Skip to content

acp-agent-sidebar.md

Latest commit

 

History

History
90 lines (70 loc) · 4.11 KB

acp-agent-sidebar.md

File metadata and controls

90 lines (70 loc) · 4.11 KB
Raw

Embedded Claude agent

The macOS workbench embeds Claude in the left sidebar. Use the Experiment / Agent segmented control to switch between deterministic campaign controls and the conversation. The earlier local-model prompt authoring form is intentionally absent; natural-language experiment work now goes through the agent and the app's MCP boundary.

Runtime boundary

The app launches the pinned official Claude ACP adapter:

npx --yes @agentclientprotocol/claude-agent-acp@0.61.0

It negotiates ACP v1 over newline-delimited JSON-RPC stdio, opens one session, and passes fips-wind-tunnel as a session-scoped stdio MCP server. The MCP process still connects to the authenticated loopback endpoint owned by the visible app. ACP is a product integration boundary; it does not define engine time, ordering, graph representation, fidelity, or provenance.

Every new session selects ACP mode bypassPermissions, which is the adapter's equivalent of Claude Code's --dangerously-skip-permissions. Exceptional permission callbacks are resolved automatically and never interrupt the sidebar conversation.

Claude is the only reasoning layer in the embedded path. The ACP session disallows wind_tunnel_start_experiment and wind_tunnel_amend_experiment because those tools ask a second local model to interpret natural language. Before session/new, the app loads the complete checked-in skills/fips-wind-tunnel/SKILL.md and appends it to the Claude Code system prompt. The embedded session also disallows wind_tunnel_get_skill, so Claude cannot waste a tool round trip reloading guidance already in context. Instead, Claude calls structured tools directly:

  • wind_tunnel_set_parameters for direct-control changes;
  • wind_tunnel_run_campaign for a complete declarative campaign;
  • wind_tunnel_inject_event for a forward-only change at the cursor;
  • save, list, and rerun tools for durable exact-Campaign experiment reuse;
  • playback, state, wait, analysis, and explain tools for observation.

The two natural-language authoring tools remain available to other MCP clients; they are excluded only from this embedded ACP session.

Routine direct controls, playback, and saved-experiment operations must not search the knowledge catalog. Targeted knowledge reads remain available for protocol facts, raw schema authoring, unfamiliar evidence, and source-backed architecture or fidelity claims.

Responses stream into the transcript. Agent messages use a full Markdown renderer, including headings, lists, links, emphasis, tables, and fenced code blocks. Tool calls remain compact status rows rather than being folded into the model's prose.

The composer stays open while Claude is responding. A message sent mid-turn is injected into the running turn through the ACP steering extension (_session/steering) instead of being queued as a separate prompt, so Claude adapts immediately rather than waiting for the current turn to finish.

Requirements

  1. Claude Code must already be authenticated on the Mac.

  2. Node.js and npx must be installed.

  3. Install the Wind Tunnel MCP helper:

    scripts/install-mcp-server.sh --app "/absolute/path/to/FIPSD.app"

The app searches the active PATH, Homebrew locations, and installed NVM node versions for npx. These environment overrides are available for development:

  • FIPS_WIND_TUNNEL_NPX
  • FIPS_WIND_TUNNEL_MCP
  • FIPS_WIND_TUNNEL_WORKSPACE

The sidebar reports an actionable error and offers Retry when either executable is unavailable or ACP negotiation fails. Starting a new conversation terminates the current adapter, creates a fresh ACP session, reattaches the MCP server, and reapplies bypass mode.

Model selection

A menu in the sidebar header lets you choose Automatic, Sonnet, Haiku, or Opus. The choice persists across launches and is applied by setting ANTHROPIC_MODEL in the adapter process's environment before launch; Automatic leaves it unset so the adapter uses its own default. Because the model is only read at process start, changing the selection while a conversation is active starts a new conversation (equivalent to pressing the new-conversation button) to apply it.