Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

26 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Codesemble — repository evidence transformed into a small native Codex team

Turn your project and goal into the smallest useful Codex agent team.

Codesemble audits your project, proposes the specialists it needs, and gives you an exact preview, human approval, and rollback.

Quick start · How it works · Safety · Documentation

Open the interactive project guide →


Why Codesemble?

  • Understands the project — recommendations come from repository evidence and your goal.
  • Builds only justified roles — no fixed team, no role padding, no team to design by hand.
  • Keeps changes reviewable — project-local setup is previewed before apply and can be rolled back.

Codesemble is for builders and teams who want useful specialists without maintaining a multi-agent framework. It configures native Codex project files; it does not replace Codex.

Quick start

1. Install, then restart Codex

codex plugin marketplace add VAMFI/codsemble --ref 0aa84be36209e454fadfb34b96c8c3d5b3a16caf
codex plugin add codsemble@codsemble

Open a fresh Codex session so its four Codesemble skills are discovered. The pinned commit is the locally validated v0.2 privacy-hardening candidate. It is not merged, tagged, or released yet; use it only after that commit is available from the configured Git source.

2. Audit, choose, and approve

$codsemble:initialize-team Audit this workspace and help me choose the smallest useful team.

Codesemble presents Focused, Recommended, and Extended options, then shows the exact project files and configuration diff. Choose an apply-capable mode if you want changes: a preview-only plan is read-only and cannot later be applied.

Nothing is written until you approve the exact current plan with its confirmation ID or complete six-word voice challenge.

3. Restart and give the team a goal

After apply, open another fresh Codex session so the generated agents are discovered. You continue working through the primary thread:

You → primary Codex thread → bounded specialists → primary integrates and reports
Use the installed specialists where useful to review the checkout flow, fix the
highest-risk issue, and validate the result.

Specialists run only when the primary delegates separable work; they are not always running. The primary remains accountable for decisions and final claims.

Maintain the team with $codsemble:update-team, $codsemble:team-doctor, and $codsemble:rollback-team.

How it works

Project evidence + your goal → required work → project team → preview → apply
  1. Audit project signals locally.
  2. Compile required work into evidence-backed team options.
  3. Preview and apply only the exact fresh plan you confirm.

For example:

Project evidence
  GitHub Actions + delivery goal + security requirement

Recommended Codex team
  Delivery specialist       owns the delivery work
  Security specialist       reviews the security boundary
  Independent validator     verifies the high-risk path

Spawned-worker ceiling: 4

The names and number of agents change with the project and goal. Codesemble explains why every role is present before apply.

Why 111 does not mean 111 agents

The 111-entry catalog is a library of reusable capability primitives—not a team size, recommendation ceiling, or concurrency limit. Codesemble generates only the roles justified by the current project: Focused covers required work, Recommended adds independent verification only for high-risk required work, and Extended adds only closed-rule optional lifecycle work; it may equal Recommended.

Installed roles and concurrent workers are separate. A project can install 12 useful specialists while allowing only 4 workers to run at once; the primary thread is not counted in that worker ceiling.

Read the compiler contract → · Explore the primitive library →

What it creates

AGENTS.md                         bounded orchestration guidance
.codex/agents/<role>.toml         native specialist definitions
.codex/config.toml                optional project worker ceiling
.codex/codsemble/manifest.json    selected team and ownership record
.codex/codsemble/transactions/    reversible transaction history

Safety and control

Audit and preview are read-only. Apply refuses the plan if relevant evidence, runtime capabilities, ownership lineage, or output preimages changed. Codesemble does not edit personal Codex settings, mark projects trusted, collect credentials, control Codex runtime permissions, broaden the active sandbox, or publish anything.

Configuration safety → · Voice-friendly approval → · Privacy → · Threat model →

Documentation

Project status

The v0.2 code line is present on main, but the privacy-hardening candidate documented here is one local commit ahead of the recorded origin/main. It is validated locally and has not been pushed, merged, tagged, or released; v0.1.0 remains the latest release.

Validation claims remain tied to the exact tested payload and environment. See the evidence →

Contributing

Try Codesemble in an isolated project and report where its audit, proposed roles, or explanations were incomplete. See Contributing, Support, and Security.

Codesemble is an independent open-source project licensed under the Apache License 2.0.

About

Repository-aware multi-agent team builder for Codex

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages