refactor: make Delegate Kit a universal orchestration policy

[tomastaker]

Summary

This rewrites Delegate Kit from a Claude/Codex-specific execution package into a compact, host-neutral orchestration policy for capable coordinator models.

The current session remains the coordinator. Delegate Kit now teaches it to choose the smallest useful execution shape:

• DIRECT for localized or tightly coupled work;
• SCOUT for one bounded read-only investigation;
• SINGLE for one coherent delegated outcome;
• PARALLEL for 2–3 genuinely independent outcomes;
• SEQUENTIAL when dependencies or shared state forbid parallelism;
• independent review for non-trivial or risky changes.

Design decisions

• Prefer the current host's native subagent, worker, worktree, messaging, and lifecycle capabilities.
• Do not hard-code Claude, Codex, model names, providers, subscriptions, or cross-vendor routing.
• Keep planning, integration, verification, and the user conversation with the root coordinator.
• Derive worker count from independent outcomes rather than repository size or apparent task size.
• Default to at most 3 active workers, 2 writers, 1 scout, and 1 reviewer.
• Require self-contained worker briefs, explicit write ownership, bounded nesting, and evidence-backed returns.
• Treat a different model/provider as an optional independence upgrade rather than a universal requirement.

Removed

The active skill package no longer ships:

• external claude -p / codex exec orchestration;
• routing presets and provider-specific model tables;
• custom run ledger, quota fallback, delivery, and resume runtime;
• installed role definitions and Codex agent configuration;
• custom worktree CLI and shell safety hooks;
• panel/lead review machinery and the large JSON result schema.

Those mechanisms remain available in Git history. Modern harnesses already provide much of this runtime functionality, while keeping it in the core skill made the policy heavier and less portable.

Documentation

The README now explains the decision model, scaling behavior, coordinator ownership, worker briefs, independent review, optional model tiers, installation, and examples for small repositories and large monorepos.

The design borrows selected principles from Superpowers, Orca, and Hyperskills without adopting large default swarms or a mandatory development ceremony.

Validation

CI now validates that:

• the compact SKILL.md exists and has required frontmatter;
• the core decision concepts are present;
• the routinely loaded skill remains under 12 KB;
• provider-specific runtime terms do not leak back into the core policy.

Expected outcome

Delegate Kit becomes suitable for routine, effectively always-on use across Claude Code, Codex, T3 Code, Orca, Pi, OpenCode, and future harnesses. Small tasks should remain single-agent; large repositories should trigger targeted scouting; parallel workers should appear only where outcomes and mutable state are genuinely independent.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 96ba1d62-db2e-4e50-a54d-5186ee8f2904

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-08-29T22:03:32.381833Z	ee4b1db	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ee4b1dbf32

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Quote the description so the frontmatter parses

When a skill host parses this frontmatter as YAML, the embedded shape: work contains a colon followed by a space inside an unquoted scalar, causing parsing to abort with mapping values are not allowed in this context; consequently, Delegate Kit cannot be discovered or loaded by standards-compliant skill loaders. Quote or fold the description, and consider having CI parse the frontmatter as YAML rather than checking substrings.

Useful? React with 👍 / 👎.

[tomastaker]

Superseded by the policy + external-runtime rework (new PR from rework/policy-plus-external): keeps the shape vocabulary from this PR, moves the runtime behind references instead of removing it, and makes cross-family review a preference with a same-family fallback.