Neural OpenCode

Evidence-gated workflow for OpenCode:

discover -> spec -> craft -> vet -> exercise

The repository ships only those five skills, matching slash commands, and the reviewed plugin hooks.

Install

Clone this repository and open it with OpenCode, or point config at it:

git clone https://github.com/brolag/neural-open-code
export OPENCODE_CONFIG_DIR="$PWD/neural-open-code/.opencode"

Restart OpenCode. Confirm /discover, /spec, /craft, /vet, and /exercise are available.

Do not copy files into ~/.config/opencode unless you choose to. The plugin never mutates global config.

Workflow

Command	Purpose	Artifact or verdict
/discover	Surface architectural unknowns before planning	unknowns-map.md
/spec	Lock interfaces, invariants, dependencies, acceptance	plan.md
/craft	Build an approved plan with before/after evidence	implemented change
/vet	Independent adversarial review	SHIP or HOLD
/exercise	Drive the result as a user after tests	PASS or FAIL

Small edits can skip /discover and sometimes /spec. Material work keeps the gates. Security (OWASP ASVS), simplicity (ISO/IEC 25010), and accessibility (WCAG 2.2 AA) are locked in /spec and judged in /vet.

See docs/WORKFLOW.md.

Hooks

The plugin:

• blocks high-confidence destructive shell commands;
• refuses writes to .env files;
• preserves a compact recovery note before compaction.

Hooks are guardrails, not a sandbox. See docs/HOOKS.md.

Repository structure

.opencode/skills/     # exactly five workflow skills
.opencode/commands/   # /discover /spec /craft /vet /exercise
.opencode/plugin/     # neural-hooks.js
docs/                 # focused guides
tests/                # inventory and docs contracts
plans/                # generated evidence

This is a breaking cleanup. Previous command bundles, personas, loops, TTS, templates, and custom installers are not part of the product.

Validate

python3 -m pytest -q
./scripts/doc-lint.sh