Feed the model diffs and file stats, not just commit subjects

[tabkram]

Why

parseCommits in src/git.ts:52 gives the model only the commit subjects. On most repositories those are fix stuff and wip, and the model is asked to write a customer-facing note from them.

This is the largest quality gap in the tool, and where the closest competitors are ahead: @entro314labs/ai-changelog-generator and github/copilot-release-notes both read diffs. There is no diff access anywhere in this codebase.

Prerequisite

getChangelog returns string[] and parseCommits hardcodes hash: "" — the hash is thrown away, so there is nothing to hang a diff on. Commits must travel as objects first. No user-visible change, but it touches src/git.ts, src/generator.ts, src/types.ts and the tests.

Side effect: --changelog and --changelog-file produce hashless commits, so they are never enriched. Degradation for free.

Scope

1. File stats — --numstat: which files changed and by how much. Cheap, high signal. One git log --numstat <from>..<to> for the whole range, not one process per commit.
2. Small-commit diffs — full diff under a line threshold, where the code is the summary.
3. Budget — degrade instead of blowing up a large release.

Config

git:
  maxCommits: 200

  # What the model receives for each commit.
  #
  # auto (default): messages + files everywhere, diff wherever the message is
  #                 poor (wip, fix stuff…), within diffBudget.
  #
  # Or an explicit list:
  #   messages — the commit subject
  #   files    — files touched, with +/- line counts
  #   diffs    — full diff of commits under the threshold
  commitDetail: auto

  diffBudget: 30000

A list rather than a single level so it can subtract: on a repo where every subject is wip, [files, diffs] drops the noise and keeps the code.

auto spends the budget on the commits whose message says least — feat(auth): add SSO login already explains itself. An explicit list is reproducible to the token, which the #6 fixtures need. Ties break on hash.

CLI: --commit-detail files,diffs.

This also removes git.commitFormat — declared at src/types.ts:90, read nowhere. Backward compatible, since Zod strips unknown keys.

Safety

• Diffs are untrusted input: same labelled blocks and neutralizeDelimiters as the changelog (source files can contain =====).
• Redact hunks against SECRET_CONTENT_PATTERNS in src/context.ts — a diff adding an API key would otherwise ship it to the LLM.
• Skip lockfiles, dist/, minified and binary files.

Acceptance criteria

• Commits carry their hash end to end; existing tests pass unchanged
• commitDetail and diffBudget documented in docs/configuration.md, defaulting to auto
• commitFormat removed; a config still containing it loads without error
• Diffs labelled as untrusted and redacted against SECRET_CONTENT_PATTERNS
• Budget degradation is deterministic and reported
• Add an eval harness for release-note quality #6 fixtures show measurably better output on the messy-history fixtures
• --dry-run shows what was included, the estimated cost, and what was dropped

Files

src/git.ts, new src/git-context.ts, src/generator.ts, src/prompts/builder.ts, src/types.ts, src/bin/cli.ts, docs/configuration.md

Relates to

The mechanism — hash plumbing, collection, budget, redaction, --dry-run — is unit-testable and does not depend on #6. Only the tuning does: the small-commit threshold, whether auto beats an explicit list, how stats read best in the prompt. Land the mechanism, tune once the harness exists.