optimize

cclens optimize — act on the findings in this session

cclens optimize normally launches a new claude session seeded with an advisor prompt. This skill runs the same analysis but adopts that prompt in the current session instead, so the user stays where they are.

1. Resolve the binary and the store

• Binary: use cclens if command -v cclens finds it. Otherwise run every command below through Nix instead: nix run github:lambdalisue/cclens -- <subcommand …>. If neither cclens nor nix is available, stop and tell the user how to install it (nix profile install github:lambdalisue/cclens, or cargo install --git https://github.com/lambdalisue/cclens).

• Store: if ./cclens.db exists, use it and omit --db. Otherwise use the per-user store so no file is dropped into the current project:

  DB="${XDG_CACHE_HOME:-$HOME/.cache}/cclens/cclens.db"
  mkdir -p "$(dirname "$DB")"
  

  and pass --db "$DB" to every command below.

2. Analyze and fetch the advisor prompt

cclens analyze --db "$DB"
PROMPT="$(mktemp)"
cclens optimize --frozen --print --db "$DB" > "$PROMPT"

If the user asked to optimize their global setup or one specific project, add --scope global or --scope project:<slug> to the optimize command — the prompt then pins the work to that config layer.

--print writes the full advisor prompt — the prescribed instructions plus the complete findings briefing — to stdout. It goes through a temp file (mktemp creates it private, 0600), not your tool-output stream, because the briefing carries real paths and error excerpts from the user's transcripts.

3. Follow it

Read the temp file and adopt its contents as your instructions for the rest of the task: investigate each finding to a root cause yourself (using cclens sql against the store, and the user's actual config files), conclude with a prioritized fix-plan naming specific file edits, and pause only for the user's approval of that plan before editing anything.

Delete the temp file (rm -f "$PROMPT") as soon as you have read it. Do not paste the briefing wholesale back to the user — deliver the conclusions.