Agent skill · Public v2
Organized Chaos
Give each agent a bounded job. Use evidence to choose the worker, review the patch, and accept verified results in order.
The ZIP includes the skill, references, Python adapters, and offline tests. Scripts target macOS or Linux and require Python 3; this release was tested on macOS with Python 3.14. Model accounts, credentials, and a Jev gateway are configured separately.
One vertical: from goal to accepted patch
Start with one task. Add parallel workers only when the work is independent and review capacity supports them.
1 · Coordinator
Define the task
Set the goal, permitted writes, dependencies, and evidence needed to accept each result.
2 · RepoToire + source inspection
Gather evidence
Supply structural context and impact. The coordinator checks freshness, coverage, and missing dependencies.
3 · Jev advice → coordinator → selected models
Choose and draft
Prefer available local models for implementation and routine work. Choose among admitted plans using observed usage and shared capacity. Workers propose changes for coordinator approval before implementation.
4 · Independent reviewer + coordinator
Review and correct
Review a fixed patch, verify concrete findings, and send bounded corrections to an available worker.
5 · Integration owner
Accept in order
Check each result on the cumulative accepted revision. Earlier failures hold later results provisional.
RepoToire supplies evidence. The coordinator checks it.
RepoToire can help identify owners, affected symbols, and verification scope. Its output carries a source revision, coverage, trust, and caveats. Source inspection and behavior checks still establish the conclusion.
Transfer admission or execution responsibilities only after the installed runtime path is supported, qualified, and authorized. Use its existing owner instead of creating a competing scheduler. This skill does not establish concurrent Machine execution.
Local work first. Balance the remaining work.
You select models and roles. The coordinator checks availability, context, workload, and shared quota, preferring local models for implementation and routine work. One valid plan is selected locally; with multiple valid plans, Jev chooses using the supplied usage evidence. Model-quality ranking is optional.
Reviewers use a fresh context and a different model from every patch author. Unknown quota stays pending. If an authorized task exhausts its allowance, preserve unfinished work, stop retries, and use another model only within existing authorization and after fresh capacity checks. The coordinator owns this recovery procedure.
Design reference: TypeSafe’s System One guide. The scheduling policy is this skill’s design; it is not a provider guarantee.
Try it on one authorized task
- Download and extract the ZIP. Review the complete
organized-chaosfolder, including its references and scripts. - Preserve any existing installation, then place the folder in your agent’s supported skills directory.
- Follow the onboarding instructions to create a private inventory, select your models, and check access. The local editor opens an existing inventory.
- Ask: “Use organized-chaos for this task. First show its scope, dependencies, model availability, and acceptance checks.”
Installing the package grants no tool or provider permissions. Jev adapters are included; gateway hosting and an account are not. No model provider is contacted by this page or download.
What the package establishes
- Includes a coordination method, adapters, and offline tests. Passing those tests does not establish your provider access or readiness.
- Keeps drafts provisional until review and checks match the cumulative accepted revision. A clean model review is not correctness proof.
- One local comparison did not establish a throughput or cost benefit. Measure complete comparable runs, including setup, routing, retries, review, and integration, before claiming savings.
- Separates batch acceptance from commits, merges, deployment, and Machine retirement. Each needs its own authority and evidence.
Review the package instructions
The entry instructions below come from the v2 source. The ZIP contains all 40 files, including references, adapters, and tests.
Read public v2 instructions
--- name: organized-chaos description: Coordinates delivery when users ask to split a large task across agents or models, run independent issues in parallel, or land dependent changes in order. Includes model onboarding, bounded delegation, independent review, and ordered acceptance. --- # Organized Chaos Turn an authorized goal into independently verifiable tasks. Use the user's selected models and access methods. The lead owns scope and acceptance. Jev chooses a staffing plan from options admitted by the coordinator, using availability and usage. The coordinator owns validation and dispatch. 1. **Select available models.** On first use, follow [onboarding](references/onboarding.md). Keep model choices, role preferences, billing policy, and routing policy in user configuration outside this package. Read `~/.config/organized-chaos/routing.md` when present; preserve it during updates. Refresh access and shared-usage evidence before staffing. Authentication, model availability, remaining quota, and task suitability are separate facts. Unsupported or unknown access stays pending. 2. **Define tasks before staffing.** Record outcomes, owned paths, acceptance checks, dependencies, and retirement order. Use the table below. Record required tools, context size, and execution constraints. Use available Symbol FS evidence to bound the code context; it does not establish model quality. Independent tasks may overlap when custody and capacity permit; dependent tasks start from accepted prerequisites. Verify consequential repository claims against source. 3. **Assign bounded work.** Follow [routing](references/routing.md). Default to local-first implementation and routine work, then balance overflow and reviews across the user's selected models. Check access, context capacity, current workload, and shared quota. Give the adapter capacity-valid plans; it selects a sole plan locally and asks Jev to choose among multiple plans based on usage balance. Do not require model-quality proof, qualification criteria, or a suitability confidence threshold. Exclude the patch's author model from its review. Jev suitability experiments remain separate opt-in work. Give each writer an outcome, owned paths, limits, and required evidence. Every model worker first proposes code changes without editing; the coordinator approves the proposal and scope before implementation. Follow the [proposal and approval procedure](references/delivery.md#proposal-before-implementation). For an already reviewed exact source replacement, the coordinator may assign the [single-file gate](references/worker-gate.md) to apply approved bytes deterministically; this executor does not dispatch a model. 4. **Verify, review, and accept in order.** Follow [delivery](references/delivery.md). Run behavior checks, freeze the patch, and obtain independent review. Validate findings against source. A completed worker result remains provisional until its acceptance checks pass on the cumulative accepted state. Retire only the next eligible task; a failed prerequisite blocks dependent acceptance. 5. **Report evidence and stop.** Use the receipt below. Separate edited, tested, installed, merged, and deployed states. Report unavailable checks and unresolved findings. Claim cost savings only from comparable measured runs, including coordination and failures. Skill updates require user authorization. | Order / ID | Outcome and acceptance check | Prerequisites | Owned paths | Model / budget | State | | --- | --- | --- | --- | --- | --- | | 1 / A | Define contract; boundary checks pass | None | Contract files | Configured writer | Ready | | 2 / B | Consume contract; journey passes | A accepted | Consumer files | Configured writer | Waiting for A | | 3 / C | Independent documentation verified | None | Documentation | Configured writer | Ready | ```text Decision: accepted | rejected | needs-evidence Accepted tasks and exact revision: Verification and independent review: Unresolved findings, owner, next action: Files changed and installation/delivery state: ```
Read formatted instructions
Organized Chaos
Turn an authorized goal into independently verifiable tasks. Use the user's selected models and access methods. The lead owns scope and acceptance. Jev chooses a staffing plan from options admitted by the coordinator, using availability and usage. The coordinator owns validation and dispatch.
- Select available models. On first use, follow onboarding. Keep model choices, role preferences, billing policy, and routing policy in user configuration outside this package. Read
~/.config/organized-chaos/routing.mdwhen present; preserve it during updates. Refresh access and shared-usage evidence before staffing. Authentication, model availability, remaining quota, and task suitability are separate facts. Unsupported or unknown access stays pending. - Define tasks before staffing. Record outcomes, owned paths, acceptance checks, dependencies, and retirement order. Use the table below. Record required tools, context size, and execution constraints. Use available Symbol FS evidence to bound the code context; it does not establish model quality. Independent tasks may overlap when custody and capacity permit; dependent tasks start from accepted prerequisites. Verify consequential repository claims against source.
- Assign bounded work. Follow routing. Default to local-first implementation and routine work, then balance overflow and reviews across the user's selected models. Check access, context capacity, current workload, and shared quota. Give the adapter capacity-valid plans; it selects a sole plan locally and asks Jev to choose among multiple plans based on usage balance. Do not require model-quality proof, qualification criteria, or a suitability confidence threshold. Exclude the patch's author model from its review. Jev suitability experiments remain separate opt-in work. Give each writer an outcome, owned paths, limits, and required evidence. Every model worker first proposes code changes without editing; the coordinator approves the proposal and scope before implementation. Follow the proposal and approval procedure. For an already reviewed exact source replacement, the coordinator may assign the single-file gate to apply approved bytes deterministically; this executor does not dispatch a model.
- Verify, review, and accept in order. Follow delivery. Run behavior checks, freeze the patch, and obtain independent review. Validate findings against source. A completed worker result remains provisional until its acceptance checks pass on the cumulative accepted state. Retire only the next eligible task; a failed prerequisite blocks dependent acceptance.
- Report evidence and stop. Use the receipt below. Separate edited, tested, installed, merged, and deployed states. Report unavailable checks and unresolved findings. Claim cost savings only from comparable measured runs, including coordination and failures. Skill updates require user authorization.
| Order / ID | Outcome and acceptance check | Prerequisites | Owned paths | Model / budget | State |
|---|---|---|---|---|---|
| 1 / A | Define contract; boundary checks pass | None | Contract files | Configured writer | Ready |
| 2 / B | Consume contract; journey passes | A accepted | Consumer files | Configured writer | Waiting for A |
| 3 / C | Independent documentation verified | None | Documentation | Configured writer | Ready |
Decision: accepted | rejected | needs-evidence
Accepted tasks and exact revision:
Verification and independent review:
Unresolved findings, owner, next action:
Files changed and installation/delivery state:
Package ZIP SHA-256
077e63128bece3c6b66a67777ccb104ef66d68138b123c8e638233cea751e051