[BUG] Regression v2.1.126→v2.1.179: Sessions fail with '400 Input too long' despite being 3.5x smaller than working sessions

Description

Regression between v2.1.126 (working) and v2.1.179 (failing): sessions fail with "400 Input is too long for requested model" despite being significantly smaller than working sessions in older versions.

Evidence

Working Session (v2.1.126)

• UUID: ee1e3fc4-2456-4020-9664-67b207344156
• Size: 3,704 lines, 6.5MB
• Status: ✅ Works perfectly
• Compaction events: 83 detected

Failing Session (v2.1.179)

• UUID: 9a54abcb-1b9f-4d32-bfba-9818f181caff
• Size: 1,051 lines, 1.5MB
• Status: ❌ Fails with 400 error
• Compaction events: Only 5 detected
• Error message: "context_management": null in error response

Key observation: Failing session is 3.5x smaller than working session but fails immediately.

Reproduction

1. Start session on v2.1.179
2. Load multiple markdown documentation files via Read tool (~3 files, total ~2000 lines)
3. Work for several hours (file edits, bash commands)
4. Session grows to ~1000 lines in .jsonl
5. Next user message triggers: API Error: 400 Input is too long for requested model
6. Session becomes permanently unusable
7. context_management: null in error response suggests auto-compaction is not triggering

Expected Behavior

Auto-compaction should trigger before hitting limit (as it did in v2.1.126).

Environment

• Claude Code versions: 2.1.179 (failing) vs 2.1.126 (working)
• OS: Linux 6.17.0-35-generic
• Model: Sonnet 4.5
• Entrypoint: claude-vscode (VS Code extension)

Related Issues

Appears related to:

• #60141 - MCP tool schemas consuming excessive context (v2.1.138)
• #14472 - Cannot resume when context exceeds limit
• #31313 - Cowork session resume loop

Regression likely introduced between v2.1.126-v2.1.138.

Impact

Sessions become permanently unusable mid-work with no recovery path. Lose all conversation history and accumulated context.