docs: how a reply to a contributor is written, and the older bug #13 surfaced

[rokbenko]

Two things #13 turned up that were not in its diff.

CONTRIBUTING gains "How the reply is written"

The reply to a contributor goes out under a person's name, to a person. The rules that keep it reading that way were being carried in somebody's head rather than written down, which meant they had to be said out loud again:

• no dashes, at all, except inside things that genuinely have one (google-genai, gemini-3.8-flash, --provider, branch names, identifiers)
• none of the assistant tells: "Great work!", "Overall,", "In summary,", tidy parallel triples, a closing line restating the paragraph above it
• when something is declined, say what it lost to. Usually timing, not judgement. If they were right when they wrote it, say so in as many words
• thank them for the specific thing, not for "the contribution"

The section is written to its own rules, so the file carries a worked example of the thing it is asking for. Verified: it contains no em dash, no en dash, and its only hyphens are inside a package name, a model id and a CLI flag.

The CHANGELOG says how old the Gemini bug is

Rule 5 says older breakage a review surfaces gets fixed and said out loud in the same release, rather than left for later.

google-genai has validated function declarations since 2.x, so the exclusiveMinimum bug refused every verb call on every robot for every Gemini model, not only Gemini 3. It survived because 0.8.0's own "non-Anthropic default model IDs are unverified" bullet was true to the letter: Gemini's default had never been sent a request from here. A --provider nobody runs is a --provider nobody finds the bugs in.

Credit added to @Bayway (#13), who ran it.

Gate

tests/test_docs.py and tests/test_catalogue.py pass, which are the two suites that read these files.

🤖 Generated with Claude Code