Help automate coding

[keen peak]

Hi.
Im using my Openclaw to code in Claude. But it is anything but self running. It does one task then stops. It does not informe me when its done or during th eprocess. I constantly have to ask for a progess report. SOmetimes its still working, sometimes it has been ready for minutes and have not told me Claude was done. I woild like to kick it off with the very detailed task markdown files we create before we start. But he is not in any way self going. I ask for prrogress. He tells me clade is ready with the task. And "Should we go on with next task" I answer "Yes" and off he goes to fix th enext task. Then stops again and dont inform me of progress or blockers or anything. Untill i ask for progress.

I have tried usung settings in agents.md and other files but i just cant have him work...

Any ideas on how i can get this running "it self" once we done all the planning and researching?

[tranquil berry]

This usually is not an AGENTS.md problem. It’s usually 3 separate things:

1. You’re spawning Claude as a one-shot instead of a bound persistent ACP session

   • In chat, use:
     • /acp spawn claude --bind here
     • or /acp spawn claude --mode persistent --thread auto
   • If it’s not bound/persistent, it will finish a turn and basically wait.

2. ACP tasks default to done_only, not progress updates

   • That means “tell me when done”, not “keep me posted”.
   • After spawn, switch the task to:
     • openclaw tasks notify <taskId> state_changes
   • Use /tasks to find the active task.

3. Your top-level instruction is too turn-based

   • Tell it something like:
     • “Work through the entire markdown plan sequentially until all tasks are done or you are truly blocked. Do not stop after each subtask to ask for confirmation. Only report blockers, major milestones, and final completion.”

Also check these config bits if you want Discord thread/bound behavior to actually work:

• acp.enabled: true
• acp.dispatch.enabled: true
• session.threadBindings.enabled: true
• channels.discord.threadBindings.enabled: true
• channels.discord.threadBindings.spawnAcpSessions: true (needed if OpenClaw should create child threads for ACP)

If it finishes but doesn’t post back, that’s usually a delivery/binding issue, not Claude itself. Then check:

• /tasks
• openclaw tasks show <runId-or-sessionKey>

First thing though, please paste:

• openclaw --version
• /acp status
• /acp doctor

If you paste those, plus how you currently start Claude, I can probably tell you the exact missing piece.

Docs:

• https://docs.openclaw.ai/help/faq
• https://docs.openclaw.ai/tools/acp-agents
• https://docs.openclaw.ai/automation/tasks
• https://docs.openclaw.ai/gateway/configuration-reference

[keen peak]

IM not using Discurd to chat with Openclaw. I use the Openclaw gateway from a workstation next to my mini mac with Open claw on.

openclaw --version
OpenClaw 2026.4.14 (323493f)

ACP error (ACP_SESSION_INIT_FAILED): Session is not ACP-enabled: agent:main:main
next: If this session is stale, recreate it with /acp spawn and rebind the thread.

ACP doctor:
configuredBackend: acpx
activeRuntimeSessions: 0
runtimeIdleTtlMs: 0
evictedIdleRuntimes: 0
activeTurns: 0
queueDepth: 0
turnLatencyMs: avg=97479, max=127398
turnCounts: completed=23, failed=3
errorCodes: ACP_TURN_FAILED=3
registeredBackend: acpx
runtimeDoctor: ok (embedded ACP runtime ready)
runtimeDoctorDetail: agent=codex
runtimeDoctorDetail: command=npx @zed-industries/codex-acp@^0.11.1
runtimeDoctorDetail: cwd=/Users/martinsaldert/.openclaw/agents/rufferto
runtimeDoctorDetail: protocolVersion=1
healthy: yes
capabilities: session/set_config_option, session/set_mode, session/status

OpenClaw starts automatically via a macOS LaunchAgent at user login/boot session startup, using a plist file.
"/opt/homebrew/opt/node@22/bin/node /opt/homebrew/lib/node_modules/openclaw/dist/index.js gateway --port 18789"

[tranquil berry]

Ah, this helps a lot.

Your Gateway itself looks basically healthy. I don’t think the macOS LaunchAgent is the real problem.

The big clue is this:

• ACP_SESSION_INIT_FAILED: Session is not ACP-enabled: agent:main:main

That means you’re targeting a normal OpenClaw session as if it were an ACP harness session.

So the likely issue is: you’re mixing up these two modes:

• openclaw acp = ACP bridge mode (editor/client talks ACP to OpenClaw)
• /acp spawn ... / ACP agents = OpenClaw launches Claude Code / Codex / etc

agent:main:main is your regular main session, not a spawned ACP runtime session, so it won’t behave like a self-running Claude worker.

A few specific things I’d call out:

• activeRuntimeSessions: 0 means there is no live ACP runtime right now.
• runtimeDoctorDetail: agent=codex means your ACP backend currently resolves to Codex, not Claude.
  • If you expected Claude, check acp.defaultAgent or make sure you spawn/configure claude explicitly.
• If you want live progress, make sure acp.stream.deliveryMode is "live", not "final_only".

So my best guess is the real fix is:

1. Do not point this workflow at agent:main:main.
2. Create/use a real ACP session for the coding harness.
3. Use Claude explicitly if that’s what you want, instead of falling through to Codex.
4. Keep the task prompt explicit:
   • “Continue through the full plan until done or blocked. Don’t stop after each subtask.”

If you want, paste these 3 things and I can tell you the exact fix instead of guessing: