openclaw - 💡(How to fix) Fix Feature request: slim/bare mode for claude-cli backend to support sub-opus models [1 participants]

Problem

The claude-cli provider backend invokes the underlying Claude Code CLI without --bare, so every session rebuilds the full Claude Code harness: CLAUDE.md auto-discovery up the directory tree, hooks, plugin sync, auto-memory, background prefetches. On a reasonably active workstation this preload easily exceeds 200k tokens before any user prompt is added.

Consequence: any model in the Claude family with < 1M context — i.e. sonnet-4.6 (195k) and haiku-4.5 (200k) — immediately fails with:

[agent/cli-backend] cli exec: provider=claude-cli model=claude-haiku-4-5 promptChars=2
[agent/cli-backend] claude live session start: provider=claude-cli model=claude-haiku-4-5
[agent/cli-backend] claude live session turn failed: error=FailoverError
FailoverError: Prompt is too long

A 2-character prompt ("hi") to haiku reproduces the failure. Opus-4.7 (1024k ctx), gpt-5.4, and gemini-2.5-flash work because their windows tolerate the preload.

On my host I confirmed (by moving skills out of skills.load.extraDirs temporarily) that skills are not the primary bloat — even with zero SKILL.md files present, haiku still overflows. The bulk is coming from Claude Code's default context discovery/hooks, not from OpenClaw's own configuration.

This means on workstations like mine, any cron job or agent invocation that tries to use sonnet/haiku for cost reasons silently routes to the fallback (openai-codex/gpt-5.4 in my case), which also breaks .sh execution because gpt-5.4 + toolsAllow:[exec] triggers an interactive approval prompt in cron context and the script never runs.

Proposal

Expose a config (and/or per-invocation flag) that causes the claude-cli backend to pass --bare — or equivalently --no-context-files, --no-skills, plus explicit --system-prompt-file / --mcp-config / --add-dir / --agents / --settings where OpenClaw actually wants content — to the underlying claude binary.

Something like:

"agents": {
  "defaults": {
    "claudeCli": {
      "slimMode": true
    }
  }
}

or a per-agent-turn opt-in:

{
  "payload": {
    "kind": "agentTurn",
    "model": "claude-cli/claude-sonnet-4-6",
    "slimMode": true
  }
}

That would let cron jobs and other minimal-context workloads pick sonnet/haiku without fighting the auto-discovery.

Impact

Without this, users with non-trivial workstation configs are locked into claude-cli/claude-opus-4-7 for everything, even short wrapper-style cron jobs. That has real cost implications — and silently-wrong fallbacks to other providers that don't honour the same toolsAllow semantics.

Environment

• OpenClaw 2026.4.22 (00bd2cf)
• Linux (Ubuntu 6.8.0-110-generic)
• Claude Code CLI via @anthropic-ai/claude-code
• Reproduces with a fresh openclaw capability model run --model claude-cli/claude-haiku-4-5 --prompt "hi" when the invoking user's home has a non-trivial ~/CLAUDE.md or workspace attachments.

Happy to test a proposed fix if useful.