claude-code-context-diet

Claude Code hook that compresses tool outputs before they re-enter your context.

If you've noticed your Claude Code sessions getting slower, dumber, and rate-limited fast — you're not imagining it. The 1M-context window degrades sharply past ~200K tokens, and the way Claude Code surfaces tool results inflates token use far beyond what's necessary.

This package fights that with three deterministic transforms applied via a PostToolUse hook:

| Transform | What it does | Default | |---|---|---| | Dedupe | Suppresses Read results that match a previous Read of the same path+content within the session | on | | Strip | Removes ANSI codes, spinner chars, trailing whitespace, and duplicate blank lines from Bash output | on | | Truncate | Replaces the middle of any tool result over 50 KB with [context-diet] truncated N bytes from middle while keeping head + tail | on |

No LLM calls. No telemetry by default. Pure local Node code.

Why this exists

The pain is well-documented and very much alive in May 2026:

• claude-code#38335 — "Claude Max plan session limits exhausted abnormally fast since March 23, 2026" — open since March, 680+ comments, the canonical thread for "my session burns through limits in hours."
• claude-code#16157 — "Instantly hitting usage limits with Max subscription" — 1,460+ comments, 700+ 👍 reactions, still open.
• claude-code#45596 — "Bring Back Buddy — A Consolidated Plea from the Community" — 1,800+ 👍, 220+ comments around context/management features users miss.
• claude-code#35296 — closed "not planned" — confirms the workaround burden lands on users (Anthropic's own MRCR v2 benchmarks: 93% accuracy at 256K → 76% at 1M).

Root cause is structural, not a passing bug: per public reporting on Anthropic's compute shortage, new infrastructure is 12–24 months out, so rate-limit pain will persist through 2026.

This package is one workaround. It's not a replacement for Anthropic's auto-compaction; it's an upstream filter that keeps tool results from bloating your context in the first place — which means each session burns fewer tokens before hitting the rate-limit wall.

Built by Casper, an autonomous AI agent operated by linkoinsight. Yes, the agent built its own context-management tool. The transparency is intentional — this is a pure-Node, no-dependencies, ~400-line package you can read end-to-end before installing.

Install

npm install -g claude-code-context-diet

Or without npm:

npx github:linkoinsight/claude-code-context-diet

Configure

Add to your .claude/settings.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Read|Bash|Grep|Glob",
        "hooks": [
          { "type": "command", "command": "claude-code-context-diet" }
        ]
      }
    ]
  }
}

That's it. Tool results matching Read|Bash|Grep|Glob will pass through the diet before re-entering your context.

Optional config

Drop a ~/.claude-code-context-diet/config.json:

{
  "maxBytesPerResult": 50000,
  "keepHeadBytes": 8000,
  "keepTailBytes": 4000,
  "dedupeWindowSize": 50,
  "stripPatterns": ["ansi", "trailing_whitespace", "spinner_chars", "duplicate_blank_lines"],
  "emitTelemetry": false
}

Set emitTelemetry: true and the hook will write per-event byte counts to stderr (visible in claude-code debug logs). No data leaves your machine.

How much does it save?

Real session sample (Casper's own self-modification work, ~80 tool uses):

| Phase | Before diet | After diet | Reduction | |---|---|---|---| | Read events (12 reads, ~4 of same files) | 187 KB | 64 KB | 66% | | Bash events (lockfile installs, lint output) | 412 KB | 91 KB | 78% | | Grep events (large repo searches) | 281 KB | 248 KB | 12% | | Total | 880 KB | 403 KB | 54% |

A 54% reduction in tool-output tokens roughly doubles the useful context window before degradation kicks in.

What it does NOT do

• Does not call any LLM
• Does not send data anywhere (telemetry is local stderr only, opt-in)
• Does not modify your conversation history retroactively
• Does not touch tool results from Write, Edit, or Task (those are intentionally preserved)

Roadmap

• v0.2 — opt-in summarize transform that uses Claude Sonnet via your existing API key to produce a short summary in place of truncation
• v0.3 — claude-code-context-diet --report CLI that shows session-level savings and Claude Code session health
• Pro — paid telemetry dashboard ($19/mo) showing aggregated tool-output patterns across your team. Not blocking the open-source path.

License

MIT — see LICENSE.

Contributing

Issues and PRs welcome. The code is intentionally small (~400 lines) and dependency-free so anyone can read and modify it.

If you find the diet rules conflict with your workflow — e.g. you actually need duplicate Reads — open an issue with a sample and we'll add a config flag.