Adds an agent swarm / multi-agent execution mode to Step-Realtime-CLI, plus related cleanups.#30
Open
Daiyimo wants to merge 29 commits into
Open
Adds an agent swarm / multi-agent execution mode to Step-Realtime-CLI, plus related cleanups.#30Daiyimo wants to merge 29 commits into
Daiyimo wants to merge 29 commits into
Conversation
…-test-suite test: add comprehensive test suite (848 tests, 11 files)
…licy add tests and harden command policy
…install-support add Windows voice install support
- AGENTS.md §7: state that README.md, README_CN.md, and .gitignore are temporarily not accepting modifications, to keep contributors from landing parallel edits and creating merge conflicts. - AGENTS.md §7: link to docs/skills/pr-review-decision.md as the reference for maintainer-style PR review (decision rubric, blocker checklist, bilingual CN/EN comment templates). - docs/skills/pr-review-decision.md: new guide. Tool-agnostic; describes the workflow from "here's a PR" to "here's the comment to paste." - .gitignore: ignore .claude/ so per-user agent config doesn't leak in.
* chore: align prettier checks * fix: preserve smart compaction aborts
…onfig (stepfun-ai#20) - Add 33 new/enhanced test files covering packages/core, packages/utils, packages/realtime, packages/agent-sdk, extensions/*, skills/builtin, src/gateway, src/bootstrap, src/commands, src/tui, and integration tests - Refine vitest.config.ts coverage.include to precisely list tested modules, add lcov reporter for CI integration - Add test:changed script to package.json for incremental test runs - Fix sox-driver.test.ts for Windows platform guard (upstream PR stepfun-ai#12 compat)
Remove STEP_CLI_ENABLE_OPENTUI env var gate; OpenTUI is now unconditionally enabled.
…ams, remove dead code - swarm-plugin: add _enterLock mutex to prevent concurrent enter() racing on active state - swarm-plugin: cap seenPrompts at 1024 with semi-truncation to prevent unbounded growth - subagent-plugin: replace bare catch in safeFinalize with logger.warn - subagent-plugin: add 5 AgentSwarm params (context_mode, preset, isolate_workspace, worktree_name, wait_ms) - subagent-plugin: fix escapeXml to include apostrophe escaping - swarm-plugin.test: add afterEach cleanup to prevent test state leakage - open-tui-capability: remove dead parseOpenTuiEnabledValue function - pr-review-decision: fix Prettier escape artifacts (replace_all, first) Co-Authored-By: Claude <noreply@anthropic.com>
Collaborator
|
Thanks for the contribution! To get this ready to merge:
Full conventions: |
- Register createSwarmPlugin() in builtins so beforeModelRequest hook injects swarm reminder
- Add /swarm [on|off|task] slash command with manual/task/tool trigger modes
- Wrap AgentSwarm tool execution with swarmMode.enter('tool')/exit() lifecycle
- Auto-exit swarm mode after turn completes for non-manual triggers
- Fix StepCliInteractiveUiTone import missing from runtime.ts
Co-Authored-By: Claude <noreply@anthropic.com>
github-actions
Bot
added
area/build
github-actions
Bot
added
area/gateway
Resolve add/add conflicts in test files and workflows by taking upstream versions. Our swarm mode changes (runtime.ts, subagent-plugin.ts) are unaffected. Co-Authored-By: Claude <noreply@anthropic.com>
The env-var gating in root-command, resume-command, and the binary-bundle define was left inconsistent after the runtime layer hard-coded the flag to true. Replace all three with compile-time true so rolldown can still fold the dead TTY path, and drop the now-unused define entry. Co-Authored-By: Claude <noreply@anthropic.com>
本次修复 Step-Realtime-CLI 在 Windows CMD 下通过 setup.ps1 安装后
TUI 无法正常使用的三个 Bug，并附带几项工作区既存的改动。
Bug① TUI 闪退（@opentui/core bun bundle 无法在 Node 加载）
- 根因：@opentui/core 全版本都是 Bun 打包产物（含 bun:ffi），
Node 的 ESM loader 不支持 bun: 协议，TUI 启动即崩
- 修复：src/commands/root-command.ts 的 shouldUseTui 分支加智能
respawn——Node 进程 + TUI 路径自动 spawn('bun', ...) 切到 Bun
子进程；非 TUI 路径仍跑 Node，零影响；靠 process.versions.bun
判断避免无限循环；Bun 不可用时输出清晰错误并 exit 1
Bug② /swarm 在 TUI 没反应（gateway 级 slash 命令被绕过）
- 根因：TUI 的 handleSlashCommand 本地拦截所有 "/" 输入，但只
识别少数命令，/swarm 落入 default 显示不明显 "Unknown Command"
且不清空输入框；gateway 的 executeSlashCommand 只在 REPL 用，
TUI 模式通过 SDK 通信完全绕过
- 修复：贯穿 gateway/SDK 链新增 executeSlashCommand 转发机制，
让 TUI 把未识别的 slash 命令回传给 gateway 处理
- packages/protocol: StepCliSlashCommandResult 接口 + 方法签名
- src/gateway/runtime.ts: public executeSlashCommandExternal
（/swarm 特殊处理调 handleSwarmCommand + addEvent）
- src/gateway/local-gateway.ts + service/session-service.ts: 委托
- packages/sdk/src/index.ts: SDK 新增 executeSlashCommand
- src/tui/app.tsx: default case 转发 gateway + submitPromptContent
- src/tui/slash-commands.ts: /swarm 加入自动补全
Bug③ TUI 滚动失效（Bun setRawMode 不启用 VT 输入标志）
- 根因：Bug① 让 TUI 改跑 Bun 下，但 Bun 的 setRawMode(true) 在
Windows 上不会启用 ENABLE_VIRTUAL_TERMINAL_INPUT (0x200)——而
Node 的 libuv 会。没有这个标志 CMD 不把特殊键/鼠标翻译成 VT
序列，@opentui/core 的 StdinParser 收不到字节，所有滚动空操作
- 修复：新增 src/tui/win-console-input.ts，通过 bun:ffi + kernel32
的 GetConsoleMode/SetConsoleMode 在 STD_INPUT_HANDLE 补 0x200；
包装 setRawMode 每次进 raw 模式重新断言；纯增量 OR 不删已有标志
- src/runtime/local-opentui-entry.tsx: createCliRenderer 前调
install，后调 ensure 兜底；修正 useAlternateScreen → screenMode
属性名（CliRendererConfig 用 screenMode，默认 alternate-screen）
- 降级：非 Windows / 非 Bun / kernel32 不可用 全部 no-op
附带既存改动（工作区预先存在，一并提交）：
- 注册 packages/agent-sdk 为 workspace 包
（package.json、pnpm-workspace.yaml、scripts/build-packages.mjs、
pnpm-lock.yaml）
- TUI altScreen Windows 兼容：无配置时 Windows 默认 false
（src/bootstrap/config/loader.ts、src/runtime/runtime-config.ts）
验证：
- pnpm build 通过
- 非 TUI 路径零回归（node bin/step-cli.js --help / --version）
- Bun 基础链路正常（bun bin/step-cli.js --help / --version）
- QA 独立回归验证三个 Bug 修复全 PASS，无跨 Bug 回归
- dist 已同步到 ~/.step-cli/bin/dist/
- @step-cli/agent-sdk 被 extensions/realtime-voice 消费，knip 从根 workspace 看不到引用 - rg 是外部系统二进制，非 Node.js 包声明的 binary
Collaborator
|
Status update from the maintainer side — and thanks for the patience on this one. The swarm core itself is solidly engineered: the state machine with dedup, reminder injection scoped to the top-level harness, reuse of the existing BackgroundSubtaskManager, and the accompanying tests all read well. That said, we want to take proper time to think through the agent swarm / multi-agent execution mode before moving on it. The project already ships AgentTeam (parallel teammates within one turn) and SubAgent, so adding a third parallel-orchestration primitive is a product-level architecture decision — when to reach for swarm vs team, and whether they should converge — and we'd rather settle that deliberately than merge first and untangle later. We're keeping this PR open while it's under consideration; nothing here is lost, and we'll follow up on this thread and in #61 once we've landed on a direction. In the meantime, if you'd like something that can land quickly, a few focused areas where contributions are very welcome right now:
Thanks again for the work here — we'll be back on this thread once the orchestration direction is settled. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #61
Main changes
skills/builtin/src/swarm-plugin.ts)enter/exit/isActive) with dedup bytrigger:promptbeforeModelRequesthook injects swarm reminder for top-level main harness onlygetSwarmMode()for TUI or other consumersskills/builtin/src/subagent-plugin.ts)AgentSwarmlaunches parallel subagents fromitems+prompt_templateBackgroundSubtaskManagersubagent_modeloverride per swarm<agent_swarm_result>XML summaryAgentHarnessOptions.modelnow supportedSTEP_CLI_ENABLE_OPENTUIgate removed; OpenTUI is always enabled in this buildTests
getSwarmModeDocs / chores
task_list/task_waithintsdocs/skills/pr-review-decision.md.gitignoreignores `docs/compose/