Skip to content

README.md

Latest commit

 

History

History
99 lines (63 loc) · 4.63 KB

README.md

File metadata and controls

99 lines (63 loc) · 4.63 KB

opus-context

Teaches Claude Code Opus 4.6 to use its full 1M token context window instead of defaulting to conservative behaviors designed for smaller models.

Summary

Claude Code defaults to reading files in small chunks, delegating searches to subagents, and re-reading files it already has in context. These habits make sense for 200K-token models but waste Opus 4.6's capacity. This plugin provides always-on behavioral rules that override those defaults: read whole files, read directly instead of delegating, pre-load dependencies before editing, and track what's already loaded.

Principles

[P1] Read More, Not Less: With 1M tokens, a 500-line file costs ~0.1% of context. The cost of missing context from partial reads exceeds the cost of loading the full file.

[P2] Own Your Context: Read files directly rather than delegating to subagents that return summaries. Summaries lose nuance; source code doesn't.

[P3] Load Before You Leap: Read a file's dependency graph (imports, callers, tests) before editing it. Never discover mid-edit that you need context you don't have.

[P4] Budget Awareness: Read aggressively early; shift to selective reading as context fills. Don't run out of context on long sessions.

Requirements

None beyond Claude Code with Opus 4.6 (1M context).

Installation

/plugin marketplace add L3DigitalNet/Claude-Code-Plugins
/plugin install opus-context@l3digitalnet-plugins

For local development or testing without installing:

claude --plugin-dir ./plugins/opus-context

How It Works

flowchart TD
    Task([Task received]) -->|apply| Baseline[Apply baseline rules:<br/>whole-file reads, no re-reads,<br/>parallel reads, direct over delegation]
    Baseline --> Complex{Complex task?<br/>3+ files, debugging,<br/>new feature, refactor}
    Complex -->|No| Execute[Execute with baseline rules]
    Complex -->|Yes| Scope[Step 1: Scope file graph<br/>via Glob/Grep]
    Scope --> Load[Step 2: Load strategically<br/>in priority-ordered parallel batches]
    Load --> DeepExec[Step 3: Execute with<br/>full context loaded]
    Execute --> Budget{Context pressure?}
    DeepExec --> Budget
    Budget -->|Low| Continue((Continue<br/>aggressively))
    Budget -->|High| Conserve((Shift to<br/>selective reads))
Loading

Usage

This skill is always-on. It activates automatically via skill description matching whenever Claude Code processes a task. No manual invocation needed.

Optional CLAUDE.md reinforcement (add to your global ~/.claude/CLAUDE.md):

## Context Management

When running on Opus 4.6 (1M context), always consult the `opus-context:deep-context` skill for file reading, delegation, and pre-loading decisions.

Skills

Skill Loaded when
deep-context Always, via broad description matching. Governs file reading, subagent delegation, and pre-loading decisions.

Hooks

All hooks are registered declaratively via hooks/hooks.json.

Hook Event What it does
session-start.sh SessionStart Prints a terminal banner confirming 1M context mode is active. Terminal-only; does not inject into AI context.

Design Decisions

  • Always-on via description, not SessionStart injection: SessionStart hook stdout goes to the terminal, not AI context. The skill's broad description field is the primary activation mechanism, with optional CLAUDE.md reinforcement as a second trigger path.

  • Heuristic budget thresholds instead of percentages: Claude can't measure its exact context usage. Budget awareness uses observable signals (file count, conversation length, compaction events) instead of precise percentages.

  • No model detection gate: The skill's rules are framed around having abundant context. A smaller-context model that loads the skill will naturally hit budget awareness thresholds and shift conservative sooner.

Planned Features

  • PreCompact hook: Save a context summary before compaction so post-compaction sessions retain awareness of what was loaded.
  • Read tracking: PostToolUse hook that counts Read calls and warns when context is getting heavy.

Known Issues

  • No runtime model detection: The skill cannot programmatically verify it's running on Opus 4.6. If loaded on a smaller-context model, the aggressive reading rules may fill context faster than intended. The budget awareness heuristics mitigate this.

Links