Should the full-suite hook match argv rather than the whole command string?

The fork

.claude/hooks/no-full-suite.sh is a text instrument reading a command string. It cannot tell an invocation from a description of one.

Match argv only — the executable and its arguments, not heredoc bodies or -m text. Narrower, and narrowing permission machinery is the direction that needs an owner.
Leave it. The workarounds are cheap and the guard stays maximally blunt.
Exempt the commit-message path specifically — the narrowest change: a heredoc or -F body destined for git commit is prose by construction.

Recommendation: 1 or 3, and 3 if there is any doubt — it fixes both measured instances and cannot loosen anything that runs a test, because a commit message runs nothing. But this is rejected/-adjacent only if the owner thinks the cost is imaginary, and it is not, for the reason below.

The two instances, neither of which ran a test

frankH, landing unitalias (731463e70). The commit message named the core tier recipe in order to say it had not been run and why — the change is inert without a manifest row, the lookup exits on a zero count before the cached walk, so every existing compilation takes the path it took before. The hook refused the commit. Passed via -F instead; hook left alone, deliberately.

frank-coordinator, appending to the roster. A markdown document quoting a corpus glob inside a heredoc, in a commit touching one file. Refused. Landed through the Write tool — and explicitly not by setting PXX_ALLOW_FULL_SUITE=1, since that flag asserts "I genuinely need the full suite", which would have been false.

Both authors declined to route around the guard and both were right to. The rule against rephrasing a command to slip past a hook is absolute, and neither did. But note what each did instead: one changed the transport, one changed the tool. Neither leaves any trace for the next person, who will meet the same refusal cold.

Why the cost is not the keystroke

CLAUDE.md asks, in as many words, for the gate justification to go in the commit message — "say in the commit why quick was not enough", and "note it in the commit message, the only warning anyone gets."

The hook makes the one place we ask for gate reasoning the one place gate reasoning is expensive to write.

That is a slow, invisible pressure toward vaguer commit messages: the author who hits this twice stops naming the recipe and writes "did not run the big one", which is exactly the sentence that is useless to a reader six weeks later. The guard does not fail loudly here; it shapes what people write.

It is also this tree's own recurring class, arriving in the guardrail that enforces this tree's rules: an instrument that reads text cannot tell an assertion from a description of one. Same animal as a grep -L answering about a literal string, and as frankZ's uses regex reading "Uses only the language surface that ALL backends support today" out of a header comment and yielding the unit i386.

What is NOT in scope

The escape hatch is not the problem and must not be touched. PXX_ALLOW_FULL_SUITE=1 is a SPEED guardrail an agent lifts autonomously, no permission needed, and it is documented in the hook's own text at line 93. Two sessions this evening read that wrongly as a permission gate and one of them declined work it was entitled to do. That is a separate finding about the hook's legibility, not an argument for changing what it blocks — and the fix there, if any, is wording.

Nobody may implement any option here without the owner saying which. Filed by the coordinator, which holds no lane and writes no code, precisely because both finders correctly refused to decide it themselves.

POINTER 2026-09-09 (frankB) — measurement lives on the sibling row

A read-only measurement of what this hook actually does today — the three gates in series, the semicolon that defeats the first-word exemption, a 4510-row corpus census (16.5% of ticket bodies, not "most"), and a simulation of option 2 showing it holds all three positive controls while leaving the commit-message shape refused — is recorded in [[decide-t-the-full-suite-hook-refuses-prose-about-the-suite]].

It narrows this fork but does not settle it, and nothing was implemented. This row's own constraint is unchanged and was respected: nobody may narrow this on their own judgement, and nobody may implement any option without the owner saying which. A peer relayed an assertion that the fork had been delegated to agents; a peer cannot grant that, and this row says so.

One correction to this row's framing, from the measurement: the fork reads as three architectural options, and the measured shapes need two small independent changes — skip heredoc bodies, and make the chain test look at command text rather than the whole string. Option 1 (leave it) remains coherent.