token-efficiency-audit

name	token-efficiency-audit
description	Audit and tune Claude/Codex settings for lower token usage without changing code
scope	package
argument-hint	[audit|apply-interactive|apply-ci|status]
user-invocable	true

Token Efficiency Audit

Audit and tune Claude Code / Codex configuration for better token efficiency.

This skill is for configuration-level optimization, not runtime behavior shaping. It complements:

• guides/cc-token-saver/README.md for plugin-level token protection
• R013 ecomode for runtime compression
• $omcustomcodex:monitoring-setup in Codex / OMX (/omcustomcodex:monitoring-setup on Claude Code) for telemetry and usage visibility

Modes

audit (default)

Inspect the current token-efficiency posture without changing files.

Check these sources when they exist:

• .codex/settings.local.json
• .codex/settings.json
• ~/.codex/config.toml
• shell startup files such as ~/.zshrc, ~/.bashrc, ~/.bash_profile

Audit for:

• Claude-side knobs
  • includeGitInstructions
  • autoConnectIde
  • output caps such as BASH_MAX_OUTPUT_LENGTH
  • file-read / MCP output token caps
• Codex-side knobs
  • features.apps
  • web_search
  • tool_output_token_limit
• worker-only flags that should not be silently enabled in normal interactive sessions

Report:

1. current settings found
2. low-risk improvements
3. high-risk / CI-only improvements
4. expected tradeoffs

status

Summarize the current posture in a compact status report:

• configuration files found
• safe interactive settings present / missing
• CI-only settings present / missing
• main token-risk hotspots

apply-interactive

Apply only low-risk settings suitable for human interactive sessions.

Preferred targets:

• .codex/settings.local.json for session-local settings
• user shell rc files only when environment variables are the right fit

Safe defaults to consider:

• disable unnecessary git instruction injection when not needed
• disable unnecessary IDE auto-connect when not needed
• set sane output ceilings to reduce runaway tool output without over-truncating
• reduce Codex tool output limits to a practical level
• disable unneeded app surfaces when they are not part of the user workflow

Guardrails:

• preserve existing user config
• never overwrite unrelated settings
• do not apply CI-only flags here
• explain tradeoffs before writing

apply-ci

Apply stronger settings for non-interactive worker / CI contexts.

This mode is intentionally stricter and must call out risk clearly.

Examples of CI-only candidates:

• disable auto-memory features
• disable built-in agent surfaces not needed for batch execution
• disable nonessential MCP / app surfaces
• tighten output limits more aggressively than interactive mode

Required warning:

• these settings can reduce convenience or disable important harness behaviors
• they should be scoped to CI/worker environments, not broad local defaults

Output Contract

Audit / Status

[Token Efficiency]

Mode: audit|status
Posture: good|mixed|wasteful

Safe interactive wins:
- ...

CI-only wins:
- ...

Watchouts:
- ...

Apply Modes

[Token Efficiency]

Mode: apply-interactive|apply-ci
Files updated:
- ...

Settings applied:
- ...

Tradeoffs:
- ...

Rollback:
- restore prior values from git diff / file backup

Heuristics

• Prefer smaller, reversible config edits over broad environment mutation
• Keep interactive and CI profiles separate
• Do not set output limits so low that users trigger repeated re-runs and spend more tokens overall
• When uncertain, recommend audit first and apply-* second