Chromatic #
Chromatic is an installable coding-agent skill that combines reviewed implementation planning, dual-loop BDD, and fresh-context adversarial code review.
Repository: https://tangled.org/hyl.st/chromatic
PLAN → REVIEW → integration RED → unit RED/GREEN/REFACTOR → integration GREEN → BLUE
It records intended outcomes before coding and actual outcomes after verification, so the plan remains a useful decision record instead of becoming stale process debris.
Install #
Clone this repository into the skill directory for your coding agent:
# OpenCode
git clone https://tangled.org/hyl.st/chromatic ~/.config/opencode/skills/chromatic
# Claude Code
git clone https://tangled.org/hyl.st/chromatic ~/.claude/skills/chromatic
# Codex CLI
git clone https://tangled.org/hyl.st/chromatic ~/.codex/skills/chromatic
Restart the coding agent after installation or update. To update:
git -C ~/.config/opencode/skills/chromatic pull --ff-only
Use #
Request Chromatic by name with a concrete task or issue:
Use Chromatic to implement issue #42.
Chromatic reads repository policy, writes and independently reviews a plan, delivers one observable behavior at a time through nested BDD/TDD loops, runs independent BLUE reviews, and records the verified outcome in the plan.
Repository profiles #
Optional files under profiles/ adapt issue selection, worktrees, verification lanes, and handoff rules for a repository without forking the core workflow. Profiles may tighten gates but may not weaken dual-loop BDD or adversarial review.
The initial automerge-atproto profile preserves its persistent epic worktree, Chainlink lifecycle, canonical protocol contracts, and required conformance lanes.
Development #
The root SKILL.md is the installed skill. Keep it host-agnostic and concise. Put repository-specific policy in profiles/ and verify that every profile referenced by the skill is committed.
The Tangled push remote is git@knot.commonscomputer.com:did:plc:jx5khyyfjknpff236v6dotmy.
When changing the workflow, check that the skill still requires:
- reviewed, persisted implementation outcomes before production edits;
- one failing outer test per behavior increment;
- evidence-bearing inner red-green-refactor cycles;
- fresh plan and implementation reviewers;
- complete verification and an updated outcome record.