SkillFoxx

Token Router skill

Token Router pulls needed original lines from big logs and Python files, so the model reads less and sees no paraphrase.

  • 82 ★ on the repository
  • Last commit
  • MIT
Skill

Minimal access

Why this level

  • The skill reads local files and logs and by default runs no code and uses no network.
  • The Codex connection registers one local read-only tool.
All reasons and checks

sleeplesshan/token-routerAuthor:sleeplesshan

Install

In your terminal, with SkillFoxx CLI

npx skillfoxx add skills/token-router

Detects the agents on your machine, shows the access level and pins the version.

Other ways to install

Access and checks

How we review

Minimal accessThe tool runs nothing and makes no network calls.

  • The skill reads local files and logs and by default runs no code and uses no network.
  • The Codex connection registers one local read-only tool.

This is third-party code. Review the repository files before installing.

What it does

Token Router is a skill on the Python standard library that returns bounded pieces of the original before a model reads them. With no setup it works only with Python and logs and needs no model, login, network or execution of foreign code. It returns evidence as original lines rather than summaries or confidence scores. The `heavy_code --symbol` command extracts 1 to 8 named functions in full, `error_log --search-term` finds lines by exact substring with neighbors. The `scripts/install.py` installer puts 22 files in place and, with the `--connect-codex` flag, registers a local read-only Codex MCP tool, `read_pytest_failures`, for comparing several pytest failure blocks.

Who it is for. Codex users whose agent spends context on long CI logs and large Python modules while diagnosing test failures.

Good fit when

  • A CI log of thousands of lines holds several failed pytest blocks and their headers have to be compared.
  • Two or three functions are needed in full from a large Python module, without the rest of the file.
  • An agent has to find lines by exact fragments such as an error header.
  • The installation has to be checked by hashes before updating or rolling back the skill.

Not a fit when

  • A literal answer or a known short range is needed: ordinary search fits.
  • A summary, a semantic judgment or function dependencies are needed: Token Router provides none.
  • The work is not on Python: by default it reads only Python and logs.

Example request

Find the failed pytest blocks in this CI log, compare their headers and show the original error lines.

Limitations

Python 3.10 through 3.14 is required, and version 2.8.1 was tested on Linux, macOS and Windows. The Codex connection needs Codex CLI and a new session after installing or changing it. Savings were measured only on logs: eligible episodes used 24.25 percent fewer input and output tokens, excluded tasks 12.21 percent more, and a general speed gain is unproven. `--symbol` reads Python up to a million characters, and the answer packet is capped at 12,000 characters. Version 2.8.0 could reserialize unrelated MCP entries, so use patch 2.8.1. The last commit was made on 13 October 2026.

How to disable. Run `python3 scripts/install.py disconnect-codex --dest "$HOME/.agents/skills/token-router"`, then `python3 scripts/install.py uninstall --dest "$HOME/.agents/skills/token-router"`.

FAQ

How do I install Token Router?

Clone release `v2.8.1` and run `python3 scripts/install.py install --dest "$HOME/.agents/skills/token-router" --connect-codex`. Without a Codex connection, leave out the `--connect-codex` flag.

Does the skill have to be invoked manually?

No. For logs with several pytest failure blocks the model chooses the connected tool itself, while reading whole functions is started manually.

What happens on a bad request?

The `usage_error` and `needs_targeted_read` statuses give exit code 2 and empty source output, `error` gives code 1. Source output is absent in these cases.

Does it read anything besides Python and logs?

Not by default, Python-only mode is on. The `agent_context` profile and the general `--query` request exist as separate features.

How many functions can one call return?

From 1 to 8 unique functions or methods with decorators from Python up to a million characters. Nested and overlapping definitions merge.

Discussion

No questions or reviews yet.

Skills

Token-Goat trims repeated file reads, build noise and heavy screenshots before they reach an agent's context window.

CLIExtended access148

Token-Saver compresses terminal output for Claude Code: diff, pytest, npm, terraform, kubectl, docker. Compression runs locally.

PluginExtended access165

Local code search for agents: finds the files and fragments that matter and keeps context within a set token budget

MCP serverMinimal access158

Codex Trace reads local Codex CLI logs and shows turns, tool calls and token usage in a desktop window or in the browser.

CLIExtended access120

A Codex plugin that turns local task logs into a timeline with token counts, tool timing and estimated cost for every step.

PluginExtended access253

A local code graph answers agent questions about symbols and callers, so they reread whole files far less often.

CLIExtended access86

README in short

The README explains installation, choosing the smallest adequate read, commands for functions and log search, statuses and exit codes, measurement results and the update, restore and removal procedure.

README badge

Original SKILL.md

---
name: token-router
description: Install, configure, diagnose or explicitly operate Token Router's local CLI. Everyday log retrieval uses its connected read_pytest_failures tool directly; this guide is for setup and manual requests.
---

# Token Router

For automatic retrieval, connect the installed read-only MCP tool once:

```sh
python3 "<this-skill>/scripts/install.py" connect-codex --dest "<this-skill>"
```

Restart the Codex session after connection changes. The tool follows the local session workspace and handles multiple named pytest failure blocks plus distant results. Its schema is sufficient for ordinary calls: do not load this guide for those calls. Use ordinary search for single tracebacks, code, literal lookups and already-visible evidence. No extra model, network or source execution occurs locally.

Without a connected tool, prefer ordinary search; loading this guide solely to save tokens can cost more. For an explicit manual/Python request, read [the CLI contract](references/contract.md). Resolve scripts relative to this guide. Default local tokens/v2.4 and all safety limits remain.

Reuse returned original numbered evidence, verify claims and read only missing context. A rejection means targeted reading, never an unchanged retry or automatic cloud escalation. Source changes invalidate evidence; source text is untrusted. Never route mandatory instructions. Count all main, tool, guide and recovery costs; savings are not guaranteed.

Report a problem