Witch

A fixed-lane delegation harness for OpenCode. Route routine volume to Luna, integration to Terra, and consequential judgment to Sol.

Apache-2.0 · Install · Architecture · Provenance

Witch gives OpenCode a deterministic multi-agent workflow instead of leaving model choice, permissions, and delegation shape implicit. One Sol orchestrator routes work into seven named subagents with fixed models and permissions, then owns verification and final acceptance.

witch (Sol / medium)
├── routine work       → witch-routine       (Luna / high)
├── complex work       → witch-complex       (Terra / max)
├── research           → witch-researcher    (Luna / max, read-only)
├── security-sensitive → witch-security      (Sol / high, restricted)
├── task review        → witch-task-reviewer (Luna / max, read-only)
├── repair review      → witch-re-reviewer   (Luna / high, read-only)
└── final acceptance   → witch-final-reviewer (Sol / high, read-only)

There are no escalation profiles, generic-agent fallbacks, dispatch-time model overrides, or Sol/max lane. A failed task is repaired inside its declared lane or stopped for human direction.

• Predictable cost — inexpensive agents handle bounded volume while Sol is reserved for orchestration, security, and final judgment.
• Explicit permissions — implementation agents receive bounded edit access; research and review agents cannot edit, delegate, run shell commands, or access external directories.
• Deterministic routing — security, research, complex integration, and routine work have an enforced precedence.
• Evidence before acceptance — task review, bounded repair/re-review, and a fresh final review are explicit workflow stages.

git clone https://github.com/muratmirgun/witch.git
cd witch
npm install
npm run verify

Start OpenCode from the repository root for project-local loading. To install the same agents, skills, plugin, and model variants into the global OpenCode configuration, run:

npm run install:opencode

The installer creates a timestamped backup before replacing Witch-owned files and preserves unrelated user configuration. See .opencode/INSTALL.md for model preflight, global installation, git-backed loading, and troubleshooting.

OpenAI's standard API list prices changed on 2026-07-30:

For an identical input/output token shape, Luna is approximately 10x cheaper than Terra and 25x cheaper than Sol. That ratio explains Witch's lane choices; it is not a claim that every Witch task costs 10x or 25x less. Real workflow cost depends on lane mix, token volume, prompt caching, reasoning effort, tool loops, and whether a task needs repair.

Source: OpenAI, “Advancing the price-performance frontier with GPT-5.6”.

Before implementation or research, the controller calls witch_route with worktree-mutation, security, and integration signals plus a non-empty reason. The first matching rule wins: security, research, complex integration, then routine implementation.

The detailed route contract, review cycle, permission boundaries, and file layout are documented in docs/architecture.md.

npm test       # routing, permissions, installer, plugin, and provenance
npm run verify # tests plus TypeScript plugin syntax validation

Witch is an independent OpenCode-only implementation derived from selected workflow material in obra/superpowers. It vendors its selected source locally and never fetches or installs the upstream package at runtime.

The selected upstream skills are vendored and adapted under the MIT license; these four prompt templates remain byte-for-byte preserved:

• skills/subagent-driven-development/implementer-prompt.md
• skills/subagent-driven-development/task-reviewer-prompt.md
• skills/subagent-driven-development/re-review-prompt.md
• skills/requesting-code-review/code-reviewer.md

Witch is released under the Apache License 2.0. The pinned upstream revision, hashes, and attribution boundary are recorded in THIRD_PARTY_NOTICES.md and vendor/upstream.json.