claude.yml: the `opened` mention gate is formatting-blind, so discussing the bot dispatches a run

[d-morrison]

Split out of #298 before merging #552, so the remainder is not lost when the parent closes.

#298 raised two coupled problems. #552 ships the first — assignment as a sufficient trigger, via dispatch-on-assignee. This is the second, quoted from #298's own problem statement:

opened genuinely needs a content signal, since every new issue fires it. But the substring match is formatting-blind, so an issue that merely discusses the bot in a code span dispatches a full run (four instances so far; most recently ai-config#682 -> #683, which ran ~8 minutes and then died at the push step).

Why it is not simply "reuse the gha#342 fix"

gha#342 already solved this class inside the job: detect-bot-mention pipes the body through strip-non-invoking-markup.sh, which removes blockquotes, fenced blocks, indented blocks, and inline code spans, so a quoted mention no longer dispatches an agent run or a review re-dispatch.

What it does not fix is the job-level if:, and gha's own CLAUDE.md already says why:

Note what it does not fix: claude-bot.yml's job-level if: still tests the raw body, because a GitHub expression cannot strip Markdown, so the job still starts and the runner still spins up.

So the remaining cost is bounded but real: a runner spins up, the checkout runs, and the job stands down at the gate. What #342 removed was the billed agent run; what is left is the runner time and a confusing skipped job on an issue that was only talking about the bot.

Directions, none obviously right

1. Accept it. The expensive half is already fixed, and a GitHub expression genuinely cannot strip Markdown. Document the residual and close.
2. Cheap pre-job filter. A tiny first job that runs the stripper and gates the real job on its output — moves the decision out of the expression language at the cost of one extra job per event.
3. Narrow the expression heuristic. e.g. require the mention at a line start, or not immediately preceded by a backtick. Partial by construction, and Bot mention trigger matches inside code spans and blockquotes, so writing about the bot invokes it #342's own history is a warning about approximating CommonMark: it tracks the spec closely because under- and over-stripping land on different callers.

Direction 3 looks worst — it re-opens the approximation problem #342 deliberately closed, in a language that cannot express the fix. Between 1 and 2, the question is whether the residual runner time is worth an extra job on every triggering event, which is a maintainer call rather than a technical one.

Related

• claude.yml: make issue assignment a sufficient trigger, independent of the mention substring #298 — the parent; its assignment half ships in fix(claude.yml): make issue assignment a sufficient trigger #552.
• gha#342 — the in-job fix, and the source of the "a GitHub expression cannot strip Markdown" constraint.
• gha#553 — separately, the audit rule that fix(claude.yml): make issue assignment a sufficient trigger #552 showed names one direction out of three.

Not claimed.