openclaw - 💡(How to fix) Fix Feature: slim mode to pass --bare to claude-cli backend (unlock sub-opus models on heavy workspaces) [1 participants]

Problem

On a workspace with many skills / a rich CLAUDE.md / ShieldCortex-style plugin hooks, the OpenClaw claude-cli backend builds a pre-turn harness (auto-discovered CLAUDE.md up the tree, plugin sync, hooks, skill catalog, auto-memory) that can exceed the context window of smaller Claude models.

Concrete reproduction on our E.D.I.T.H. host (OpenClaw 2026.4.22, Claude Code subscription):

$ openclaw capability model run --model claude-cli/claude-sonnet-4-6 --prompt "hi"
[agent/cli-backend] cli exec: provider=claude-cli model=sonnet promptChars=2
[agent/cli-backend] claude live session start: provider=claude-cli model=claude-sonnet-4-6
[agent/cli-backend] claude live session turn failed: 
  provider=claude-cli model=claude-sonnet-4-6 durationMs=3933 error=FailoverError
FailoverError: Prompt is too long

• Same command with --model claude-cli/claude-haiku-4-5 and a 9-character prompt → same FailoverError: Prompt is too long.
• --model claude-cli/claude-opus-4-7 (1024k ctx) works fine — the harness fits in opus's window but not in sonnet's 195k / haiku's 200k.

Result: even though the Claude Code subscription allows sonnet/haiku model IDs to pass through the claude-cli provider, sub-opus models are effectively unusable on heavy workspaces — requests silently fail over to fallback (gpt-5.4), which for .sh script invocation then blocks on interactive approval.

Claude Code already supports this

claude CLI has a --bare flag:

Minimal mode: skip hooks, LSP, plugin sync, attribution, auto-memory, background prefetches, keychain reads, and CLAUDE.md auto-discovery. Sets CLAUDE_CODE_SIMPLE=1. Skills still resolve via /skill-name. Explicitly provide context via: --system-prompt[-file], --append-system-prompt[-file], --add-dir (CLAUDE.md dirs), --mcp-config, --settings, --agents, --plugin-dir.

But the OpenClaw claude-cli backend doesn't currently surface this. There's also no knob to pass --bare or equivalent trimming flags per-session or per-job.

Proposed feature

A config/flag combination that lets cron jobs (and one-shot openclaw capability model run / openclaw infer model run invocations) opt into a slim mode which maps to claude --bare under the hood, with optional explicit context via --append-system-prompt-file / --add-dir.

Ideal shape:

// openclaw.json
"agents": {
  "defaults": {
    "claudeBackend": {
      "slimMode": false
    }
  }
}

Plus per-job override in cron payload:

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

Also surface --slim / --bare on openclaw capability model run and openclaw infer model run for parity.

Why this matters

• Unlocks sonnet-4.6 and haiku-4.5 for cron jobs on heavy workspaces — meaningful cost savings for high-frequency jobs (our task-watcher fires 48×/day, school-inbox-guard 36×/day, etc.).
• Keeps opus-4.7 as the default for interactive sessions that need the full harness.
• Non-breaking: opt-in, default false.

Host details

• OpenClaw 2026.4.22 (build 00bd2cf)
• Node (from ~/.npm-global)
• Model catalog shows claude-cli/claude-sonnet-4-6 (195k), claude-cli/claude-haiku-4-5 (200k), claude-cli/claude-opus-4-7 (1024k) all as configured
• Auth: anthropic:claude-cli OAuth via Claude Code subscription
• Workspace size: 20 skills under ~/skills, 3 under ~/.openclaw/workspace/skills, plus a ~/.openclaw/workspace/memory/ dir (~1.4 MB across 58 daily files). Stripping skills alone was insufficient — harness still overflows 200k with zero skills present.

Happy to test any prototype branches on our host.