@livx.cc/agentx 0.99.122 → 0.99.124

package/README.md

@@ -87,7 +87,7 @@ Beyond file tools, the runtime ships the higher-altitude pieces too — each an
 - **Skills + memory** — VFS-backed (`skillsDir`/`memoryDir`); persistence is just the backend choice.
 - **Subagents** (`subagents`; **typed agents** via `agentsDir` — `<dir>/<name>.md` defines a persona + model + scoped tools, selected with the `Task` `agentType`), **hooks** (`hooks`: preToolUse/postToolUse/onStop — block or audit any tool call), **slash-commands** (`commandsDir`), **TodoWrite**, **MCP** (`mcpToolsToAgentTools`).
 - **A delegated child is a full agent**: it inherits the parent's **tools** (incl. host-injected/MCP ones), **system prompt**, and **providerOptions** (only when it stays on the same provider — cursor's `cwd`/`cursorSession` never leak onto an anthropic child; override with `providerOptionsFor(model)`). `Task({ model })` (and a `TaskBatch` task's `model`) runs one sub-task on another model — alias or full id, precedence **explicit `model` > `agentType` def's `model` > parent's**. A def's frontmatter `tools:` allowlist narrows the inherited set and may name a host-injected tool by name; its body still overrides the persona. The host's `preToolUse` gate is inherited too, so delegation is not a way around it (the observer/lifecycle hooks stay parent-only). Inherited tools that escape the VFS are re-bound per child (`AgentTool.withCwd`): a `isolation:"worktree"` child's real `Shell` runs in the WORKTREE, keeping any host rename (`Shell` -> `Bash`) (background shell jobs are off there, and the parent-bound `ShellOutput`/`ShellStatus`/`ShellKill` are dropped rather than pointed at the parent's jobs); an overlay child's shell writes straight to the working tree — an overlay isolates the VFS, not a process.
-- **Two-model modes** (`src/modes.ts`) — **delegate** (`delegate: { worker }`): the agent's model PLANS and may not write (Write/Edit/MultiEdit/ApplyEdits/bash/Shell are refused at execution, keyed on `ToolContext.model`), `Task`/`TaskBatch` go on (depth 1, children default to the worker, and get the host's `postToolUse` so their writes are observed). **advisor** (`advisor: { model, limit?, snapshot? }`): the agent's model does the work and an `Advise` tool asks the stronger model (≤3 calls; it sees only the request, the host's optional `snapshot()` and the question); the first write waits for one consult. Both are off when the models are equal or either is `cursor/*`; the second model's usage comes back as `RunResult.workerUsage` / `workerModel` (not in `usage`). Prompt directives switch them: `[opus->haiku] task` (delegate), `[haiku<-opus] task` (advisor), `[off]`; keyed forms `delegate:x`, `advisor:x`, `mode:off`. A host passing its own `keys` to `parseDirectives` must list `delegate` / `advisor` / `mode`.
+- **Two-model modes** (`src/modes.ts`) — **delegate** (`delegate: { worker }`): the agent's model PLANS and may not write (Write/Edit/MultiEdit/ApplyEdits/bash/Shell are refused at execution, keyed on `ToolContext.model`), `Task`/`TaskBatch` go on (depth 1, children default to the worker, and get the host's `postToolUse` so their writes are observed). **advisor** (`advisor: { model, limit?, snapshot? }`): the agent's model does the work and an `Advise` tool asks the stronger model (≤3 calls; it sees only the request, the host's optional `snapshot()` and the question); the first write waits for one consult. A `claude-code/*` agent gets the same modes as SDK options (`ccModeOptions`): a planner's native Write/Edit/MultiEdit/NotebookEdit/Bash/Task/Agent are `disallowedTools` and agentx's Task/TaskBatch reach it over the in-process bridge; an advised worker gets `Advise` bridged and a PreToolUse hook that denies native writes until one consult. Mixed pairs (anthropic ↔ claude-code) work. Both are off when the models are equal or either is `cursor/*`; the second model's usage comes back as `RunResult.workerUsage` / `workerModel` (not in `usage`). Prompt directives switch them: `[opus->haiku] task` (delegate), `[haiku<-opus] task` (advisor), `[off]`; keyed forms `delegate:x`, `advisor:x`, `mode:off`. A host passing its own `keys` to `parseDirectives` must list `delegate` / `advisor` / `mode`.
 - **Streaming** (`stream: true` → `text_delta` via `HostBridge.notify`) and **context compaction** (`compaction: { maxMessages }` → edge-safe summarize-and-boundary). Defaults preserve the original non-stream, drop-oldest behavior.
 - **Multi-turn + project context** — `Agent.send()` continues a conversation across turns (vs `run()`, which starts fresh); **project instructions** (`instructionFiles`: `AGENTS.md`/`CLAUDE.md` at the FS root) inject into the system prompt.
 - **DuplexAgent** (`src/duplex.ts`) — voice-optimized three-tier engine (reflex/act/think): a fast reflex agent streams instant replies and self-selects escalation — `Act` for standard tool work (Sonnet-class), `Think` for deep reasoning (Opus-class, configurable, default on). Results are pushed back and re-voiced by the reflex (turn mutex, coalesced completions, `TaskStatus`/`CancelTask`). See [`mind/10`](./mind/10-duplex.md).

package/dist/cli.d.ts

@@ -1,7 +1,7 @@
 #!/usr/bin/env bun
-import { u as RouterState, H as Hooks, c as AgentOptions, v as RunResult, e as ChatUsage, R as ReasoningEffort, C as ChannelMode, b as Agent } from './secondChannel-BddfVzTp.js';
+import { w as RouterState, H as Hooks, c as AgentOptions, x as RunResult, f as ChatUsage, R as ReasoningEffort, d as ChannelMode, b as Agent } from './secondChannel-GFT69ax4.js';
 import { IFilesystem } from '@livx.cc/wcli/core';
-import { M as Message, U as UserQuestion, H as HostBridge, c as ContentPart, o as MessageContent } from './tools-DZHZIyxh.js';
+import { M as Message, U as UserQuestion, H as HostBridge, c as ContentPart, o as MessageContent } from './tools-DwfeF8VQ.js';
 
 /**
  * On-disk session store for the CLI: each conversation is one JSON file at