CLI -p mode rejects valid prompts with 'Prompt is too long' below model context limit

Description

When using Claude Code CLI in print mode (-p), the CLI rejects prompts with "Prompt is too long" error even when the prompt size is well within the model's actual context window.

Environment

• Claude Code version: 2.0.52
• OS: macOS Darwin 24.6.0
• Model: claude-opus-4-5-20251101 (also tested with claude-sonnet-4-5-20250929)

Steps to Reproduce

1. Create a prompt that is ~524k characters (~120k tokens)
2. Run: claude -p --model opus "$(cat large_prompt.txt)"
3. CLI immediately returns "Prompt is too long" error

Observed Behavior

{"type":"assistant","message":{"content":[{"type":"text","text":"Prompt is too long"}],"error":"invalid_request"}}
{"type":"result","subtype":"success","is_error":true,"duration_ms":1024,"duration_api_ms":0,"num_turns":1,"result":"Prompt is too long"}

Key observations:

• input_tokens: 0 - The prompt is rejected before reaching the API
• duration_api_ms: 0 - No API call is made
• The error occurs regardless of model selected (opus, sonnet, sonnet[1m])

Expected Behavior

• Opus 4.5 has a 200k token context window
• Sonnet with extended context (sonnet[1m]) has 1M token context
• A ~120k token prompt should be accepted and sent to the API

The CLI should not have a hardcoded limit lower than the model's actual context window, or at minimum should respect extended context model variants.

Use Case

This affects automated workflows that use -p mode for code review/analysis tasks with large codebases. The same prompts work fine with other CLI tools (e.g., Codex, Gemini CLI) that don't impose artificial limits below the model's capability.

Workaround Attempted

• Tried --model opus, --model sonnet, --model sonnet[1m] - all fail
• Tried with and without --output-format stream-json - same error
• The limit appears to be in the CLI itself, not configurable via settings or environment variables

Suggested Fix

1. Remove or significantly increase the CLI's internal prompt size limit for -p mode
2. Or make it configurable via settings.json or environment variable
3. Or at minimum, respect the model's actual context window (200k for opus, 1M for extended context variants)