Free forever ♡
Portable, individually installable skills and plugins for AI coding agents by Slopware Engineer.
| Package | Type | Purpose |
|---|---|---|
| MSW Kernel | Skill and plugin | Admit only necessary work, require proof, and stop at the fixed point. |
| MSW Hook | Optional plugin | Reinforce the MSW Kernel once at session context boundaries. |
| MSL Kernel | Skill and plugin | Report only what the reader must act on, in a form they can check. |
| Timebox | Skill and plugin | Converge authorized work inside AWT and a closeout-only CGP. |
| CODER Loop | Codex skill and plugin | Optimize Codex development through bounded work, fresh evaluation, and evidence-scoped remediation. |
| Codex Voice Optimizer | Codex skill and plugin | Optimize Codex Voice for hands-free orchestration across owning work threads. |
Each package has its own documentation and remains independently installable.
A modular Codex-native stack that removes waste at every development boundary. Install one layer or compose the stack. Free forever.
| Component | Stack role | Contribution |
|---|---|---|
| Codex Voice Optimizer | Control plane | Hands-free intent, routing, authority, and material progress |
| CODER Loop | Development engine | Bounded implementation, independent evaluation, remediation, acceptance, and evolving model routing |
| MSW Kernel | Scope kernel | Necessary coordination, work, findings, repairs, and proof |
| Timebox | Convergence envelope | One explicitly authorized AWT/CGP clock around the work |
| MSL Kernel | Communication kernel | The final facts the user must understand or act on |
The product story is simple: CVO controls it. CODER builds it. MSW keeps it necessary. Timebox makes an authorized run converge. MSL makes the result clear. Every package works alone; together they form an evidence-evolving, high-efficiency agentic development system without a hard dependency or shared runtime.
All necessary work. Nothing beyond it.
Package documentation · Skill source
Minimum Sufficient Work is the principle. The MSW Kernel is the compact instruction set and program that applies the principle to agent work.
The MSW Kernel is a lightweight alternative to heavier, more invasive agent-control systems. Its small instruction surface can be just as effective, and arguably more effective, because it gives the agent one necessity test, a proof obligation, and a clear stopping condition without layering on another workflow.
It applies to implementation, debugging, review, research, planning, and documentation, not just code size. The original MSW Kernel post introduced the idea.
This uses a free-form Git marketplace directly from this repository. It does not require submission to or verification for an official marketplace.
codex plugin marketplace add transcendr/slopware-skills
codex plugin add msw@slopware-skillsAdd automatic reinforcement only if you want it:
codex plugin add msw-hook@slopware-skillsStart Codex, open /hooks, and trust the single SessionStart hook. The
msw plugin remains hook-free.
claude plugin marketplace add transcendr/slopware-skills
claude plugin install msw@slopware-skillsAdd automatic reinforcement only if you want it:
claude plugin install msw-hook@slopware-skillsThe msw plugin remains hook-free.
This installs the portable, hook-free Agent Skills version:
npx skills add transcendr/slopware-skills --skill msw -gThe installer supports Codex, Claude Code, and other Agent Skills-compatible clients.
git clone https://github.com/transcendr/slopware-skills.git
mkdir -p ~/.agents/skills
cp -R slopware-skills/plugins/msw/skills/msw ~/.agents/skills/mswThe separate msw-hook plugin injects one compact MSW reminder when a session
context starts, resumes, clears, or compacts. It does not run before tools,
after edits, or on every user prompt. There are no modes, daemons, dependencies,
status lines, or state files.
Install only msw when you want explicit invocation with no hook.
Ponytail and similar systems can introduce modes, decision ladders, persistent reinforcement, invasive hooks, and highly opinionated conventions. The MSW Kernel stays at a lower level: define the contract, test necessity, prove the required work, and halt. This avoids overthinking, process doom spirals, and conventions that can clash with your own while remaining applicable beyond coding tasks.
The contract converts an open-ended request into the smallest provable outcome. The deletion test forces every proposed action to justify itself against that outcome instead of habit, reviewer authority, or imagined future needs. The proof obligation prevents minimalism from becoming corner-cutting. The fixed point gives the agent a concrete stopping condition, so it neither stops early nor keeps expanding the task.
Absolutely, if you want MSW as a standing project or global instruction. It
will not permanently change model behavior; MSW remains principally opt-in and
usually benefits from reinforcement during normal work and discussion with the
agent. The optional lightweight msw-hook can reduce how often you need to
reinforce it, but evaluate the hook carefully in your own environment before
enabling it.
Apply the MSW skill at the start of the task. Whenever the agent is planning, implementing, working through a test and remediation loop, or considering new code, artifacts, or process, reinforce it with one line:
Remember to follow the MSW deletion rule for all claims: no exceptions.
Repeat that line when needed during normal work and discussion. It restates the kernel's necessity test without adding persistent machinery.
Everything the reader must act on. Nothing they must wade through.
Package documentation · Skill source
Minimum Sufficient Language is the principle. The MSL Kernel is the compact instruction set and program that applies the principle to anything an agent writes for a reader.
MSL is the sibling of MSW. MSW decides what work is necessary; MSL decides what language is necessary to report it. Both are kernels rather than workflows, and neither imposes a template.
Use $msl to write this update:
<facts>
The kernel binds the reader first, including everything already on their record, then partitions every fact before a sentence is written. A fact reaches the reader only by changing what they believe about the state of the world, how tightly a risk was bound, or what decision remains. Everything describing how the work was carried out, including tools, stages, superseded designs, abandoned attempts, and cases tested, is instrument. The instrument column is closed while drafting.
That partition is the whole trick, and it is a gate rather than a rule. Telling an agent to avoid narrating its tooling does not hold, because contrast and step-by-step feel like proof that the work was careful. Removing those facts from view before drafting does hold.
The same program serves a progress note to your own user and a closeout comment, a pull request write-up, an incident note, or an email to someone outside the company. Only the reader changes. Length is never imposed; it falls out of the partition.
codex plugin marketplace add transcendr/slopware-skills
codex plugin add msl@slopware-skillsclaude plugin marketplace add transcendr/slopware-skills
claude plugin install msl@slopware-skillsnpx skills add transcendr/slopware-skills --skill msl -ggit clone https://github.com/transcendr/slopware-skills.git
mkdir -p ~/.agents/skills
cp -R slopware-skills/plugins/msl/skills/msl ~/.agents/skills/mslRequired work inside an authorized clock.
Package documentation · Skill source
Timebox is a hook-free execution protocol for work that must converge inside a requester-supplied clock. Available Work Time (AWT) contains the substantive work and its proof. Closeout Grace Period (CGP) is a shorter buffer reserved for an already-converged deliverable, a delayed final check, narrow organization, or communicating the result.
For most users, the experience is simple:
Use $timebox with AWT 45 minutes and CGP 5 minutes.
<task>
When the AWT/CGP context is already clear, shorthand works too:
Timebox this 45/5: <task>
The agent validates the pair, records fixed deadlines, states the outcome and smallest proof, then works normally. It quietly checks whether the required result will finish inside AWT, cuts optional depth when necessary, reserves CGP for closeout, and stops with an honest status at the hard stop. If the task is complete and proven early, it stops early.
The optional independent-monitor design was created with the Codex app in mind. Codex can use a separate monitor task, task reading and messaging, and a same-task heartbeat to watch the fixed clock without interrupting the working task. The same approach can work in any agent harness with equivalent observation, wake, and messaging capabilities. Everyone else gets the lightweight self-monitoring workflow with no background process.
The package has no hook. A valid timebox depends on the requester-supplied durations, actual start, task contract, and live progress. Session and tool hooks do not have that complete context, so an always-on hook would be more invasive and less reliable than the skill itself.
codex plugin marketplace add transcendr/slopware-skills
codex plugin add timebox@slopware-skillsclaude plugin marketplace add transcendr/slopware-skills
claude plugin install timebox@slopware-skillsnpx skills add transcendr/slopware-skills --skill timebox -ggit clone https://github.com/transcendr/slopware-skills.git
mkdir -p ~/.agents/skills
cp -R slopware-skills/plugins/timebox/skills/timebox ~/.agents/skills/timeboxCodex Optimized Development, Evaluation, and Remediation.
Package documentation · Skill source
The CODER Loop turns complex implementation into coherent task families with one owner each, then sends every completed candidate through a fresh independent review. Only observations that leave an acceptance claim unmet or unsupported become repair work. After a repair, the CODER Loop rechecks the claims that could have changed instead of discarding unrelated evidence or repeating the entire review for show.
The parent remains orchestration-only and owns final acceptance. Workers own implementation, reviewers stay read-only, and delegation never expands commit, publication, deployment, or other external-write authority.
The CODER Loop uses collaboration subagents for its normal worker, repair,
integration, and review roles. It can pair inherited working context with a
selected model and reasoning effort through the fork_turns: "99" Subagents
V2 route, while fresh independent reviewers start without inherited reasoning.
With separate user permission, top-level Codex tasks remain available for genuine lifecycle or placement needs. Temporary fallback tasks are archived after their terminal result and evidence have been captured. A postmortem then uses the completed loop evidence to calibrate future model routing without weakening independent review or changing a default from one success.
The CODER Loop has no fixed agent count or hard-coded model roster. If the task is one coherent family, it uses one worker and one fresh reviewer. If Codex cannot supply a fresh review context, the skill stops and does not pretend that self-review is independent.
Start with Start the CODER Loop tutorial for a teaching-only walkthrough, or
say Show me the Slopware Dev Stack to see how MSW, Timebox, MSL, and CVO
compose around the loop. CODER announces available companions once and can
install exact requested layers with permission.
This installs from the repository's free-form Git marketplace. It does not require submission to the official plugin directory.
codex plugin marketplace add transcendr/slopware-skills
codex plugin add coder-loop@slopware-skillsnpx skills add https://github.com/transcendr/slopware-skills/tree/main/plugins/coder-loop/skills/coder-loop -g -a codexgit clone https://github.com/transcendr/slopware-skills.git
mkdir -p ~/.agents/skills
cp -R slopware-skills/plugins/coder-loop/skills/coder-loop ~/.agents/skills/coder-loopYou speak. The coordinator routes. The work threads own the work.
Package documentation · Skill source · Original X post
Codex Voice Optimizer upgrades one ChatGPT Voice task into an ear-first control surface for engineering work. Its always-active base layer keeps spoken updates concise, routes substantive work to owning threads, verifies current task state, preserves explicit authority, and lets you interrupt task waiting simply by speaking.
Free-form orchestration is the default. Named workflows install additional structure only when invoked: Freeway derives parallel work lanes from genuinely independent work, Decision Walkthrough closes material authority choices before implementation, and Tutorial explains the system before any optional live demonstration. Every workflow inherits the optimized base behavior and role contracts.
For codebase work, the coordinator can discover the relevant Codex project and place every newly authorized owning thread there. Project selection establishes placement; it never authorizes task creation.
The change-and-research topology from the original workflow remains a reference example, not a default or limit. Work threads can communicate directly when their lanes have a real dependency, while material progress returns through the voice coordinator.
Codex Voice Optimizer works alone as the voice control plane. It automatically composes available MSL and MSW kernels, exposes CODER as the opt-in development engine, and exposes Timebox as the opt-in convergence envelope.
You -> CVO routing -> CODER development when invoked
-> MSW work admission -> Timebox convergence when authorized
-> task evidence -> MSL fact admission -> CVO speech
On activation or first workstream setup, CVO announces active kernels and available development layers once and continues. With permission, it can install MSL, MSW, CODER, Timebox, or the full Slopware Dev Stack, verify the requested state, and tell you to start a new Codex task. There is no hard dependency, hook, background installer, or repeated prompt.
Companion contract · MSL Kernel · MSW Kernel · CODER Loop · Timebox
Use $codex-voice-optimizer to coordinate this workstream through my existing
Codex tasks.
This work belongs to my <project> Codex project. Keep any new work tasks there.
The full experience requires ChatGPT Voice in the Codex desktop app and access to project discovery, project-scoped task creation, and persistent task listing, reading, messaging, and interruptible waiting. The package is intentionally absent from the Claude Code marketplace because its contract is Codex-specific.
codex plugin marketplace add transcendr/slopware-skills
codex plugin add codex-voice-optimizer@slopware-skillsnpx skills add https://github.com/transcendr/slopware-skills/tree/main/plugins/codex-voice-optimizer/skills/codex-voice-optimizer -g -a codexThe direct GitHub path keeps this Codex-only skill independently installable without adding it to the Claude Code marketplace.
git clone https://github.com/transcendr/slopware-skills.git
mkdir -p ~/.agents/skills
cp -R slopware-skills/plugins/codex-voice-optimizer/skills/codex-voice-optimizer ~/.agents/skills/codex-voice-optimizerThe plugin is hook-free. The coordinator role must be bound explicitly to one voice task, a current workstream, named owners, and exact authority; a global lifecycle hook cannot determine those safely.
.agents/plugins/marketplace.json Codex marketplace
.claude-plugin/marketplace.json Claude Code marketplace
plugins/
msw/
.codex-plugin/plugin.json
.claude-plugin/plugin.json
README.md
skills/msw/SKILL.md
msw-hook/
.codex-plugin/plugin.json
.claude-plugin/plugin.json
README.md
hooks/hooks.json
msl/
.codex-plugin/plugin.json
.claude-plugin/plugin.json
README.md
skills/msl/SKILL.md
timebox/
.codex-plugin/plugin.json
.claude-plugin/plugin.json
README.md
skills/timebox/
SKILL.md
references/independent-monitor.md
coder-loop/
.codex-plugin/plugin.json
README.md
skills/coder-loop/
SKILL.md
references/
companions.md
codex.md
postmortem.md
review-kernel.md
role-contracts.md
workflows/
tutorial.md
codex-voice-optimizer/
.codex-plugin/plugin.json
README.md
skills/codex-voice-optimizer/
SKILL.md
references/
codex-app.md
companions.md
voice-coordinator.md
work-thread.md
workflows/
decision-walkthrough.md
freeway.md
tutorial.md
Each future package gets its own plugins/<name>/ directory and marketplace
entry, so skills and plugins remain individually installable while living in
one repository.
CC BY 4.0: use, adapt, redistribute, and commercialize these skills
however you want; keep credit to Slopware Engineer / @aienginerd.
Free forever ♡