Make code changes you can explain.

Rootloom gives coding agents a disciplined path from repository evidence to a focused change and verification you can inspect.

codex plugin marketplace add liyanqing90/rootloom codex plugin add rootloom@rootloom

These commands install the native Codex plugin. A separate Agent Plugins preview provides portable Change, Review, and Project Guidance.

A black cat at a loom weaving tangled risks, defects, and context into a checked result

Most coding failures happen before the diff.

An agent can write plausible code at the wrong boundary, run one convenient test, then report more certainty than the evidence supports. Rootloom changes that sequence.

Patch the symptom

The error line gets a guard; the component that owns the behavior stays broken.

Test the happy path

Reconnect passes while cancellation and clean disconnect quietly drift.

Report a result

Suggested checks and executed checks blur into the same sentence.

Rootloom makes the missing middle visible.

The patch is one step in an engineering decision, not the whole decision.

Request
Repository evidence
Risk and scope
Owning boundary
Focused change
Verified report

From request to evidence, one owned step at a time.

Ordinary work stays on the repository's normal edit-and-test path. Rootloom supplies the discipline around it.

  1. 1

    Read the project

    Inspect source, tests, local rules, and existing work before editing.

  2. 2

    Set risk and scope

    Match the workflow to the blast radius and name what should not change.

  3. 3

    Find the owner

    For a defect, trace the trigger to the boundary that owns the invariant.

  4. 4

    Make the focused change

    Use the repository's architecture and preserve unrelated work.

  5. 5

    Verify and report

    Exercise the main path, the invariant, and an adjacent path. Name what actually ran.

Rootloom loom diagram: risk, defects, and context are processed with evidence, a contract, and tests into a verified result
The illustrated loom is the product model: messy inputs become a reviewable result through evidence, scope, and tests.

One Core, four public entries.

Change routes risk and evidence modes internally, so users choose by intent instead of governance machinery.

Daily change

$operating-coding-change

Build, fix, or refactor ordinary code with scoped verification.

Review only

$operating-code-review

Inspect a diff, PR, migration, or design without changing files.

Project guidance

$project-guidance

Seed, refresh, refine, or validate concise AGENTS.md guidance.

Setup

$setup-rootloom

Install, upgrade, inspect, or roll back optional global guidance and Autonomy Rules.

Need a governed change or machine-readable evidence bundle? $operating-coding-change loads that mode explicitly and only when needed.

Compare every Skill

Completion should say what happened.

Rootloom keeps plans, command results, repository state, and human judgment as separate facts.

Plans stay labeled as plans.

A generated verification command is a suggestion until it actually runs.

State is checked after commands finish.

Exit code 0 is insufficient when the repository or captured evidence changed.

Evidence supports judgment.

It makes a review inspectable; it does not prove that code is correct or secure.

Start with the native Codex plugin.

Use these commands for the complete Codex experience. Agent Plugins-compatible clients can load the three-Skill preview from portable/rootloom; see the preview guide below.

1

Add the marketplace

codex plugin marketplace add liyanqing90/rootloom
2

Install Rootloom

codex plugin add rootloom@rootloom
3

Open a new task

$operating-coding-change
Fix the reconnect race and verify reconnect, clean disconnect, and cancellation.

Global setup is optional.

Core installation does not write ~/.codex/AGENTS.md, enable Hooks, install Rules, or install Rootloom Memory.

Read setup and rollback

Useful because it stays narrow.

Rootloom owns the execution and review boundary. Existing engineering tools keep doing their own jobs.

Rootloom is

  • A native Codex plugin plus Agent Plugins preview
  • Single-agent by default
  • Python-standard-library-only at runtime
  • Lightweight for ordinary changes

Rootloom is not

  • A specification framework or test runner
  • A secret scanner or sandbox
  • Proof that a change is correct or secure
  • Feature parity across every coding-agent client