Skip to content

Repository files navigation

GPT 5.6 CODEX MARS PLUGIN

Mars Cost Router — three deliberate routes meeting at a verified root

An independent, instruction-driven delegation policy for Codex. Route by risk and effort; verify at the root.

Independent and unofficial: “Mars” is this project's brand. OpenAI is the authority for Codex, model, and provider behavior.

Validate Version 0.3.2 Tested with Codex CLI 0.144.5 Instruction-only MIT

Mars Cost Router is a small local plugin with a three-lane policy for bounded subagent delegation. It gives the root explicit requested settings, supports bounded self-contained child messages, and keeps integration and verification with the root. The instruction-only policy makes route intent reviewable while the root evaluates returned work and, when available, native runtime metadata.

Project: userbox020/mars-cost-router

Build Week judges: follow the quickstart below, then read the Project Story.

Judge quickstart

git clone https://github.com/userbox020/mars-cost-router.git
cd mars-cost-router

Keep this shell in the cloned repository root for the install, launch, prompt, and validation steps below. Local package validation requires Python 3.10+. Node.js 24+ is required only for video-source verification and is optional for the core validator, unit tests, and compile check.

1. Install, list, and inspect features

codex plugin marketplace add userbox020/mars-cost-router
codex plugin add mars-cost-router@mars-plugins
codex plugin list
codex features list

Confirm that the list shows mars-cost-router@mars-plugins at 0.3.2, then launch a new interactive Codex session from the cloned repository root so discovery reloads the installed plugin:

codex --enable multi_agent \
  -c 'features.multi_agent_v2.enabled=true' \
  -c 'features.multi_agent_v2.hide_spawn_agent_metadata=false'

Multi-agent V2 is under development and is required for the five-field spawn surface used by this walkthrough. The fixed evidence and tested release path use Codex CLI 0.144.5. Separately, this exact command's feature configuration and codex features list were checked provider-free on an installed Codex CLI 0.144.6. The provider-free check reported multi_agent as stable and true, and multi_agent_v2 as under development and true. That check did not launch a child, runtime spawn, model request, or provider call. Open the plugin browser or @ surface and select Mars Cost Router. In interfaces that expose skills, activate it in a prompt with $mars-cost-router:mars-cost-router.

2. Copy, paste, and run one bounded prompt

$mars-cost-router:mars-cost-router You are running from the cloned mars-cost-router repository root. For this bounded demonstration, delegate one read-only child to inspect this clone's README.md and docs/INSTALL.md and report the repository-relative locations of the install, refresh, and uninstall instructions. Use a generic task label, send only the minimum necessary context, and require the child not to delegate or spawn another agent. Do not edit files, access credentials, use the network, commit, or publish. At the root, verify every reported location in this clone before answering.

Inspect the emitted spawn request before authorizing or relying on the child. For this Economy-shaped read-only task, the request should contain all five fields: a generic task_name, a bounded self-contained message, model: "gpt-5.6-terra", reasoning_effort: "low", and fork_turns: "none". The message should include the no-nested-delegation constraint. Provider/runtime metadata remains the authority for effective execution; the payload records the requested settings.

3. Validate the checkout locally

Run from the repository root:

python scripts/validate_plugin.py
python -m unittest discover -s tests -v
node video/verify.mjs
python -m compileall -q scripts tests

The Node command verifies the video source and may be skipped for core package validation; the validator, unit tests, and compile check require Python 3.10+.

4. Refresh, troubleshoot, or uninstall

Refresh the marketplace entry and plugin, then start a new session:

codex plugin remove mars-cost-router@mars-plugins
codex plugin marketplace upgrade mars-plugins
codex plugin add mars-cost-router@mars-plugins
codex plugin list

If discovery or version output is stale, confirm the marketplace entry, repeat the refresh sequence, and start another new session. Model availability, permissions, and authorization are separate runtime checks; return availability or authority problems to the root rather than substituting a lane. Refresh keeps the marketplace registered. A complete uninstall removes both the plugin and its marketplace entry:

codex plugin remove mars-cost-router@mars-plugins
codex plugin marketplace remove mars-plugins

Use the concise Playbooks for adaptable workflows, and see Install and troubleshooting for setup, refresh, discovery, availability, and permission guidance.

A reviewable delegation toolkit

Strength How it supports a clear handoff
Decision-first lane policy Root/no-delegation comes first; Premium risk takes precedence, Economy requires every low-risk condition, and Balanced is the delegated default.
Bounded handoffs Self-contained messages define one objective, scope, acceptance criteria, and an optional reviewable return shape.
Cause-aware recovery The root distinguishes weak or conflicting evidence, malformed calls, missing authority, unavailable models, and increased risk through explicit reviewed decisions.
Owned decomposition The guidance avoids overlapping parallel writer scopes, waits for reviewed prerequisites, and keeps final integration at the root.
Root acceptance The root inspects returned evidence, resolves conflicts, and runs final verification before making final claims.
Practical playbooks Four adaptable workflows cover bounded lookup, focused implementation, security review, and dependent work.
Inspectable, privacy-conscious package A compact skill and versioned policy use generic labels and minimum-necessary context in a static package.
Verified release path Local installation guidance, package validation, and hosted CI provide a direct path to inspect the release.

Three deliberate lanes

Lane Good fit Requested model and effort Context
Economy bounded inspection, lookup, low-risk checks gpt-5.6-terra · low fork_turns: "none"
Balanced focused implementation, tests, documentation, review gpt-5.6-terra · medium fork_turns: "none"
Premium security boundaries, difficult debugging, broad uncertainty gpt-5.6-sol · high fork_turns: "none"

Terra is the lower-listed-rate lane under the dated Standard API rate source. Lane choice follows the decision protocol above; recovery remains a reviewed root decision rather than an automatic retry or substitution.

How it fits

Delegation flow: policy informs the root; the root makes bounded child calls and verifies returned work.

The skill and versioned policy give the root a clear request shape. The root writes the child request, receives the result, and owns final verification. Native child metadata, when available, gives the root runtime detail to inspect alongside the returned work.

Sanitized child request shape

{
  "task_name": "focused_check",
  "message": "Inspect one bounded area. Return findings only. Do not delegate or spawn another agent.",
  "model": "gpt-5.6-terra",
  "reasoning_effort": "medium",
  "fork_turns": "none"
}

Use task_name as a short, generic, privacy-safe label. Put the bounded objective, scope, permissions, acceptance criteria, and return format in message; this keeps each handoff clear for the child and reviewable by the root.

Watch the 0.3.1 walkthrough

Watch the 0.3.1 walkthrough · Script · Captions · Recording checklist · Terminal commands

The video highlights explicit lanes, bounded requests, inspectable instruction-only architecture, sanitized presentation, verified installation/CI, and the fixed-series record. See Evidence for scope and methodology.

Fixed v1.2 descriptive evidence

Record classification: descriptive-synthetic.

Fixed v1.2 descriptive comparison with stacked child and root token bars plus separate 12 of 12 deterministic-check panels for both treatments.

Three precommitted pairs of four fixed read-only tasks recorded against Codex CLI 0.144.5:

Recorded observation Selective Terra/Sol policy All-Sol/high baseline
Deterministic checks 12 / 12 12 / 12
Observed automatic retries 0 0
Observed reroutes 0 0
Child tokens 356,116 356,494
Root tokens 372,590 412,418
Total tokens 728,706 768,912
Median wall duration 45.094 s 53.328 s

Both arms recorded 12 / 12 deterministic checks, with zero observed automatic retries and reroutes. Child-token totals were nearly flat: 356,116 versus 356,494 (-0.11% recorded). Most of the recorded total-token difference occurred at the root; total tokens were 728,706 versus 768,912 (-5.23% recorded). Evidence contains the fixed-series scope, methodology, and interpretation.

Mars’s published comparison normalizes dated Standard API listed rates as of 2026-07-17 to Terra 50 and Sol 100. Evidence covers the source, method, recheck guidance, and the distinction from ChatGPT credits.

Dated Standard API rate index: Terra 50 and Sol 100.

The sanitized machine-readable fixed-v1.2 summary and rate index provide the compact public record. Evidence links the definitions, provenance, boundaries, and Heldout-v2 context.

Evidence scope

Lane fields are instruction-driven requested settings, checked by the root before each handoff. Native child metadata, when available, is the authority for effective runtime detail. Model availability is account-, CLI-, and environment-specific. Static validation establishes package conformance; the published fixed suite supplies descriptive outcomes for its stated tasks and runtime evidence supplies execution observations.

Security and privacy by design

The minimal, inspectable package centers on an instruction skill and versioned policy. Use generic labels, bounded child messages, and sanitized evidence when presenting or sharing work. Privacy and security guidance covers safe delegation inputs and evidence hygiene.

Development

Read Architecture, Install, and Evidence before changing policy or public claims. This presentation intentionally distinguishes requested settings from observed runtime facts.

License

MIT. See the repository license.

About

Instruction-driven model lanes for bounded Codex subagent delegation.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages