Why does newSession take 11+ seconds?

[dmeehan1968]

response is the raw Typescript API response to connection.newSession.

The durationMs shows how long the response took from the request. There are no MCP servers in the request.

It does this on EVERY request, not just the first after launching the ACP process, and does not include the initialise request.

Claude CLI takes ~2 seconds to startup, creating a new session, so I don't see why the ACP should be so slow.

Materially this means there is a long delay before I can show the configOptions to the user, unless I cache them from a prior session which seems a dubious approach.

{
  "projectId": "b452e9f6-57ab-424a-b4ae-a4625cd204ce",
  "response": {
    "sessionId": "89843009-0684-42cb-a059-02d6496e8cf8",
    "models": {
      "availableModels": [
        {
          "modelId": "default",
          "name": "Default (recommended)",
          "description": "Opus 4.6 with 1M context · Most capable for complex work"
        },
        {
          "modelId": "sonnet",
          "name": "Sonnet",
          "description": "Sonnet 4.6 · Best for everyday tasks"
        },
        {
          "modelId": "haiku",
          "name": "Haiku",
          "description": "Haiku 4.5 · Fastest for quick answers"
        }
      ],
      "currentModelId": "default"
    },
    "modes": {
      "currentModeId": "default",
      "availableModes": [
        {
          "id": "auto",
          "name": "Auto",
          "description": "Use a model classifier to approve/deny permission prompts."
        },
        {
          "id": "default",
          "name": "Default",
          "description": "Standard behavior, prompts for dangerous operations"
        },
        {
          "id": "acceptEdits",
          "name": "Accept Edits",
          "description": "Auto-accept file edit operations"
        },
        {
          "id": "plan",
          "name": "Plan Mode",
          "description": "Planning mode, no actual tool execution"
        },
        {
          "id": "dontAsk",
          "name": "Don't Ask",
          "description": "Don't prompt for permissions, deny if not pre-approved"
        },
        {
          "id": "bypassPermissions",
          "name": "Bypass Permissions",
          "description": "Bypass all permission checks"
        }
      ]
    },
    "configOptions": [
      {
        "id": "mode",
        "name": "Mode",
        "description": "Session permission mode",
        "category": "mode",
        "type": "select",
        "currentValue": "default",
        "options": [
          {
            "value": "auto",
            "name": "Auto",
            "description": "Use a model classifier to approve/deny permission prompts."
          },
          {
            "value": "default",
            "name": "Default",
            "description": "Standard behavior, prompts for dangerous operations"
          },
          {
            "value": "acceptEdits",
            "name": "Accept Edits",
            "description": "Auto-accept file edit operations"
          },
          {
            "value": "plan",
            "name": "Plan Mode",
            "description": "Planning mode, no actual tool execution"
          },
          {
            "value": "dontAsk",
            "name": "Don't Ask",
            "description": "Don't prompt for permissions, deny if not pre-approved"
          },
          {
            "value": "bypassPermissions",
            "name": "Bypass Permissions",
            "description": "Bypass all permission checks"
          }
        ]
      },
      {
        "id": "model",
        "name": "Model",
        "description": "AI model to use",
        "category": "model",
        "type": "select",
        "currentValue": "default",
        "options": [
          {
            "value": "default",
            "name": "Default (recommended)",
            "description": "Opus 4.6 with 1M context · Most capable for complex work"
          },
          {
            "value": "sonnet",
            "name": "Sonnet",
            "description": "Sonnet 4.6 · Best for everyday tasks"
          },
          {
            "value": "haiku",
            "name": "Haiku",
            "description": "Haiku 4.5 · Fastest for quick answers"
          }
        ]
      }
    ]
  },
  "durationMs": 11481
}

[rohan-patra]

The biggest avoidable issue is if you have many MCPs (especially stdio transport instead of http transport) configured in Zed and Zed passes those through via the ACP connection. These each spin up standalone subprocesses with significant startup times.

The biggest unavoidable delay is the startup time for a standalone node subprocess to run the claude agent sdk (the acp adapter itself), and then the actual cli startup time added on to that.

[dmeehan1968]

@rohan-patra I don't have any MCP's configured (I'm using my own ACP client).

I appreciate the setup that needs to happen, but if the Claude CLI can start and establish a session in ~2 seconds, it seems odd that the node/adapter overhead would be so much larger.

[dmeehan1968]

Is there a way to debug the adapters startup that would give timing information?

[rohan-patra]

Vanilla startup should qualitatively take at least double the amount of time that CLI startup does without any extra configs/passthroughs

[benbrandt]

11 secs does seem excessive. The latest version of the Agent SDK which we are using now uses a binary instead of having Node run a giant JS file. We'll see if this helps with startup times or not. But unlike the CLI, we are spawning a process for every session because that is how Anthropics SDK API works

[dmeehan1968]

@benbrandt I'd be happy to do some debugging on this to get relevant timings. I had a dig through the code and it seems there isn't logging for that, perhaps other than enabling the raw SDK metadata if that could reveal progress. I get that a new claude process is spawned for each, but it still seems way too slow (especially compared to other ACP's I've tried, even though they have similar approaches). I've also not noticed this sort of lag in other agent clients that use the Claude SDK directly.

The acp-inspector project seems like vapour-ware, not implementation, only features and not updated for 3+ months.

Right now this is negatively impacting my ACP client, so would really like to optimise if thats at all possible.

[benbrandt]

New version of the SDK is here: #572

I have never seen 11+ seconds... but hopefully this helps. Are you doing anything different in your acp client that might be causing a difference? anything about the runtime you can share?

[dmeehan1968]

@benbrandt Using 0.30.0 gets it down to ~6s for each new session. A big improvement, bit still seems to be a bit more than it could be.