Context Window and Response Truncation Issue (Claude Code)

[dhabedank]

Environment

Provider: Claude Code (Opus 4 20250514) (MAX 20x)
Current Context Window: 326.6k/200k (exceeding limit by 126.6k tokens)
Project Type: Next.js application with TypeScript

Issue

Experiencing a critical issue where either:

The context window overflow (326.6k/200k) is causing API responses to be truncated, OR
Truncated API responses are preventing proper context window management/clearing

The exact causality is unclear, but the result is a complete inability to continue working once this state is reached.

Observed Behavior

• API response gets truncated mid-stream when creating/modifying files
• Response ends abruptly: "text":"Now I'll create the signup page:\n\n"
• Context window shows 326.6k/200k (63% over limit)
• Automatic context condensing fails to meaningfully reduce size (329,718 >> 326,620)
• /smol command has no effect
• System enters an unrecoverable state

Similar Issue Context

The claude-task-master project encountered a related truncation bug (Issue #913, PR #920):

From their documentation:

"The Task Master CLI relies on the @anthropic-ai/claude-code SDK for local LLM execution. A bug in the Claude-Code CLI layer truncates stdout when the response size crosses several hard cut-off lengths (≈4k, 6k, 8k … 16k characters). When the SDK reads that output line-by-line it attempts JSON.parse(line) and throws a SyntaxError (e.g. 'Unterminated string in JSON at position 8000'). Any downstream consumer—including Task Master—crashes."

This suggests the truncation might be happening at the CLI/SDK layer rather than being purely a context window issue.

Potential Root Causes

• CLI Layer Truncation: Similar to claude-task-master #913, responses may be getting truncated at specific character thresholds
• Context Window Management: The system may not be properly handling context when approaching/exceeding limits
• Circular Failure: Truncated responses might prevent context clearing, which in turn causes more truncations

Impact

• Severity: Critical - Complete work stoppage
• Recovery: None - Must start new conversation, losing all context

Current Workarounds

✗ Automatic context condensing (minimal effect)
✗ /smol command (ineffective)
✗ Manual recovery (impossible due to truncation)

Key Questions for Investigation

• Are responses being truncated at specific character count thresholds (like in claude-task-master)?
• Is the context window genuinely over 326k tokens, or is this a measurement error?
• Why does context condensing reduce so little (329k to 326k)?
• Could implementing claude-task-master's truncation fix help here?

Recommendations

Given the similarity to claude-task-master's issue, investigating their fix (PR #920) may provide insights. Their solution for handling CLI stdout truncation at specific character thresholds could be applicable to this context window/truncation problem.

The issue significantly impacts productivity by requiring complete conversation restarts and loss of working context.

[dhabedank]

Here is how it looks for me - I thought I was having good luck by lowering the context condensing trigger threshold (70%), and it seemed to compact nicely - but then ballooned WAY up. Perhaps it has something to do with the number of tabs I had open which, post-compact, then repopulated the context window massively? Going to try again with another conversation but lower the default of 20 open tabs context limit down to half.

[chrarnoldus]

Perhaps it has something to do with the number of tabs I had open which, post-compact, then repopulated the context window massively?

That sounds plausible. Have you had any luck lowering the max number of open tabs?

[github-actions]

Closing this one out since it's been inactive for a while — not because it's a bad issue, just that older items tend to lose context over time and we'd rather start fresh if this is still a problem.

Please feel free to reopen (or open a new issue) if you're still running into this. We're happy to take another look! 🙏