[stn1slv]

Problem

lazycommit commit and lazycommit pr send the whole diff to the model with no size limit. With a large change set, OpenAI rejects the request (context length or tokens-per-minute limit). Retry and model fallback then resend the same oversized request, and the user just sees an error.

What this changes

• Diff.Compact(budget) in internal/domain: a diff that fits the budget is sent unchanged. A larger diff becomes a list of changed files with +/- line counts, followed by every file diff that fits in full. Leftover space goes to the first file that did not fit, cut at a line boundary and marked [truncated].
• Lock files (go.sum, package-lock.json, yarn.lock, ...), minified files, source maps, and binary files are listed without their diffs.
• New max_diff_chars setting in prompts.yaml, default 40000 (about 10k tokens). It layers like num_suggestions, so a repo-local lazycommit.prompts.yaml can override it. Values below 1000 are raised to 1000 so the file list always fits.
• The git adapter now passes --no-color --no-ext-diff --src-prefix=a/ --dst-prefix=b/, so settings such as diff.noprefix, diff.mnemonicPrefix, diff.external, or color.ui=always cannot change the format the parser reads.

Behavior change: diffs larger than 40000 bytes are now shortened by default. If you use a large-context model, raise max_diff_chars. There is no "off" value yet. I can add one (for example -1) if you want it.

Testing

• Unit tests for compaction (budget limit, generated and binary files, quoted and renamed paths, long file lists, input without git headers), config layering, the pipeline, and the git format flags.
• go test ./internal/... and golangci-lint run pass.
• On a real 385 KB diff from this repo's history, the compacted prompt was under 40 KB, listed all 50 files, and kept 17 full file diffs plus one truncated file.