Harness support

super-backlog v1 supports two agent harnesses (design decision D1): OpenCode and Claude Code. sbl init targets both by default; use --harness opencode or --harness claude to restrict.

OpenCode — native

• opencode.json gets a merged plugin entry: "plugin": ["superpowers@git+https://github.com/obra/superpowers.git"]. Existing keys and other plugin entries are untouched; the entry is only added if the exact kit spec string is absent.
• The glue skills (spec-to-backlog, backlog-status-report, task-review-gate, architecture-summary) are installed as files at .opencode/skill/<skill>/SKILL.md.

Claude Code — file-based skills + marketplace

• The same glue skills are installed file-based at .claude/skills/<skill>/SKILL.md — this always works, with or without CLI scripting support.
• A short pointer section is written to CLAUDE.md. It references the managed block in AGENTS.md and its delegation rule. sbl init and sbl update refresh this section whenever its text differs from the current kit version. Put your own CLAUDE.md content under its own heading, because text inside the pointer section is replaced.
• Marketplace installation: init does not run claude commands. After writing skills, it prints the exact command to paste inside Claude Code: /plugin install superpowers@claude-plugins-official. Because this step must be run manually, init pushes a warning (claude plugin install must be run manually) and finishes as success-with-warnings (exit code 4). File-based skills work immediately either way.

Model tiers and delegation (all harnesses)

The managed AGENTS.md block carries a harness-neutral "Model tiers and delegation" section. The agent does the work itself by default; when it delegates, the subtask goes to the cheapest model that solves it reliably, picked from four tiers: Light (mechanical: searching, listing, extracting), Standard (clearly specified routine work: research, routine code, tests, docs), Strong (complex implementation, debugging, reviews) and Top (analysis, design, architecture, decisions where a mistake is costly or hard to reverse). When unsure, the agent takes the higher tier. The section also says when delegation pays off, how to brief a subagent (goal, context, limits, done criterion, return format), that returns stay brief, and that a result that falls short goes one tier up instead of being retried. Short examples name the per-harness setting: for Claude Code the Agent tool's model parameter (haiku, sonnet, opus, strongest available), for OpenCode the agent's model field. With the router enabled, the examples name the configured models instead.

The rule applies regardless of the session model and is always on. Projects that want different behavior add their own rule below the block. The opt-in model router (sbl models) writes one subagent per tier with a fixed model and renders the block's examples from its configuration.

What is file-based vs. delegated

Artifact	Mechanism	Managed how
.opencode/skill/spec-to-backlog/	file-based, written by init	fingerprint header line
.claude/skills/spec-to-backlog/	file-based, written by init	fingerprint header line
opencode.json plugin entry	file-based merge into your config	byte-equality of entry
Superpowers runtime for OpenCode	delegated to OpenCode's plugin loader via the git-backed spec string	canonical spec string only
Superpowers marketplace install for Claude Code	instructed (manual one-time step)	init prints the exact command; never executed automatically
AGENTS.md workflow block	file-based injection	start/end markers; written with the file's line endings
CLAUDE.md pointer	file-based append, refreshed in place	recognized heading; the section ends at the next heading; written with the file's line endings

The manifest is designed so later harnesses (Cursor, Codex, Gemini CLI — v2 backlog) can be added without changing the ownership model.