[ZackLiuzeyu]

Summary

This turns the web dashboard from a read-only monitor into a control center for running a team of agents from the browser. It also adds a fourth agent type that runs on any OpenAI-compatible API. Everything the dashboard does goes through the same orchestrator code the MCP tools use, so the MCP / CLI workflow keeps working as before.

54 files, about +13.3k / −3.0k lines. The changes are grouped by area below. Each area is one or more self-contained commits, so it can be reviewed commit by commit, or I can split this into smaller PRs if you prefer.

What's in it

1. Web control center

• Actions from the browser:
  • message and broadcast;
  • pause/resume a session or an agent; release all agents, or only approved ones;
  • add, remove, rename and change the role of an agent;
  • create a new team; create, relaunch, archive, end, export and delete a session;
  • edit the plan and the knowledge store; manage file locks and ownership zones; guardrail steppers.
• One orchestrator for both hosts:
  • The orchestrator's tools moved to orchestrator/runtime.ts, which both orchestrator-server.ts (MCP) and dashboard/server.ts host. The web runs the same create_team / add_agent / release_agent / end_session code instead of a copy.
  • dashboard/control.ts is the API (POST /api/<action>). Every mutation names its session and checks that the slot belongs to it.
• Human gate: session config human_gate (also a create_team option) keeps a fully approved session waiting for a person instead of ending itself.
• Broker: plan item add/edit/delete, plus a merging config_patch on session update. Messages from the browser use the sender id human.

2. Workspace layout

• Three panes:
  • left: sessions and a file explorer;
  • middle: the team chat;
  • right: file preview and the team roster.
  • Side panes collapse and resize, and become drawers on narrow screens.
• Chat:
  • Markdown is rendered safely.
  • File names an agent mentions become links.
  • Enter sends without breaking IME input.
• Preview:
  • Formats: Markdown, code, images, PDF, .docx, CSV, sandboxed HTML, audio and video.
  • It follows the open file as agents edit it.
• File serving: dashboard/files.ts serves the project read-only (GET /api/fs/*). Every path goes through realpath, so neither .. nor a symlink can leave the project. HTML and SVG get a sandbox CSP, and .docx is rendered server-side with a small zip reader, with no new dependency.
• Themes: light, dark or follow the system.

3. Activity, approvals, models and limits

• Activity timeline: one live timeline per agent, covering thinking, tool calls, diffs, command output and turn costs. Claude stream-json, Codex app-server items and Gemini stream-json map onto one shape (orchestrator/activity.ts, broker /activity/*).
• Approvals:
  • Per-agent modes: bypass, auto, accept_edits and ask.
  • Claude now runs with --permission-prompt-tool stdio. Codex approval requests are no longer auto-accepted when a handler is set.
  • Requests are stored in the broker (/approvals/*), so any dashboard can answer them: Allow, Always allow, or Deny with a note.
  • Team-protocol tools never ask.
• Models and effort: set per slot and synced to running drivers. Claude switches live (set_model / set_permission_mode); Codex picks the change up on its next turn.
• Usage limits: session config.limits, in tokens or USD, per agent or per session. An agent that reaches its limit is paused and interrupted.

4. Plan limits, talking to agents, live thinking

• Plan limits: read without spending a model turn and cached for 2 minutes.
  • Claude: 5-hour and weekly windows, via a get_usage control request.
  • Codex: 5-hour and weekly windows, via account/rateLimits/read.
  • Gemini: daily quota per model, via Code Assist retrieveUserQuota with the stored Google login.
• Talking to agents:
  • Agents can message human and are told the id to use.
  • Interrupt stops one agent's turn so it answers first.
  • Q&A mode holds the whole team: running turns stop and agents only hear the person. Only read-only tools are allowed, teammates' messages wait, and nudges, respawns and auto-end pause.
• Live view: Claude runs with --include-partial-messages. A LiveTracker turns Claude, Codex and Gemini deltas into a typing indicator, the current step and the thinking as it forms.

5. Chinese / English UI and customization

• dashboard/i18n.js: tr(english, vars) with a Chinese dictionary, pattern rules for server-generated text, and in-place translation of the static HTML. Chinese is the default for zh browsers. tests/web-i18n.test.ts checks that every UI string has an entry.
• Avatars: default to each CLI's icon and can be changed to initials, a color or an image.
• Page background: a background image with adjustable opacity and blur, kept per browser.

6. "Other" agents on any OpenAI-compatible API

• Agent type api: DeepSeek, Qwen, Kimi, GLM, OpenRouter, a local Ollama, or anything else that speaks the chat-completions API.
• Providers: defined in a dialog with presets and model fetching. Fields: name, base URL, key, default model, optional prices, context window and extra request fields.
• Keys:
  • Stored in ~/.multiagents/providers.json (mode 0600); the browser only gets a masked hint.
  • A key written as $ENV_NAME is read from that environment variable.
  • Slots keep only the provider id.
• orchestrator/api-driver.ts:
  • Runs the agent in-process. It spawns the team-tools MCP server in driver mode, adds local tools with Claude's names (Read, Write, Edit, LS, Glob, Grep, Bash, WebFetch), all confined to the project, and streams completions (reasoning_content becomes thinking).
  • It emits Claude stream-json events and implements a new ClaudeSession interface, so the following all reuse the Claude paths: delivery, approvals, activity, the live view, cost accounting, Q&A, interrupts and live model switches.

7. Fixes

• Non-Latin session names: slugify kept only a-z0-9, so a session named in Chinese got the id "" and could not be opened or ended. Slugs now keep letters of any script; the broker refuses an empty id and migrates rows stored under "".
• Outdated broker: the broker daemon kept its old code across updates, so new features silently did nothing.
  • /health now lists BROKER_FEATURES. The dashboard restarts an idle outdated broker at startup, and otherwise shows a restart banner.
  • A restart no longer kills the agents: current brokers exit on SIGUSR2 without the SIGTERM peer cleanup, and older ones get SIGKILL (WAL keeps the data).
• Finished agents deleted: the dead-slot loop deleted every slot that had been disconnected for 5 minutes, so agents that exit when done (e.g. Gemini after approval) vanished. deadSlotVerdict now deletes only orphan slots; finished agents stay listed as offline.
• Codex rate limits: older app-servers reject params on account/rateLimits/read; it now sends null.
• CLI detection: detectAgent now honors MULTIAGENTS_CLAUDE_COMMAND / MULTIAGENTS_CODEX_COMMAND, so CI without the real CLIs can launch teams.
• Dashboard stuck on "connecting…": this happened when an older server served the new files. It now falls back to English and asks for a restart.
• CI: workflow_dispatch lets CI be run by hand from the Actions tab.

Behavior changes to be aware of

• Claude flags: driver-launched Claude agents now get --permission-prompt-tool stdio and --include-partial-messages.
• Broker restarts: the broker handles SIGUSR2 as "exit without touching peers". SIGTERM behaves as before.
• Background cleanup: it no longer deletes finished agents; they stay until a person removes them.
• Browser-started teams: these default to human_gate: true.

Testing

• HOME=$(mktemp -d) bun test: 352 pass, 2 skip (live tests, gated by MULTIAGENTS_LIVE_TESTS), 0 fail.
• tsc --noEmit reports no errors.
• New suites:
  • web control, files and i18n: tests/web-control.test.ts, tests/web-files.test.ts, tests/web-i18n.test.ts;
  • activity, approvals and settings;
  • plan usage, Q&A mode and the live view;
  • supervision verdicts;
  • tests/api-agent.test.ts, which runs a scripted fake OpenAI-compatible server through a real broker and the dashboard. It covers masking, thinking, tools, replying to the person, cost accounting, approvals, model switching and the symlink boundary.
• Driver tests still run against tests/fakes/fake-claude.ts and tests/fakes/fake-codex-app-server.ts.
• README and CLAUDE.md are updated for all of the above.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-22T11:30:18.100185Z We couldn’t complete this review. Comment @codex review to try again.	a6e7226	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.