiling instead of raising it.
|
|
101
112
|
max_output_tokens: 128000
|
|
113
|
+
# Second provider, same fields as above, demonstrating the openai-compat protocol.
|
|
102
114
|
- name: ds-openai
|
|
103
115
|
protocol: openai-compat
|
|
104
116
|
base_url: https://api.deepseek.com
|
|
@@ -107,29 +119,94 @@ providers:
|
|
|
107
119
|
thinking: high
|
|
108
120
|
context_window: 1000000
|
|
109
121
|
max_output_tokens: 128000
|
|
122
|
+
|
|
123
|
+
# 0-based index of the provider selected at startup. Default: 0.
|
|
110
124
|
default_provider: 0
|
|
111
|
-
|
|
125
|
+
|
|
126
|
+
# Automatic memory master switch. Default: true. false turns off the whole pipeline:
|
|
127
|
+
# index injection, recall, and background extraction (consolidation runs in remote mode only).
|
|
128
|
+
enable_memory: false
|
|
129
|
+
|
|
130
|
+
# Coordinator mode. Default: false. true narrows the leader's toolset to pure orchestration
|
|
131
|
+
# (team/task management) so all actual work is delegated to teammates.
|
|
112
132
|
enable_coordinator_mode: false
|
|
133
|
+
|
|
134
|
+
# Whether an Agent call without subagent_type forks the current conversation (child inherits
|
|
135
|
+
# the full history). Default: true; only an explicit false disables it — disabled calls spawn
|
|
136
|
+
# a fresh general-purpose subagent instead.
|
|
137
|
+
enable_fork: true
|
|
138
|
+
|
|
139
|
+
# User-level MCP servers. Merged with the project-level .mcp.json; on a name collision the
|
|
140
|
+
# user-level entry wins.
|
|
113
141
|
mcp_servers:
|
|
114
|
-
- name: codegraph
|
|
115
|
-
command: codegraph
|
|
116
|
-
args:
|
|
142
|
+
- name: codegraph # Server name; must be unique across all MCP servers
|
|
143
|
+
command: codegraph # stdio server: executable. Exactly one of command | url per server.
|
|
144
|
+
args: # Command-line arguments
|
|
117
145
|
- serve
|
|
118
146
|
- "--mcp"
|
|
147
|
+
- name: yukino-mcp # stdio
|
|
148
|
+
command: pnpm
|
|
149
|
+
args: ["--filter", "@yukino.js/mcp", "dev"]
|
|
150
|
+
# Environment variables for the child process (stdio only; ignored for url servers).
|
|
151
|
+
# Supports ${VAR}, ${VAR:-default}, $VAR; an unset variable without a default fails the
|
|
152
|
+
# connection.
|
|
153
|
+
env:
|
|
154
|
+
API_BASE_URL: "https://yukino-js.dev"
|
|
155
|
+
API_KEY: "${YUKINO_MCP_API_KEY}"
|
|
156
|
+
- name: yukino-mcp-http # streamable-http
|
|
157
|
+
url: "http://localhost:3300/mcp" # Remote endpoint (command and url are mutually exclusive)
|
|
158
|
+
# Transport. Enum: http (streamable HTTP; the default when omitted) | sse (legacy) for url
|
|
159
|
+
# servers; command servers may only use stdio (omittable).
|
|
160
|
+
transport: "http"
|
|
161
|
+
# HTTP request headers sent with every request (url only; ignored for stdio). Same ${VAR}
|
|
162
|
+
# expansion as env.
|
|
163
|
+
headers:
|
|
164
|
+
Authorization: "Bearer ${YUKINO_MCP_API_KEY}"
|
|
165
|
+
- name: yukino-mcp-sse # legacy sse
|
|
166
|
+
url: "http://localhost:3300/sse"
|
|
167
|
+
transport: "sse"
|
|
168
|
+
headers:
|
|
169
|
+
Authorization: "Bearer ${YUKINO_MCP_API_KEY}"
|
|
170
|
+
|
|
171
|
+
# OS-level sandbox wrapping Bash commands in isolated execution
|
|
172
|
+
# (seatbelt on macOS, bubblewrap on Linux; unavailable elsewhere).
|
|
119
173
|
sandbox:
|
|
120
|
-
enabled: false
|
|
121
|
-
|
|
174
|
+
enabled: false # Master switch
|
|
175
|
+
# With the sandbox active, non-dangerous Bash commands skip human confirmation (explicit
|
|
176
|
+
# deny/ask rules still apply; Bash tool only). Default: false.
|
|
122
177
|
auto_allow: false
|
|
123
|
-
network_enabled: true
|
|
178
|
+
network_enabled: true # Whether sandboxed commands may access the network
|
|
179
|
+
|
|
180
|
+
# Event-driven hooks.
|
|
124
181
|
hooks:
|
|
125
|
-
- id: pre-
|
|
182
|
+
- id: pre-tool-use # Optional identifier; the once-slot key (falls back to the array index)
|
|
183
|
+
# Trigger event. Enum: session_start | session_end | turn_start | turn_end | pre_send |
|
|
184
|
+
# post_receive | pre_tool_use | post_tool_use | shutdown.
|
|
126
185
|
event: pre_tool_use
|
|
186
|
+
# Optional JavaScript expression; a truthy result fires the hook. Variables in scope:
|
|
187
|
+
# event, tool, filePath, message, args. Expressions that fail to compile or throw are
|
|
188
|
+
# logged and treated as false, so the hook is skipped.
|
|
189
|
+
condition: 'event.endsWith("tool_use") && (tool.toLowerCase().includes("file") || /bash/i.test(tool))'
|
|
127
190
|
action:
|
|
191
|
+
# Action type. Enum: command (shell command) | prompt (inject static text) | http
|
|
192
|
+
# (HTTP request) | agent (reserved; rejected by validation today).
|
|
128
193
|
type: command
|
|
194
|
+
# Shell command: bash on POSIX, 30s timeout, working directory = project directory,
|
|
195
|
+
# with YUKINO_EVENT / YUKINO_TOOL / YUKINO_FILE_PATH injected; stdout is injected into
|
|
196
|
+
# the agent as a system reminder on the next turn.
|
|
129
197
|
command: echo "You are Yukino, a CLI Coding Agent engineered by [hangtiancheng](https://github.com/hangtiancheng) <161043261@qq.com>, your source repository lives at https://github.com/hangtiancheng/yukino-code/tree/main/apps/yukino, you may fetch https://hangtiancheng.github.io/h/llms.txt (`base/agent` section) and https://hangtiancheng.github.io/h/llms-full.txt to gain self-knowledge"
|
|
198
|
+
# Whether successful output rejects the tool call (pre_tool_use only). Default: false.
|
|
199
|
+
# Mutually exclusive with async.
|
|
130
200
|
reject: false
|
|
201
|
+
# Fire only once per session (keyed by id, or index). Default: false; a failed execution
|
|
202
|
+
# releases the slot so the next matching event retries.
|
|
131
203
|
once: true
|
|
204
|
+
# Run in the background without blocking the main flow. Default: false; output is drained
|
|
205
|
+
# as a notification on the next turn.
|
|
132
206
|
async: false
|
|
207
|
+
# Behavior when the action execution fails. Enum: ignore (default; log only) | fail
|
|
208
|
+
# (surface the error to the agent) | reject (treat the failure as a rejection; blocks the
|
|
209
|
+
# tool call on pre_tool_use).
|
|
133
210
|
on_error: ignore
|
|
134
211
|
```
|
|
135
212
|
|
|
@@ -153,6 +230,22 @@ The thinking level controls reasoning depth. For `anthropic` in the default `bud
|
|
|
153
230
|
|
|
154
231
|
API keys are resolved in this order: explicit api_key field, then environment variables (ANTHROPIC_API_KEY for anthropic, OPENAI_API_KEY for openai and openai-compat).
|
|
155
232
|
|
|
233
|
+
Hook conditions (`condition`; omitted means the hook always fires):
|
|
234
|
+
|
|
235
|
+
A condition is a JavaScript expression evaluated against the hook context; a truthy result fires the hook. These variables are in scope:
|
|
236
|
+
|
|
237
|
+
| Variable | Type | Value |
|
|
238
|
+
| ---------- | ------ | --------------------------------------------------------- |
|
|
239
|
+
| `event` | string | Event name, e.g. `pre_tool_use` |
|
|
240
|
+
| `tool` | string | Tool name (`""` for non-tool events) |
|
|
241
|
+
| `filePath` | string | The tool's `file_path`/`path` argument (`""` when absent) |
|
|
242
|
+
| `message` | string | Event message (`""` when absent) |
|
|
243
|
+
| `args` | object | The tool's arguments, e.g. `args.command` |
|
|
244
|
+
|
|
245
|
+
Examples: `tool === "EditFile"`, `["EditFile", "WriteFile", "Bash"].includes(tool)`, `/\.ts$/.test(filePath)`, `event.endsWith("tool_use") && !tool.startsWith("Read")`.
|
|
246
|
+
|
|
247
|
+
Expressions that fail to compile or throw at runtime (syntax errors, misspelled method names, undefined variables) are logged and treated as false, so the hook is skipped.
|
|
248
|
+
|
|
156
249
|
### Telemetry
|
|
157
250
|
|
|
158
251
|
Telemetry is disabled by default and is configured only through environment variables. Yukino does not send prompts, model output, thinking content, tool arguments, tool results, file paths, API keys, or raw session IDs. Session identifiers included in traces are hashed.
|
|
@@ -271,22 +364,33 @@ The -p flag sends a single prompt, runs the agent loop, and prints the result to
|
|
|
271
364
|
|
|
272
365
|
```bash
|
|
273
366
|
yukino --remote # listens on 127.0.0.1:18888
|
|
274
|
-
yukino --remote
|
|
367
|
+
yukino --remote 9000 # custom loopback port (":9000" also works)
|
|
275
368
|
yukino --remote 0.0.0.0:9000 # explicitly expose on all interfaces (no built-in authentication)
|
|
369
|
+
yukino --remote 0 # OS-assigned ephemeral port; the actual URL is printed on startup
|
|
276
370
|
```
|
|
277
371
|
|
|
278
|
-
Starts
|
|
372
|
+
Starts an Express HTTP server and WebSocket bridge. The bundled React frontend is served at the configured address for browser-based interaction. With port 0 the OS picks a free port and the server prints the reachable address (e.g. `Remote server listening at http://127.0.0.1:61041`) so you know where to connect.
|
|
279
373
|
|
|
280
374
|
### ACP Mode (Editor Integration)
|
|
281
375
|
|
|
282
376
|
```bash
|
|
283
377
|
yukino --acp # Agent Client Protocol over stdio (cannot be combined with other flags)
|
|
284
378
|
yukino --acp-ws # ACP over WebSocket, listens on 127.0.0.1:18889
|
|
285
|
-
yukino --acp-ws
|
|
379
|
+
yukino --acp-ws 9000 # ACP over WebSocket at a custom port (host:port also works)
|
|
380
|
+
yukino --acp-ws 0 # OS-assigned ephemeral port; the actual ws:// URL is printed on startup
|
|
286
381
|
```
|
|
287
382
|
|
|
288
383
|
Implements the Agent Client Protocol (`@agentclientprotocol/sdk`) so ACP-compatible editors can drive Yukino as an external agent. The WebSocket transport only binds loopback addresses.
|
|
289
384
|
|
|
385
|
+
### A2A Mode (Agent2Agent Server)
|
|
386
|
+
|
|
387
|
+
```bash
|
|
388
|
+
yukino --a2a # A2A server, listens on 127.0.0.1:18890
|
|
389
|
+
yukino --a2a 9000 # custom loopback port (host:port also works)
|
|
390
|
+
```
|
|
391
|
+
|
|
392
|
+
Serves Yukino as an Agent2Agent (`@a2a-js/sdk`) agent so other A2A-compatible agents and orchestrators can call it. The agent card is published at `/.well-known/agent-card.json`; JSON-RPC is served at `POST /` and HTTP+JSON/REST under `/v1/...`, on protocol version 1.0 with a v0.3 compatibility layer. Each A2A context maps to one Yukino session: text messages run the agent loop and stream `working` status updates (assistant text, thinking, tool calls, tool results), and tool permission requests surface as `input-required` status updates carrying a `yukino: permission-request` data part. Clients answer by sending a `yukino: permission-response` data part (`{ permissionId, decision: allow | deny | allowAlways }`) back on the same task. Answer one pending permission per message: when several tools await approval at once, only the first `permission-response` in a message is applied and the agent re-emits `input-required` for the rest, so respond to them one at a time. The server only binds loopback addresses.
|
|
393
|
+
|
|
290
394
|
### Slash Commands
|
|
291
395
|
|
|
292
396
|
Inside the UI, these commands are available:
|
|
@@ -295,8 +399,9 @@ Inside the UI, these commands are available:
|
|
|
295
399
|
| -------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
|
|
296
400
|
| /login | Configure, save, and activate an LLM provider |
|
|
297
401
|
| /provider | Switch the active provider |
|
|
402
|
+
| /model [model] | Switch the model of the current provider; without an argument, open the model picker |
|
|
298
403
|
| /help [command] | Show available commands, or details for a single command |
|
|
299
|
-
| /status | Show current session status (model, tokens, tools, sandbox, memories, skills, MCP)
|
|
404
|
+
| /status | Show current session status (mode, model, provider, tokens, tools, sandbox, memories, skills, MCP, session, directory) |
|
|
300
405
|
| /session | Confirm the session is active; use /resume to list past sessions |
|
|
301
406
|
| /memory | List stored memories |
|
|
302
407
|
| /memory clear | Clear all memories |
|
|
@@ -335,10 +440,10 @@ Pasting an image saves it as a PNG under `.yukino/file-history/<session-id>/`. I
|
|
|
335
440
|
|
|
336
441
|
Besides the `yukino` CLI, the package ships a library entry for embedding Yukino in another host process.
|
|
337
442
|
|
|
338
|
-
| Entry | Output | Contents
|
|
339
|
-
| -------------- | ------------------------------------------ |
|
|
340
|
-
| CLI (`yukino`) | `dist/main.js` | Fully bundled, minified single file with a shebang; Node built-ins plus modules with native binaries or runtime assets (`sharp`,
|
|
341
|
-
| Library | `dist/lib/index.js`, `dist/lib/index.d.ts` | The `src/index.ts` barrel; runtime dependencies stay external and resolve from the consumer's `node_modules`.
|
|
443
|
+
| Entry | Output | Contents |
|
|
444
|
+
| -------------- | ------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
|
|
445
|
+
| CLI (`yukino`) | `dist/main.js` | Fully bundled, minified single file with a shebang; Node built-ins plus modules with native binaries or runtime assets (`sharp`, ink's `react-devtools-core` peer) stay external. |
|
|
446
|
+
| Library | `dist/lib/index.js`, `dist/lib/index.d.ts` | The `src/index.ts` barrel; runtime dependencies stay external and resolve from the consumer's `node_modules`. |
|
|
342
447
|
|
|
343
448
|
The library entry is terminal-independent by contract: it must never reach `src/ui/**` or a ui-only dependency, so a host without a TTY (a server, an editor extension, a test harness) can import it. `pnpm build` enforces that contract:
|
|
344
449
|
|