Claude Code — context window exhausted mid-task

Quick fix (TL;DR)

Real error messages you'll see

These are the exact strings returned by the Claude API service and its SDKs when this error occurs. Copy-paste-searching any of them should land on this page.

⚠ Context window at 92% (183,254 / 200,000 tokens)
Run /compact to summarise older messages, or /clear to reset the session.

Error: Context window exhausted (204,832 tokens > 200,000 limit).
Auto-compact could not free enough space. Run /clear or start a new session.

anthropic.BadRequestError: Error code: 400 - prompt is too long: 208912 tokens > 200000 maximum

Reference

What consumes context in a Claude Code session

Session-management commands (Claude Code)

Root causes, ranked by frequency

Based on developer reports across Claude API forums, GitHub issues, and Anthropic community during 2025–2026.

• 24%
  Full-file reads of large sources. Read(./src/main.py) on a 3K-line file adds ~15K tokens. Multiple such reads exhaust context before the task begins.
• 18%
  Verbose bash output cached in context. Running npm test or a large build dumps 20-50K tokens of logs. All retained in context.
• 14%
  Long-running task without /compact. 50-turn coding session with no compaction; context creeps toward the limit.
• 10%
  Multiple MCP tools returning large payloads. Database queries returning full rows; Github API returning entire diff.
• 8%
  Screenshots in Computer Use mode. Every screenshot is 1.5-2.5K tokens; 20-iteration session accumulates 30-50K.
• 7%
  SKILL.md over-loading. Many skills loaded proactively even when the task uses one.
• 8%
  Repeated file re-reads. Agent forgets it read a file already and reads it again — every duplicate wastes tokens.
• 11%
  Deep tool-use chains preserved verbatim. Long tool_use → tool_result → tool_use sequences; each result adds context that stays for later turns.

Fixes — copy-paste solutions

# Session flow that keeps context healthy

# --- Phase 1: implement feature A ---
> Read the existing auth module and add a new SSO provider.
[Claude reads files, edits code, runs tests]
> Great. Now /compact "keep the file paths I edited and the new provider name for future reference"
# Claude summarises phase 1 in a few hundred tokens

# --- Phase 2: implement feature B ---
> Add a settings page for the new SSO provider.
[Claude works on phase 2 with room to spare]
> Ok. /compact "keep the file paths and route names"

# --- Phase 3: run the full test suite and fix regressions ---
> Run npm test and fix anything that broke.
[Long output, many edits — but we have context because we compacted]

# --- If a truly unrelated task ---
> /clear    # full reset, start fresh
> Now, unrelated: refactor the analytics module.

# Guide the agent toward line-scoped reads

# --- BAD: whole-file read ---
> Read src/big_module.py     # → 15K tokens

# --- GOOD: grep, then range-read ---
> Find the definition of process_batch in src/big_module.py, then read a
> 40-line window around it.
[Claude runs: rg -n "def process_batch" src/big_module.py
 → line 847
 Reads src/big_module.py, offset=827, limit=40  → ~1K tokens]

# --- Codify in .claude/CLAUDE.md ---
# CLAUDE.md is auto-loaded per project. Add:

```
When reading files larger than 500 lines, first use grep or ripgrep to
locate the relevant symbol, then Read with offset/limit for a 30-50 line
window. Do NOT read whole large files.
```

# Now every session in this project follows the pattern automatically

# Pattern for multi-session tasks

# End of session 1:
> We're at 85% context. Before we hit the limit, write a handoff file
> ./HANDOFF.md that documents:
>   1. What we've done (files edited, tests written)
>   2. What remains (with specific file paths and line refs)
>   3. Any decisions we made and why
>   4. Commands to re-establish the working state (branch name, npm version)

[Claude writes HANDOFF.md]

> /exit

# Start of session 2 (fresh /clear):
> Read HANDOFF.md and continue the task from where we left off.

[Claude reads a few KB of structured handoff, starts fresh with full context budget]

# For very long tasks, do this every 2-3 hours:
> Update HANDOFF.md with today's progress. Then /compact.

Prevention checklist

Ship these seven safeguards once and this error stops appearing in your logs.

• Run /compact at every natural task boundary — before you need it.
• For files > 500 lines, always grep first and Read with offset/limit.
• Redirect large bash output to a file (> /tmp/build.log) and only Read the tail.
• Add a project CLAUDE.md encoding your team's context-hygiene rules.
• For multi-session work, checkpoint state to a HANDOFF.md file explicitly.
• Watch the token counter — /context tells you where the tokens went.
• Split unrelated subtasks across separate sessions with /clear; do not mix them in one long session.

Frequently asked questions

Related errors & hubs