fix: feat(acp): Honor MCP servers passed from ACP clients

[github-actions[bot]]

Closes #6111

Summary

Changes made for #6111 (feat/acp: Honor MCP servers passed from ACP clients)

Implementation

• Updated crates/goose-cli/src/commands/acp.rs
  • Imports
    • Now import Envs, ExtensionConfig, ExtensionEntry, and set_extension to construct MCP extensions from ACP input:
      • use goose::agents::extension::Envs;
      • use goose::agents::{Agent, ExtensionConfig, SessionConfig};
      • use goose::config::{get_all_extensions, set_extension, Config, ExtensionEntry};
  • new_session implementation
    • Before creating the Goose session, iterate over args.mcp_servers:
      • For each acp::McpServer::Stdio { name, command, args, env }:
        • Convert env (Vec of name/value pairs from ACP) into HashMap<String, String>.
        • Construct an ExtensionEntry with:
          • enabled: true
          • config: ExtensionConfig::Stdio { name, description: String::new(), cmd: command.display().to_string(), args, envs: Envs::new(envs), env_keys: Vec::new(), timeout: None, bundled: None, available_tools: Vec::new(), }
        • Call set_extension(entry) to persist/update the MCP extension configuration.
    • Non-stdio MCP server variants are ignored (no-op), keeping behavior minimal and aligned with the requirement to support stdio MCP servers.
    • Session creation remains the same except for removal of the prior inline comment about maybe_update_name.

Behavior

• ACP-provided stdio MCP servers from NewSessionRequest.mcp_servers are now merged into Goose's configured extensions:
  • They are stored via set_extension, using ExtensionConfig::key()/name_to_key, so name conflicts are resolved by last-writer-wins.
  • This matches the requested “union with ACP wins on conflict” behavior: an ACP-provided server with the same logical name will override an existing configuration entry.
• Existing configured extensions remain unchanged and are still loaded in GooseAcpAgent::new via get_all_extensions.
• MCP capabilities advertised in initialize remain unchanged:
  • http: false, sse: false.
• Only stdio transport from ACP is honored; other transports are skipped.

Verification

• source bin/activate-hermit
• cargo check
  • Succeeded.
• cargo test -p goose-cli
  • Fails due to missing system dependency libxcb when linking test binaries; this is an environment limitation, not a code issue.
• cargo fmt
  • Run and completed.
• ./scripts/clippy-lint.sh
  • Run via redirected log; no code changes required after linting.

Notable considerations / potential issues

• Environment variables from ACP mcp_servers.env are passed through Envs::new, which filters out disallowed/sensitive keys as defined in Envs; this may silently drop some ACP-provided env vars (by design for security).
• set_extension mutates the global extensions configuration:
  • ACP-provided MCP servers will persist for subsequent runs that reuse the same configuration store, consistent with other extension configuration flows.
  • If a client repeatedly sends different command/arg/env combinations for the same server name, later sessions will overwrite earlier ones.
• Non-stdio transports in mcp_servers are ignored rather than rejected:
  • This is intentional to provide graceful degradation while still meeting the “MUST support stdio MCP servers” requirement.

---

Generated by goose Issue Solver

[codefromthecrypt]

Ps the sketch i had in the linked issue was based on the current sdk not the old 0.4. I would recommend work after this to avoid drift https://github.com/block/goose/pull/6109

[codefromthecrypt]

/goose my sketch in #6111 was imprecise as it subtly affects global state. The goal for ACP is to be like --with-extension which scopes an MCP server addition, not persisting it forever. You should verify something like self.agent.add_extension(config).await.map_err is the correct way to accomplish this goal, and change your code accordingly, and in the same style as other code in this file.

[tlongwell-block]

This is not a great PR. Going to close it.