claude-performance

claude-performance

A Claude Code plugin that turns session telemetry into self-improving behavioral rules.

Reads JSONL session data from ~/.claude/projects/, computes six effectiveness metrics, and when diagnostics fire, writes permanent rules into ~/.claude/CLAUDE.md so future sessions adapt. Measurement layer for self-improving AI workflows.

A static rule is a wish. A measured rule is a system.

Companion to claude-daily-journal and claude-insights.

What it measures

Every seven days, reads all Claude Code session files from the last week and computes:

1. Activity distribution: Coding, Exploration, Debugging, Delegation, Planning, Conversation
2. One-shot edit rate: percentage of file edits that land without a retry cycle
3. Agent spawn analysis: how many subagents fire and how many turns each takes
4. Model mix: Opus vs. Sonnet vs. Haiku across all turns
5. Project allocation: which codebases consume the most attention
6. Hookify firings: which behavioral guardrails actually trigger

What it prescribes

Six diagnostic rules. When a rule fires, one of two things happens:

• Behavioral prescriptions (verbose agents, model routing, low one-shot rate, exploration overhead) get written directly into ~/.claude/CLAUDE.md as permanent rules Claude reads on every future session start.
• Investigation prescriptions (recurring tool errors, hookify repeats) get appended to a Claude To-dos list for the user to review.

Next week the digest re-measures. If the number moved, the rule worked. If it did not, the rule fires again with updated numbers. If it sits at target for weeks, the rule can probably retire.

Install

Open Claude Code, paste:

/plugin marketplace add adelaidasofia/claude-performance
/plugin install claude-performance@claude-performance

Then copy scripts/claude_performance_digest.py into your vault at <vault>/⚙️ Meta/scripts/. The script self-locates via __file__ and expects that path.

claude plugin add github.com/adelaidasofia/claude-performance

Or clone manually:

git clone https://github.com/adelaidasofia/claude-performance ~/.claude/plugins/claude-performance
claude plugin add ~/.claude/plugins/claude-performance

Usage

Run manually:

python3 "<vault>/⚙️ Meta/scripts/claude_performance_digest.py"

Or schedule weekly (Monday 1am UTC example):

0 1 * * 1 /usr/bin/python3 "/path/to/vault/⚙️ Meta/scripts/claude_performance_digest.py"

Flags

• --days N: lookback window (default 7)
• --dry-run: print report to stdout, do not write files
• --no-report: skip the markdown report, apply prescriptions only

Configuration

At the top of scripts/claude_performance_digest.py:

THRESHOLDS = {
    "one_shot_min": 0.75,
    "exploration_max": 0.35,
    "agent_turns_max": 5,
    "opus_max": 0.70,
    "hookify_repeat": 10,
    "recurring_error_sessions": 3,
}

PROJECT_LABELS = {
    # Optional: map project directory substrings to clean display labels
    # "my-company": "CompanyName",
}

Output

A dated markdown report at ⚙️ Meta/Performance/weekly-YYYY-MM-DD.md. And if a behavioral diagnostic fires, one or more rules appended to ~/.claude/CLAUDE.md like:

- [VERBOSE AGENTS fix](performance_verbose_agents.md) | Agent briefings must
  include: specific file paths, expected output format, and scope boundary.
  Target: <8 turns per agent. Current avg: 22.0. (updated 2026-04-16)

Why this exists

Self-improvement by memory alone has a failure mode: a rule gets written, Claude reads it at session start, and under the wrong context the behavior recurs anyway. Without measurement, you cannot tell whether a correction actually worked.

This plugin closes the loop. Rules come with a number attached. Baseline, target, check-in. The script runs weekly on its own and the next session reads the new rule.

Full story: I Taught Claude to Write Its Own Rules (coming soon).

Telemetry

This plugin sends a single anonymous install signal to myceliumai.co the first time it loads in a Claude Code session on a given machine.

What is sent:

• Plugin name (e.g. slack-mcp)
• Plugin version (e.g. 0.1.0)

What is NOT sent:

• No user identifiers, names, emails, tokens, or API keys
• No file paths, message content, or anything from your work
• No IP address is stored after dedup processing

Why: Helps the maintainer know which plugins people actually install, so attention goes to the ones that get used.

Opt out: Set the environment variable MYCELIUM_NO_PING=1 before launching Claude Code. The hook will skip the network call entirely. Already-pinged installs leave a sentinel at ~/.mycelium/onboarded-<plugin> — delete it if you want to reset state.

License

MIT