brain-stats — L0 accounting layer

Framework context

This is an L0 skill in brain-mcp's SHELET stratification. L0 operations report on raw state without interpretation. Outputs are structural facts (counts, dates, sources) — not synthesized claims. No citations required because the numbers are the citation.

When to invoke

• User asks "how much is in my brain" / "how many conversations" / "what's the corpus size"
• Onboarding — show what was successfully ingested
• Before any L2/L3 synthesis call, to decide if there's enough data to synthesize

Input

view: "overview" | "domains" | "pulse" | "conversations" | "embeddings" | "github" | "markdown"

Default: overview.

Output contract

Markdown report with quantitative facts only. Every number is exact (no "approximately"). Structure:

• Total message count
• Total conversation count
• Embedding count (if L1 pipeline run)
• Summary count (if L2 pipeline run)
• Per-source breakdown
• Date range (min/max msg_timestamp)
• Top 10 domains (for domains view)
• Thinking-stage matrix (for pulse view)

Does NOT do

• Interpret what the numbers mean
• Recommend actions based on numbers
• Hide zero-count layers (if no embeddings, say "0" — don't skip the row)
• Fabricate counts if a data layer is missing (return "unavailable")

Execution

Internally wraps brain_mcp.server.tools_stats.brain_stats(view). The MCP tool is the implementation; this skill is the governance contract.

Verification checklist

• [ ] Every count is an exact integer (no commas in the internal value, only in display)
• [ ] Missing layers report "unavailable" rather than crashing
• [ ] No prose claims beyond structural facts
• [ ] Output is idempotent — same corpus → same numbers