Cockpit Multi-Agent Support

AoE’s cockpit (the web dashboard’s structured-rendering substrate) speaks the Agent Client Protocol (ACP) and supports every coding agent that ships an ACP server. Claude has the most polished surface today; other agents are first-class but lighter on per-tool polish.

Supported Agents

Agent	ACP entry	Install
Claude	claude-agent-acp (Zed adapter, requires >=0.37.0)	npm install -g @agentclientprotocol/claude-agent-acp@0.37.0
Codex (OpenAI)	codex-acp (Zed adapter)	npm install -g @zed-industries/codex-acp
OpenCode (SST)	opencode acp (native)	curl -fsSL https://opencode.ai/install | bash
Gemini (Google)	gemini --acp (native)	npm install -g @google/gemini-cli
Vibe (Mistral)	vibe-acp (native)	See https://github.com/mistralai/mistral-vibe
Pi	pi-acp (adapter)	npm install -g pi-acp (plus @earendil-works/pi-coding-agent)
aoe-agent	bundled	shipped with aoe

Adding a new ACP-capable agent: see docs/development/adding-agents.md, step 8 (Cockpit Profile).

Feature Matrix

Each cockpit feature either fires for any ACP agent, fires only when the agent’s profile opts in, or is currently claude-only.

Feature	Claude	Codex	OpenCode	Gemini	Other ACP
Streaming agent text	✓	✓	✓	✓	✓
Tool-call cards (execute / read / edit / search / fetch)	✓	✓	✓	✓	✓
Generic tool card fallback	✓	✓	✓	✓	✓
Permission / approval flow	✓	✓	✓	✓	✓
Mode picker	✓	depends	✓	depends	depends
Slash command palette	✓	depends	✓	✓	depends
Usage / context-window display	✓	depends	✓	✓	depends
MCP tool grouping	✓	claimed*	claimed*	claimed*	claimed*
/clear boundary divider	/clear	/new	/new	none	none
TodoWrite card	✓	—	—	—	—
Skill card	✓	—	—	—	—
ExitPlanMode synthesis	✓	—	—	—	—
ScheduleWakeup (/loop)	✓	—	—	—	—
Subagent indentation	✓	—	unverified	—	—
Session resume across aoe serve restart	✓	depends	✓	depends	depends

* All profiles default to the mcp__ prefix. If your agent uses a different MCP naming scheme, file an issue or PR adjusting the profile’s mcpPrefixes.

Notes on the matrix

• TodoWrite / Skill / ExitPlanMode / ScheduleWakeup are claude-only tools today, so the cards stay quiet on other agents. The cockpit doesn’t fire those cards based on coincidental tool names; gating happens in the agent profile.
• /clear is detected server-side by matching the user’s prompt against the profile’s clearAliases. Claude uses /clear; codex and opencode use /new. Gemini has no slash command verified as a conversation-clear boundary; /restore is a different semantic, so the cockpit doesn’t treat it as a clear. The composer’s / palette also surfaces each profile’s clear aliases as suggestions, since the adapters’ own available_commands_update channel does not always advertise them.
• Subagent indentation requires the adapter to emit a _meta.<namespace>.parentToolUseId field on child tool calls. claude-agent-acp emits _meta.claudeCode.parentToolUseId. OpenCode’s task tool spawns subagents but its parent-linkage convention hasn’t been verified, so the cockpit doesn’t render the indent until the contract is observed.
• Mode picker / slash palette / usage display depend on whether the adapter advertises the matching channels (available_modes, available_commands_update, usage_update). When the adapter doesn’t emit them, the UI simply stays empty rather than showing stale state.

How the Profile Works

Each agent has two profile sources, kept aligned by registry key:

• Server (Rust): src/cockpit/agent_profiles.rs. Carries parent_meta_namespaces, clear_aliases, and the supports_exit_plan_mode / supports_wakeup_tools capability gates.
• Frontend (TypeScript): web/src/lib/agentProfiles.ts. Carries the card-classifier alias map (shell → execute card, read_file → read card, etc.), the claude-specialised capabilities (todos, skills, wakeup), the MCP prefix list, and the special-title patterns matched only when the capability is on.

Profile data is conservative on purpose: where an adapter’s tool surface hasn’t been verified hands-on, the entry is omitted rather than guessed. The cockpit then renders the generic tool card, which is the right fallback. The user can file a PR adding the alias once they’ve used the agent and confirmed the wire shape.

Known Limitations

• Codex / opencode / gemini cockpit support has been built from adapter docs and code reading rather than hands-on session walkthroughs. Some tool aliases may need adjustment once each agent has been exercised end-to-end. File an issue with the wire tool.kind + tool.name observed on the cockpit side and we’ll update the profile.
• Gemini’s save_memory tool lands on the generic card today. A dedicated card is a follow-up.
• Multimodal input (image upload in the composer) is not implemented; gemini is the canonical target when it lands.
• Runtime capability discovery from the ACP InitializeResponse isn’t wired yet; profiles are the static source of truth. Runtime discovery is tracked separately.

Diagnosing Profile Issues

If a tool call on a non-claude agent renders as a generic card when you expect a specialised one:

1. Open browser devtools, find the tool-start WebSocket frame, note tool.kind and tool.name.
2. Check the agent’s profile in web/src/lib/agentProfiles.ts. The alias map only fires when tool.kind is "other" or not a concrete card kind; if the agent sends a real kind, that drives dispatch directly.
3. If the cockpit fired a claude-only card (TodoCard / SkillCard) on a non-claude agent, that’s a bug; file an issue with the wire shape.

For server-side gates (clear-boundary detection, plan / wakeup synthesis), the debug log (AGENT_OF_EMPIRES_DEBUG=1) captures which profile resolved for a session.