knowledge-mcp: Skill for Claude Code

Propose: $ARGUMENTS

For universal orchestration discipline (background spawning, user touch points), reference /orchestrate. This skill is proposal-discovery-specific.

The discriminator for every candidate: "does this create a capability that doesn't exist?" Yes → it belongs here. Makes an existing capability faster, safer, or better-tested → it is not a proposal; exclude it (note it in one line so the signal isn't lost).

Output feeds /brainstorm. Proposals are INPUTS to a future brainstorm, never work items: this skill creates NO tickets and NO projects.

Step 0: Resolve Scope + Index Freshness

Scope resolution (in order):

1. $ARGUMENTS names a repo or subsystem → that's the focus.
2. The session has already been working in a specific repo → that repo.
3. Otherwise → the repo of the current working directory.

Freshness: run manage({ "operation": "status" }). If the focus repo's index is behind HEAD, run collect({ "type": "code", "id": "<absolute-repo-path>" }) before spawning lenses — this skill is mostly non-interactive, so reindex without asking. Stale evidence produces stale proposals.

Step 0.5: Build the Dedup Set (inline, before spawning)

Proposals that collide with existing work are noise. Assemble three lists:

thoughts({ "operation": "recall", "mode": "context", "query": "<focus repo/product area>" })
query({ "type": "ticket", "limit": 50 })
query({ "type": "project", "limit": 30 })
thoughts({ "operation": "recall", "session": "propose-rejected" })

• Already planned — open tickets/projects covering a candidate → drop it.
• Already rejected — the propose-rejected session lists proposals the user dismissed in past runs → don't re-propose without new evidence.
• Already explored — prior decisions/findings that settled an area → candidates there need to engage that history, not ignore it.