dario

# dario — notes for Claude Code (and other LLM coding agents)

## Commits and PRs

- **Every PR that touches `src/` adds a bullet under `## [Unreleased]` in `CHANGELOG.md`, in the same PR.** CI fails the PR otherwise (`scripts/check-changelog.mjs`). The bullet is what becomes the release notes; a PR without one ships silently — v6.0.22 listed two of its six changes for exactly this reason. A pure refactor or test-only change with nothing user-visible takes the `no-changelog` label instead.
- **Do not append `https://claude.ai/code/session_*` URLs** to commit messages or PR descriptions. They rot fast (the session expires), they leak the source of the contribution into git history forever, and they're not what readers want when they `git blame` a line three years from now.
- **Do not append `Co-authored-by: Claude` or `Generated with Claude Code` trailers** unless explicitly requested. Same reasons.
- The `Co-authored-by:` line GitHub auto-adds for human PR authors during a squash-merge is fine — that one is generated by GitHub, not by Claude Code.

## Style

- **No `Generated by Claude` / `Drafted by AI` markers** in source files, docs, or PR bodies. The maintainer treats those as noise that has to be cleaned up before review.
- **Match the existing prose voice** in docs (terse, active, scar-tissue references to issue numbers when relevant). Don't introduce bullet-arrow patterns (`**Foo** →`), parenthetical inventories, or "Two checks worth running on the way to" framing.

## Scope

- The repo's CI enforces SHA-pinned GitHub Actions (`.github/workflows/actionlint.yml` + org policy reinforced in #209). Add new actions in tag form on first commit only — Dependabot's `github-actions` ecosystem entry pins the SHA on the next Monday run.
- The cc-drift bot (`cc-drift-watch.yml` + `cc-drift-auto-release.yml`) auto-drafts version bumps for `SUPPORTED_CC_RANGE.maxTested`. Don't compete with it: bumping `maxTested` in a manual PR will collide with the bot's next run.

## Codex drift runner

`codex-drift-watch.yml` runs daily on the `dario-drift` self-hosted runner against a dedicated live-test ChatGPT account; the `DARIO_CODEX_LIVE_HOME` repository variable holds the absolute HOME whose `.dario/codex-accounts/` directory contains that account. Without the variable the workflow is inert by design. The workflow reads the account in place, masks its ChatGPT account id, and never logs or commits its tokens; `test/fixtures/codex-models.snapshot.json` is the seeded baseline it compares against.

## When in doubt, ask

The maintainer reads PR bodies. Surface trade-offs and unknowns explicitly rather than picking a default and shipping.