knowledge-mcp: Skill for Claude Code

Research: $ARGUMENTS

For pipeline discipline reference /orchestrate. This skill is research-specific.

Step 0: Recall and freshness

thoughts({"operation":"recall","query":"<topic>"}) and a knowledge search first; what the graph holds goes into the brief as leads to reproduce, never as facts. manage({"operation":"status"}); collect if the index is behind.

Step 1: Decide the mode

• $ARGUMENTS names a ticket → VALIDATION: the researcher reproduces every premise, corrects the facts, fills the detail, stamps metadata.validated or reports what blocks it.
• Otherwise → DISCOVERY: what exists, how it works, the existing idiom, the call graph.

Step 2: Spawn the researcher (background)

The brief is discovery-framed: the problem or the ticket, never a candidate mechanism or a solution. Per the write-a-brief rulebook it carries verbatim the user's load-bearing rules, the tree to work at, absolute paths, and the isolation rule (probes in scratch; the operator's services untouched).

Agent(subagent_type: "researcher",
      prompt: "VALIDATE ticket <id> per your validation mode: reproduce every premise on <tree>, correct the facts on the ticket, fill what the draft did not know, stamp metadata.validated or report what stays unverified and what would change intent. Recall and search before every claim. Deliver the report to main.",
      description: "Validate: <ticket>",
      run_in_background: true)

Step 3: On return

Read the whole report. In validation mode: confirm the stamp landed by fetching the ticket; route intent changes to the user; a ticket left in draft goes back to /brainstorm with the researcher's evidence. In discovery mode: present the existing idiom, what exists, the call graph, what was decided and what is unclear, with the provenance labels intact. Findings that confirm or contradict a recorded thought charge it, citing the research node.