openclaw - 💡(How to fix) Fix sessions_spawn bloats subagent context even with lightContext:true [1 participants]

Bug Description

When spawning subagents via sessions_spawn with lightContext: true, the subagent still receives 11-32k tokens of context instead of the expected 1-2k tokens. This causes performance degradation for models with smaller context windows (e.g., Qwen 3.6 Plus).

Environment

• OpenClaw version: 2026.4.15
• Model: Qwen 3.6 Plus
• Spawn config: lightContext: false initially, then true after investigation

Evidence

Token counts observed:

• Simple task (lightContext: true): 11.5k input tokens
• Complex task (lightContext: true): 21.6k - 32.9k input tokens
• Expected with lightContext: 1-2k tokens

Comparison:

• Claude Code CLI + OmniRoute + Qwen: Works perfectly (~2k tokens)
• OpenClaw + OmniRoute + Qwen: Fails (11-32k tokens)

This isolates the issue to OpenClaw's sessions_spawn implementation, not the model or router.

Root Cause

OpenClaw's sessions_spawn forks parent context instead of starting fresh:

1. Parent context inheritance: Loads full conversation buffer (15-25 turns)
2. Memory auto-recall: Injects 2-7k tokens from MEMORY.md
3. Environment dump: Adds file tree, canvas state (~1.5k tokens)
4. Full tool manifest: Sends 30+ tools with schemas (12-14k tokens)
5. lightContext applied too late: Checked AFTER context is built

Pseudocode of current behavior:

ctx = buildFullContext(parent) // already 15k
if (lightContext) {
  ctx.tools = minimalTools // removes 5k
}
// Still has 10k+ from history/memory/system

Proposed Fix

1. Apply lightContext BEFORE building context

if (lightContext) {
  ctx = buildMinimalContext(task)
} else {
  ctx = buildFullContext(parent)
}

2. Disable parent context inheritance for subagents

// For subagent spawns:
inheritParentMessages: false,
skipMemorySearch: true,
minimalEnvironment: true

3. Filter tools per agent

if (agentId === 'qwen-coder') {
  tools = ['read', 'write', 'exec'] // Only essential
}

Reproduction Steps

1. Spawn qwen-coder with lightContext: true
2. Provide complex task (e.g., generate 150-line Python file)
3. Observe token count in logs: 11-32k instead of 1-2k
4. Compare with Claude Code CLI using same model/router: works with ~2k tokens

Impact

• Subagents with smaller context windows fail or underperform
• Increased API costs due to bloated context
• Inconsistent behavior vs other tools (Claude Code CLI)

Workaround

Added max_history_tokens: 3000 to agent config, but this only partially mitigates the issue. Root cause remains.

Files to Edit (for contributors)

Likely locations:

• src/sessions/spawn.ts - Context building logic
• src/context/builder.ts - lightContext implementation
• src/tools/manifest.ts - Tool filtering

Additional Context

Investigation conducted with muse-planner deep analysis. Token counts verified across multiple spawn attempts.

Expected behavior: lightContext:true should result in ~1-2k token context (system prompt + task + minimal tools), matching Claude Code CLI behavior.

Actual behavior: lightContext:true still results in 11-32k token context due to inherited parent state.