TL;DR: faf_bench proves FAF's grounding lift in-session — a cold-vs-.faf benchmark with a mechanical ✪ receipt — and faf_go now bootstraps from a cold repo (init → auto → 6Ws).
What's New
faf_bench — proof, not pitch
Everyone claims context helps the model. faf_bench makes you show it, in the same session, on your own repo. It's the new lead tool of the Core tier.
Here's the loop. It derives a question set from your .faf, asks the model those questions cold — no context — then asks the same questions with the .faf loaded. The grading is mechanical: there's no LLM judge, because the .faf is the answer key. The output is a ✪ receipt showing the grounding delta between the two passes.
- Cold pass — the model answers blind, with no project context.
- Grounded pass — the same questions, with the
.fafin hand. - Mechanical grade — no judge model, no vibes. The
.fafis the answer key. ✪receipt — the grounding delta, written down. Proof you can run.
/faf-bench — the honest two-pass prompt
Paired with the tool is the /faf-bench session protocol: cold-blind first, never inventing token counts, and never presenting the cold pass alone as a verdict. The honesty is the point — a benchmark that flatters the result isn't proof.
faf_go now bootstraps a cold repo
If there's no project.faf yet, faf_go no longer asks you to go set one up first. It runs faf_init → faf_auto to lay the foundation, then walks the 6Ws. One call takes a bare repo to a real score.
Try It
npm install -g claude-faf-mcp@5.12.0 Then, in a Claude session on any repo, run faf_bench — or the /faf-bench prompt for the full honest two-pass protocol. The ✪ receipt comes back in-session. Zero-Config — one-click .mcpb, no terminal, no JSON config.
The Numbers
- v5.12.0 — Released June 22, 2026
- 36 tools — 13 Core (advertised by default), the rest Extended via
FAF_TOOLS=all - 569/569 — Tests passing, 0 fail
- faf_bench — the new lead of the Core tier