Processed Beef is a portable, skill-only delegation process for high-quality
software delivery with low parent context and progressively cheaper subagents.
Its strict project-wide design boundaries are docs/principles.md,
owned by the user and constraining plans, designs, decisions, and work.
Maintainer process records live outside installed skills: AGENTS.md
is the maintainer authority. The canonical records are docs/principles.md
for strict, user-owned project-wide design boundaries, docs/backlog.md is the Orchestrator-owned
concise prioritized list with one linked priority/dependency line per pending or
active meaningful change, and docs/decisions.md holds concise active
decisions. Supporting records include docs/changes/ for active
meaningful-change notes and docs/learnings.md for separate
historical evidence. A separate product VISION.md, when present, is goal input only, not
Processed Beef governance.
| Role | Skill | Responsibility |
|---|---|---|
| Orchestrator | processed-beef |
Long-lived frontier-model product and engineering partner; keeps priorities and material decisions aligned with the user and delegates coherent changes |
| Lead | processed-beef-orchestrate |
Cheaper change owner; scopes work, delegates bounded units, inspects actual changes and evidence, and reports outcomes |
| Worker | processed-beef-work-unit |
Cheapest capable execution tier; completes one bounded unit, verifies it, reports evidence, and never delegates |
The user has final authority over material decisions. Ordinary technical choices, implementation, and verification proceed without artificial approval gates.
- The Orchestrator turns the user's goal into an outcome, constraints, material decision boundaries, and evidence expectations.
- One Lead owns a coherent change, with a fresh Lead at each real ownership boundary. It delegates implementation, investigation, repetitive work, and output-heavy commands before loading expensive corpora.
- A fresh Worker executes each bounded unit with a terminal result and never maintains project records. The Lead accepts work only after inspecting the real diff, repository state, and current evidence.
- Independent review is added only for risk, subtle breadth, suspicious output, or repeated semantic failure.
- The Lead returns a lean terminal handover. The Orchestrator checks user alignment without repeating implementation review or tests.
Trivial requests use the request directly. Each meaningful change has one
Lead-owned docs/changes/<slug>.md combining compact specification and current
plan, plus an Orchestrator-prioritized backlog line. Update the note only for
meaningful changes. At completion, verify acceptance, promote authorized durable
decisions, retain concise outcome/evidence, archive the note, and advance or
remove its backlog line. No approval gate, log, state or recovery file, routine
progress record, retry ledger, tick, or completion transaction is required.
- One subagent at a time is the default subscription-quota control. The user may choose concurrency when work is isolated and recoverable.
- Parent context contains curated decisions and evidence, not raw implementation corpora, diffs, logs, or transcripts.
- A failed semantic approach is not retried without new evidence, a changed hypothesis, or a changed approach. Renaming agents or units creates no budget.
- Pre-effect command, environment, and record failures are repaired locally and do not consume semantic attempts.
- Destructive actions need explicit scope; irreversible external or non-versioned destruction needs user authority.
- Commits and pushes happen only when explicitly requested, and never with known acceptance gaps or unresolved serious findings.
Full operation uses nested subagents to depth two: Orchestrator to Lead to Worker. When a host cannot provide that topology, the process reports the cost or quality limitation and uses the shallowest safe alternative rather than blocking unrelated work.
Install all three skills from this canonical repository:
npx skills add beefsack/processed-beef --allgh skill install beefsack/processed-beef --all --agent <agent> --scope userOpenCode can register them directly as a plugin:
{
"plugin": [
"processed-beef@git+https://github.com/beefsack/processed-beef.git"
]
}The plugin registers skills only. It does not activate the workflow, create role agents, select models, or configure nesting and permissions.
Per-role model selection belongs in each host's agent definitions. Use a frontier model for the Orchestrator, a cheaper capable Lead, and the cheapest capable Worker. Skills express the topology but cannot portably prove which model the host selected.
Integration guides provide copyable role-agent configuration and host limits:
See architecture.md for topology, authority, context, and recovery design. framework-comparison.md retains the research precedents.
- The project is pre-release; role agent names are not stable compatibility interfaces.
- Per-role model routing and Worker delegation denial depend on host support.
- Hosts do not consistently expose context telemetry or hard per-agent limits; bounded work and deliberate handover remain process policy.