[Optimization] Streamline fffind & ffgrep promptGuidelines into concise heuristics

[Ran-Xing]

Summary

@ff-labs/pi-fff currently injects 10 separate promptGuidelines entries (~800 characters) across fffind and ffgrep into Pi's System Prompt on every session.

Suggestion

Many of the 10 bullets describe overlapping usage heuristics (e.g., bare identifiers, path include/exclude, glob usage, 1-2 term queries). These can easily be consolidated into 2 concise summary bullets:

1. ffgrep: prefer bare identifiers, literal queries, and path scoping (e.g., path: 'src/', exclude: 'test/').
2. fffind: whole-path fuzzy search (1-2 terms); use path glob (e.g., path: '**/file.h') for exact matches.

This keeps the guidance clear and effective while cutting the token footprint in half across all user sessions.

[Ran-Xing]

For anyone looking for an immediate workaround while upstream optimizations are considered, we created a lightweight middleware extension pi-prompt-diet that dynamically prunes and condenses bloated promptGuidelines across installed packages without sacrificing any tool capabilities.

[gustav-fff]

[triage-bot] NEEDS EVALUATION: not actionable as a blind text cut.

The 10 bullets in packages/pi-fff/src/index.ts:858 (grep), :1048 (find), :1189 (multiGrep) are not decorative prose. They were tuned against observed model misuse (84ba575 feat(pi-fff): Improve instructions and add more features (#414), 3298ed9 fix(pi-fff): prefix promptGuidelines with tool name (#737)). Specific bullets encode non-obvious semantics that models get wrong without them:

• find matches the WHOLE repo-relative path, not the filename — models assume filename-only.
• bare pattern is fuzzy; exact match requires a glob in path ('**/profile.h').
• directory listing needs path: 'dir/**', not pattern alone.

Your proposed 2-bullet version drops caseSensitive, exclude, the "after 1-2 greps read the top match" stop condition, and the dir/** form. That is not a lossless compression, it is a behavior change on the tool-call side, and this repo has no eval harness to measure tool-selection regression. Cutting ~400 chars of system prompt in exchange for an unmeasurable drop in grep/find accuracy is a bad trade.

Requires an additional evaluation layer before any guideline text is touched. Not fixable right now. pi-prompt-diet is the correct place for this if you want it today — it is opt-in and the cost lands on whoever chose it.

Honk-Honk 🪿