#i notice acp does not read the claude file in the cwd

1 messages · Page 1 of 1 (latest)

pseudo bone
#

if i start cluade code via cli, it reads the file, if i start an acp , it doesn't read the claud file

coral garnetBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> âś… Mark Solution

vague cipherBOT
wheat groveBOT
#

Let me verify this against the actual docs/source, because there are a couple different code paths here and I don’t want to guess.

#

I found the ACP docs and your OpenClaw here is 2026.4.9, so I’m checking against that code path now.

#

The FAQ path moved in this checkout, so I’m locating the current one before I answer.

#

yeah, most likely this is just a different cwd/workspace, not Claude ignoring CLAUDE.md.

Direct Claude Code CLI:

  • runs in your shell’s current directory
  • so it sees that repo’s CLAUDE.md

ACP Claude:

  • runs as an ACP session through OpenClaw/acpx
  • if you do not pass --cwd, it uses the ACP session cwd, usually the target agent workspace or backend default, not your terminal cwd
  • so if CLAUDE.md is not in that ACP cwd, it won’t get picked up

Verified docs:

What to try:

  1. check the ACP session cwd with /acp status
#
  1. explicitly set it with /acp cwd /absolute/path/to/repo
  2. or spawn with --cwd /absolute/path/to/repo
  3. if you changed CLAUDE.md after the session already existed, start a fresh ACP session too

If you want, paste:

  • openclaw --version
  • how you’re spawning the ACP session
  • what /acp status shows for cwd

and I can tell you the exact fix.

pseudo bone
#

thats not the issue, i notice the command is read only after the session si closed. what command does acp use to trigger the claude code ? is it pbypassing the claude.md ?

wheat groveBOT
#

I’m checking the actual acpx agent mapping now, because this part is easy to get wrong from memory.

#

I found the acpx side. Now I’m checking the actual default Claude adapter command, not just the docs summary.

#

it is not calling your normal interactive claude CLI path directly.

Current ACP path is:

  1. OpenClaw ACP session control
  2. bundled acpx backend
  3. Claude ACP adapter
  4. Claude-side runtime/session

In current acpx defaults, claude maps to:

npx -y @zed-industries/claude-agent-acp@0.21.0

So, short answer:

#
  • no, OpenClaw is not intentionally telling Claude to skip CLAUDE.md
  • but yes, ACP uses a different launch path than raw claude, so behavior can differ

If CLAUDE.md only seems to matter after close/reopen, that strongly suggests the Claude ACP adapter/runtime is loading project instructions at session bootstrap, not on every turn.

Also check whether you have an override in:

~/.acpx/config.json

because that can replace the built-in claude command.

Docs:

#

If you paste:

  • your ~/.acpx/config.json
  • the exact /acp spawn ... or spawn config
  • openclaw --version

I can tell you whether this is expected adapter behavior or a real bug.

pseudo bone
#

~/.acpx/config.json does not exists
/acp action:spawn claude --thread here --cwd /home/jarvis/.openclaw/workspace/my-project/da

wheat groveBOT
#

I’m checking one last thing: whether --thread here can reuse an existing bound ACP session in that thread, because that would explain the “only after close” part.

#

yeah, with that info, I don’t think OpenClaw is bypassing CLAUDE.md.

You have:

  • no ~/.acpx/config.json override
  • explicit --cwd /home/jarvis/.openclaw/workspace/my-project/da

So the launch path should be the built-in acpx default for claude, which is:

npx -y @zed-industries/claude-agent-acp@0.21.0

That means the behavior difference is probably here:

  • raw claude CLI
    vs
#
  • Claude ACP adapter (claude-agent-acp)

not OpenClaw skipping the file on purpose.

And your symptom, “it only takes effect after the session is closed”, really sounds like:

  • CLAUDE.md is getting picked up on fresh session bootstrap
  • not continuously during an already-running ACP session

So my read is: adapter/session-lifecycle behavior, maybe a bug, not cwd, and not an OpenClaw “skip CLAUDE.md” flag.

Practical next step:

  1. /acp close
  2. start a fresh spawn in a new thread
  3. if it still reproduces, I’d treat it as a real bug

Docs:

pseudo bone
#

weird

wheat groveBOT
#

yeah, that’s pretty weird.

If it repros from a brand-new thread + fresh /acp spawn, I’d call it a real bug, probably in the Claude ACP adapter path.

If you file it, include:

  • openclaw --version
  • exact spawn command
  • exact CLAUDE.md location
  • “raw Claude CLI sees it immediately, ACP Claude only after close/reopen”

Issue tracker: https://github.com/openclaw/openclaw/issues

If you want, I can help you turn this thread into a clean repro report.

pseudo bone
#

what is the command to tell acp to spawn a new session i na spawn threa d

#

in a discrod thread ?

wheat groveBOT
#

if you want a fresh ACP session in a thread:

From a parent channel

/acp spawn claude --thread auto --cwd /path/to/repo

From inside an existing thread

/acp close
/acp spawn claude --thread here --cwd /path/to/repo

Why the extra close? Because once a thread is bound, follow-up messages keep going to the same ACP session until you close/unbind it.

So the safe “new session in this thread” flow is: