Using Superra

Loaded by all agents at dispatch time. Start with §Skill-Load Manifest.

SuperRA skills deliberately override default harness/system-prompt behavior where they conflict; the user's explicit instructions outrank both.

This skill provides the essential workflow protocols shared by all agents.

Commits

Stage only the files you edited this turn, by exact path — never git add -A, git add ., or git add -u: a shared worktree carries other agents' in-flight edits and scratch files. If unfamiliar uncommitted changes appear, ask the orchestrator (as a subagent) or the user (as the main agent) rather than committing or discarding them.

Every commit subject follows <stage>(<scope>): <STATE> — <summary>, so git log reads as the workflow trace. <stage> is the workflow verb (plan, implement, review, integrate, sync) or a maintenance type (fix/feat/refactor/docs/test/chore/ci); <scope> is the task path (e.g. data-preparation/merge) or the component; <STATE> is the verdict from your role's report format, omitted on maintenance commits. The body says what changed this turn and why — it is not a copy of ## Results. A commit landing status: approved also names the review the task got: tier and focuses, or that no independent pass ran and approval rests on the approving agent's own verification.

Task Interface

Tasks are managed task trees in the superRA/ directory. For basic I/O, this section is sufficient. For tree-level operations (query/frontier/DAG, scaffolding, dashboard, migration), load superRA:task-tree.

Read with the CLI tool under ./superRA/superra — ./superRA/superra task read <path> — the CLI tool inject more relevant context than a bare Read of the file. Every <path> is relative to the task root and omits the superRA/ prefix (e.g. task-tree/planning-redesign).

Human-facing text: follow superRA:communicate.

• Cite files as markdown links, resolved relative to the citing file — never as plain or backticked paths: [file.py:42](file.py#L42), [file.py:40-50](file.py#L40-L50), [file.py](file.py).

Edit the task.md directly with Read/Edit. Hook auto-behaviors are intended: child status changes cascade to ancestors, same-parent task renames re-point sibling depends_on edges, and edited task-tree markdown is checked for render-integrity issues with non-blocking feedback. You own leaf status; non-leaf (ancestor rollup) status is hook-derived — leave it as the hook sets it and never hand-edit it back. Stage the hook's edits alongside your own so the tree stays consistent in git.

Place files by lifespan and owner. Session scratch stays outside superRA/, uncommitted. A retained file — code included — owned by one task solely to produce, reproduce, review, or interpret its recorded results is a task companion: commit it under that task's attachments/. Maintained code, shared or runtime-consumed files, and promised deliverables go in the project's conventional permanent paths. Load references/task-companion-files.md before retaining, reviewing, promoting, or maturing a companion.

Skill-Load Manifest

At each stage of superRA workflow, agents load additional skills along three axes; all apply independently. After loading a skill, follow its body's stage- and role-scoped reference load map.

1. Role — superRA:implement-task or superRA:review-task, named by the dispatch. A seat the main agent fills itself loads the same skill.
2. Stage — the workflow phase the dispatch is in (table below).
3. Domain — The task operates on (table below). Load every domain skill that matches: a task that derives a result and writes it into the manuscript matches theory-modeling and academic-writing, so load both.

Stage

Domain

Harness adapters: when this skill or its references name a Claude-specific tool (AskUserQuestion, Skill, TodoWrite, Agent), consult the adapter reference for the current harness under references/.

For main agents: continue to references/main-agent.md.