fusion-tea: Skill for Claude Code

Before executing this skill, read .agentic-mbse/claude.md in Claude Code or .agentic-mbse/codex.md in Codex. Resolve supporting paths from this skill’s installed directory; keep generated outputs in the project or a temporary directory. Read referenced skills from .agents/skills/<name>/SKILL.md when their guidance is needed.

Research Acquire

Purpose: ACQUIRE sources — answer one bounded research request by finding sources, bringing them into the repository with proof they are there, and returning exactly one of four classes. Input: A request file, knowledge/research/requests/<request-id>.json Output: return.json in the run directory, plus whatever was registered

Research has two stages with a wall between them. This skill is acquisition: it proves a source is in the repository. Reading is /research: it reasons only from sources already in, and it is where domain insights (DI-XXX) are made.

You decide three things: the search queries, keeper or rejected for each candidate, and three sentences about each keeper. The agentic-mbse research CLI does everything else. It keeps the run's record, captures and checks each source, writes the registry, and computes the return class from what landed on disk.

The operator guide covers forming a request, reading the four classes and acting on what comes back. It is research-seam.md in the installed toolkit's docs directory (uv run python -c 'from agentic_mbse.cli import get_docs_dir; print(get_docs_dir())').

When invoked without a request file, ask which request to run and wait.

Skills Referenced

• source-traceability: how a registered source's location is cited.

Standing Rules

These hold for the whole invocation.

1. WebFetch is for triage only. Its output is a lossy summary. It tells you whether a page is reachable and whether it looks relevant. Never quote it, cite it or paste it into a research document, and never register it. The only thing that becomes a source is what agentic-mbse research register captured.
2. You do not mint domain insights. Registering a source is not approving an insight. Insights are made only through /research and the owner's approval. If this round found something that should become one, say so in your report and stop there.
3. You do not write registry files. Not the source index, not the manifest, nothing under knowledge/sources/. If you find yourself wanting to, the CLI refused something, and the right move is to report it, not to route around it.
4. A hold-out match is never yours to waive. register returns holdout_hit with a rule id, and the candidate is queued for the operator. There is no flag that admits it, and asking for one is out of scope. The hook sees only what reaches register, so apply any quarantine rule the project's instructions state before you search or fetch.