Read adaptivity: learn when signatures, ranges, small bodies, or full reads are best

[willwashburn]

Problem

relaywash__Read has strong defaults: signatures for large known-language files, full content for small files, range follow-ups, small body inclusion, and repeated-read suppression. Those defaults are currently static. Different repos may benefit from different thresholds.

Goal

Use observed Read outcomes to tune Read defaults per repo.

Signals to capture

• Read mode requested.
• Returned byte size.
• Whether the next tool was a range Read of the same file.
• Whether the next tool was a full Read of the same file.
• Whether repeated unchanged Read returned empty content.
• Whether signature output included a matching recently searched symbol.
• Whether compaction kept or dropped the Read result once compaction attribution exists.

Tunable profile fields

• tools.read.smallFileLines
• tools.read.smallFunctionLines
• Future: tools.read.defaultMode if evidence supports it.

Learning rules

• If signature reads are frequently followed by full reads, consider increasing smallFileLines or smallFunctionLines.
• If full reads are frequently dropped at compaction, prefer lower thresholds.
• If range follow-ups are common and successful, keep signature-first behavior.
• If repeated unchanged reads happen often, surface stronger dedupe guidance.

Acceptance criteria

• PostToolUse observations include enough Read follow-up data without logging sensitive paths.
• Aggregator can recommend threshold changes with confidence and observation counts.
• Generated profiles are applied by existing Read profile hooks.
• Tests cover signature-to-range, signature-to-full, full-dropped, and repeated-read cases.