Code Optimization Plugin

A Claude Code plugin that analyses your actual session transcripts (JSONL) to surface real token costs, waste patterns, and week-over-week improvement trends — no manual tracking required.

Tools

1. cost_summary

   Actual token usage from JSONL transcripts, broken out by session, date, and grand total.

   • Input / cache_creation / cache_read / output tokens per session
   • Top sessions by total token consumption
   • Configurable rolling window (days, default 30)
   • Filter by project slug or across all projects

   When to use: "How many tokens did we use this week?" / "Which session was most expensive?"

2. cost_trend

   Week-over-week comparison of token usage and waste metrics.

   • Per-week totals: tokens, redundant reads, large-context turns, fix-loop turns, cache hit ratio
   • Delta and % change between this week and last week
   • Plain-English interpretation: ✓ Token usage down 18% / ✗ Redundant reads up 12%
   • Configurable number of weeks (weeks, default 2)

   When to use: "How much cost optimization did we achieve?" / "Are we improving week over week?"

3. suggest_optimizations

   Analyses transcripts and returns prioritised actionable suggestions.

   • HIGH — Repeated file reads (top files read across sessions)
   • HIGH — Intra-session redundant reads (re-reads with no edit between)
   • MEDIUM — Large-context turns (cache_read > 100K tokens)
   • MEDIUM — Retry/fix loops (turns with >8 tool calls)
   • INFO — Read vs Grep tool balance

   When to use: "What are the biggest waste patterns?" / "What should I fix first?"

4. file_read_analysis

   Deep-dive on file read behaviour across sessions.

   • Top 15 most-read files with redundant read counts
   • Total redundant reads, split by: re-read after edit (valid) vs re-read without edit (waste)
   • Targeted recommendations per file

   When to use: "Which files am I re-reading unnecessarily?" / "Is the redundant read problem getting better?"

5. top_repeated_tasks

   Finds recurring task patterns across sessions using bigram clustering on human messages.

   • Top-N phrase pairs by occurrence frequency
   • Token cost attributed to each repeated pattern
   • Example message for each cluster

   When to use: "What work keeps coming back?" / "What should we automate or template?"

6. summarise

   Pure-computation before/after health report against the last time summarise was called. No LLM involved — all classification is rule-based with a 5% change threshold.

   • Verdict:
     • IMPROVING
     • MOSTLY_IMPROVING
     • MIXED
     • MOSTLY_DECLINING
     • DECLINING
     • BASELINE_SET
   • Went well — metrics that improved by >5% since last run
   • Needs work — metrics that regressed by >5%
   • Unchanged — stable metrics
   • Tracks: total tokens, cache hit ratio, redundant reads, large-context turns, fix-loop turns, Read/Grep ratio
   • Snapshot persisted to ~/.claude/cost-analysis-state.json — survives across sessions

   First call captures the baseline and returns BASELINE_SET. Subsequent calls compare against it.

   All other tools (cost_summary, suggest_optimizations, file_read_analysis) also emit a vsLastRun field showing key deltas from their own previous call — same state file, per-tool snapshots.

   When to use: "Are things better or worse than last time?" / End-of-week retrospective.

7. reset

Clears all persisted plugin state so every tool starts fresh.

• target: "all" (default) — clears both baselines and read tracker
• target: "baselines" — clears only vsLastRun snapshots for all tools and summarise
• target: "tracker" — clears only the session read tracker used by the PreToolUse:Read hook

When to use: Beginning of a new sprint/week, after a major workflow change, or when summarise baselines are stale.

Hooks

The plugin registers a UserPromptSubmit hook that fires whenever your message contains cost/token/efficiency keywords:

• cost
• token
• optim
• spend
• usage
• efficien
• how much
• waste
• redundant
• cache hit
• session log
• what did we

It injects a reminder to use the plugin tools before reading manual files.

Installation

Option a — One-Liner Installer (Recommended)

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/sarins-lab/code-optimization-plugin/main/install.sh | sh

Windows (PowerShell):

irm https://raw.githubusercontent.com/sarins-lab/code-optimization-plugin/main/install.ps1 | iex

Both scripts automatically update ~/.claude/settings.json with the marketplace entry, enable the plugin, and wire the cost-keyword reminder hook. Restart Claude Code after running.

Option B — Manual Claude Code Marketplace

Add to ~/.claude/settings.json:

"extraKnownMarketplaces": {
  "sarins-lab": {
    "source": { "source": "github", "repo": "sarins-lab/code-optimization-plugin" }
  }
},
"enabledPlugins": {
  "code-optimization-plugin@sarins-lab": true
}

Option C — Direct MCP via Npm (no Marketplace Needed)

Add to your .mcp.json or Claude Code MCP settings:

{
  "mcpServers": {
    "cost-analysis": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@sarins-lab/code-optimization-plugin"]
    }
  }
}

Claude Code will pull the latest version from npm automatically on each start.

Option D — from Within Claude Code (no Terminal Needed)

1. Open Claude Code
2. Run /plugins → Navigate to Marketplace tab → select Add MarketPlace
3. Enter: sarins-lab/code-optimization-plugin
4. Navigate to Disover tab and enable the plugin
5. Restart Claude Code

This is equivalent to Option B but done entirely inside Claude Code without editing any files.

Adding New Tools

All tools live in scripts/analysis-server.mjs. To add a new tool:

1. Add the tool definition to the TOOLS array (name, description, inputSchema)
2. Implement the function — use loadSessions(project) and extractTurns(session) as the data layer
3. Register it in the fns map inside the request router
4. Bump the version in package.json and .claude-plugin/plugin.json