tokenwise – Coding skill for AI agents

TokenWise — Measurement-Driven Model Router

Overview

The upstream package advertises model-tier routing, local NDJSON usage logs and an A/B comparison command. Evaluate those capabilities only in an installed, reviewed revision with the user's authorized model budget and actual workload. Task difficulty, response-quality rubrics and reported costs remain estimates unless independently measured.

This entry describes the upstream workflow; it does not bundle its commands or measurement implementation. Verify the installed revision, supported models and actual routing behavior before relying on a report. A user's workload and billing plan determine whether routing reduces paid usage.

When to use

• Cutting Claude Code token spend without sacrificing output quality
• Validating whether Haiku/Sonnet is "good enough" for a specific task class before trusting auto-routing
• Auditing where Opus tokens are actually being burned
• Logging per-session cost data for finance or chargeback

Subcommands

• /tokenwise:install — guided installer with diff preview, automatic backups, and --dry-run mode
• /tokenwise:report — per-session token + cost summary vs all-Opus baseline
• /tokenwise:summary [--week|--month|--all] — historical aggregate with trend
• /tokenwise:ab "<task>" — A/B test the same task at multiple tiers, generates a markdown comparison
• /tokenwise:undo — restore CLAUDE.md / settings.json from backup

Routing taxonomy

| Tier | Model | Task class |

|---|---|---|

| Mechanical | Haiku 4.5 | file reads, grep, format, rename, simple edits, doc lookups |

| Scoped reasoning | Sonnet 4.6 | single-file refactor, scoped research, test writing |

| Synthesis | Opus 4.7 | architecture decisions, multi-file refactor, security review |

Safety caps:

• Haiku never spawns further subagents
• Max spawn depth = 2
• Subagents that need a smarter model return to parent — they never escalate on their own
• Tasks under 100 chars with no file context run inline (subagent overhead > savings)
• Subagent context >30k tokens bumps a tier

Privacy

Upstream describes project-local logs at .tokenwise/log.ndjson. Inspect the installed revision and dependencies for log contents, redaction, retention and network behavior; this single-file catalog entry cannot establish zero telemetry. Model routing and A/B runs send task context to the selected provider. Exclude secrets and obtain authorization for additional provider usage; keep logs out of public commits.

Install

In any Claude Code session:

/plugin marketplace add CodeShuX/tokenwise
/plugin install tokenwise@tokenwise

Then run /tokenwise:install and follow the guided prompts.

Limitations

• Treat reported costs as estimates until reconciled with actual provider usage and current pricing; no billing-error bound is established by this catalog entry
• A/B test mode costs extra tokens (one task × N tiers) — intentional one-time validation
• Anthropic-only by design (use LiteLLM, OpenRouter, or [Bifrost](https://github.com/maximhq/bifrost) for cross-vendor)
• Verify the selected model in actual execution evidence; do not assume a model parameter or an installer probe guarantees that a specific client build honored routing

Source

• Repo: https://github.com/CodeShuX/tokenwise
• License: MIT
• Author: CodeShuX