Subagents have no awareness of CLAUDE_CODE_MAX_OUTPUT_TOKENS value

Problem

Subagents (spawned via the Task tool) inherit CLAUDE_CODE_MAX_OUTPUT_TOKENS from the parent session but have no way to know the actual value. This means they cannot adapt their output — at non-default settings, Write tool calls silently truncate.

Impact by setting

| MAX_OUTPUT | Write limit | What happens |
|------------|-------------|--------------|
| 32000 (default) | ~3000 lines | Works fine |
| 16000 | ~1500 lines | Large files truncated without warning |
| 8000 | ~800 lines | Most multi-file writes fail silently |

Reproduction

1. Set export CLAUDE_CODE_MAX_OUTPUT_TOKENS=8000
2. Use a skill/command that spawns a subagent which writes a file >800 lines
3. The subagent's Write call silently truncates — no error, no warning

Current workaround

We manually include a TOKEN BUDGET block in every subagent prompt with the current value so the subagent can self-limit. This works but requires every skill/command author to remember the convention, and new skills will miss it.

Suggested improvement

Expose CLAUDE_CODE_MAX_OUTPUT_TOKENS (and optionally MAX_THINKING_TOKENS) to subagents automatically — either as:

• A system-level context variable available in the subagent's prompt (like the model name)
• A property on the Task tool response
• Documentation/guidance on the truncation behavior so skill authors know to handle it

This would eliminate the need for manual token budget forwarding in every subagent prompt.

Environment

• Claude Code: 2.1.71
• Model: claude-opus-4-6
• OS: macOS 26.3