Unicorn orchestrator

Skill andrey-learning-machines/swe-harness/plugins/swe-harness/skills/unicorn-orchestrator

Portable SWE harness plugin for Codex and Claude Code

Install

npx -y skills add andrey-learning-machines/swe-harness --skill unicorn-orchestrator

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Codex-native orchestration workflow for the Unicorn Team custom agents. Trigger when the user explicitly asks for delegation, subagents, or parallel agent work and the task benefits from specialized architecture, implementation, review, DevOps, language-ramp-up, or UI/UX help. Use for multi-step coding work, architecture planning, review pipelines, deployment prep, or mixed UI/backend efforts where bounded child-agent tasks materially help.

SKILL.md

5.3 KB, as published. Nobody here has run it

Unicorn Orchestrator

Use this skill only when the user explicitly wants delegation, parallel agent work, or a subagent-driven workflow. In ordinary turns, keep work local.

This skill targets Codex custom agents installed in ~/.codex/agents:

Agent	Name	Codex model	Use for
Architect	`unicorn_architect`	`gpt-5.5`	design packages, ADRs, API contracts, tradeoff analysis
Developer	`unicorn_developer`	`gpt-5.5`	implementation, bug fixes, TDD, refactors
QA/Security	`unicorn_qa_security`	`gpt-5.5`	review, security findings, regression risk, missing tests
DevOps	`unicorn_devops`	`gpt-5.4`	CI/CD, IaC, deployment, observability
Polyglot	`unicorn_polyglot`	`gpt-5.4-mini`	new language or framework ramp-up
UI Designer	`ui_ux_designer`	`gpt-5.4`	design systems, art direction, UX direction
UI Builder	`ui_frontend_builder`	`gpt-5.3-codex`	frontend implementation
UX Auditor	`ux_auditor`	`gpt-5.4-mini`	accessibility, interaction, responsive, perf review

Codex agents must use OpenAI model identifiers in their Tom's Obvious Minimal Language (TOML) configuration. Claude Code agents use a different Markdown agent format and may use Claude model aliases. See docs/model-routing.md for the current model-routing rationale and fallback policy.

Core Rules

Form a local plan first.
Do not delegate the immediate blocking task if the next local step depends on it.
Delegate only bounded tasks with a concrete output.
Give each writable agent a clear ownership boundary.
Require evidence back: files changed, tests run, findings, or explicit rationale.
Wait sparingly; keep doing non-overlapping work while child agents run.
For medium or large feature work, establish durable intent first with Specification Kit (Spec Kit), Gherkin scenarios, and acceptance evidence before delegating implementation.

Default Routing

Situation	Pattern
Simple question or one-shot fix	Handle locally
Architecture or multi-service change	`unicorn_architect` first, then `unicorn_developer`
Bug fix	`unicorn_developer`; add `unicorn_qa_security` if risk is high
Code review	`unicorn_qa_security`
Deploy / CI / infra	`unicorn_devops`
New language / unfamiliar framework	`unicorn_polyglot`, then implementation agent
Visual design / landing page / product UI	`ui_ux_designer`, then `ui_frontend_builder`, then `ux_auditor`
Feature files to Linear stories	Use `linear-gherkin-stories` locally for review-first story drafts, then write through Linear MCP only after approval
Parallel independent workstreams	Spawn multiple agents with disjoint scopes

Delegation Patterns

Feature

For medium or large features, route through spec-acceptance-harness first so the request has a Spec Kit feature specification, Gherkin scenarios, acceptance tests, and story matrix expectations.
If design risk is non-trivial, ask unicorn_architect for a short design package.
Hand implementation to unicorn_developer with:
- exact goal
- relevant files
- acceptance criteria
- explicit request to write or update tests first when practical
- required validation commands
If the change is security-sensitive, externally facing, or broad, run unicorn_qa_security on the result.

Linear Stories

Use linear-gherkin-stories locally when the user asks to turn scenarios into backlog items. Generate story drafts, review them with the user, and then create or update Linear issues only after explicit approval.

Bug Fix

Hand unicorn_developer a reproduction target and require:

failing test or clear reproduction
root cause, not symptom patch
regression checks
concise summary of changed files

Review

Hand unicorn_qa_security:

diff or files to inspect
areas of concern
request for findings first, with file references and severity

Frontend

ui_ux_designer produces the design system or implementation direction.
ui_frontend_builder implements with explicit file ownership.
ux_auditor reviews accessibility, responsiveness, and motion/perf issues.

Quality Gates

Before accepting delegated work, check:

requested files or artifacts were actually produced
tests or validation commands ran when expected
risky assumptions are called out
no scope drift into unrelated files
review findings are addressed or deliberately deferred

Prompting Rules For Child Agents

State ownership clearly for writable work.
Tell the agent it is not alone in the codebase and must not revert unrelated edits.
Ask for the smallest useful final payload: summary, files changed, test results, findings, open questions.
For read-only work, require citations to files, symbols, or commands.