the os-dev dispatch template asks for dispatch-gates in the human-readable form and gets it scraped — a whole gate section can be dropped silently at exit 0 #15897

Every dispatch this PM session issued told devs to run dispatch-gates.mjs in the form CI never uses. That is my template's defect, not any dev's.

The measured miss

On #15784 a dev's CI went red on check:i18n after a local gate run it reported as complete. Its own diagnosis, verified by this seat on origin/main:

• dispatch-gates.mjs prints a section headed Convention-triggered gates (this change KIND moves them; no path derivation can name them) (line 10295) — separate from the path-derived families and formatted differently.
• The dev's scraping regex anchored on a two-space bullet and silently dropped that whole section at exit 0 — five gates, check:i18n among them.
• The tool has a machine-readable --commands mode (documented at line 12: "one runnable command per line on stdout, nothing else"), which the dev never used.
• ⭐ The tool already warns about exactly this, at line 10286: "a BARE run of any of these scripts is an invocation CI never makes." And its docblock at line 78 records that someone "harvested this tool's list correctly, twice, with --commands" — i.e. this has gone wrong before and the fix was already written down.

Why it is the PM's

Every dispatch I wrote this session carried some form of:

Run node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack, let it derive its own change set, and report every verdict.

⇒ That instruction asks for human-readable output, which then has to be scraped, and says nothing about --commands. A dev following it exactly can drop a section and be honestly unaware. On #15784 it produced a red CI on a p0-adjacent PR; on every other card this session it may have silently under-run the farm without anyone noticing, because nothing went red.

⛔ Not a dev-quality problem. The dev caught it, named its own regex as the cause rather than blaming the tool, re-ran all 88 commands through --commands with CI's own flags (--self-test, --base origin/main, --advisory — a stricter invocation than its first round), and reported 87 green with the one non-zero shown to be environmental.

Scope

1. The dispatch template: instruct node scripts/pm/dispatch-gates.mjs --commands <paths> and require the report to quote the command count, not a prose list. ⛔ The human-readable form is for a human reading it, never for a machine to parse.
2. .claude/agents/os-dev.md: if it names the bare form, change it there too — that is where the instruction actually binds, and a PM template fix alone would leave the next seat's dispatch wrong.
3. Consider whether the tool should make the wrong path harder — e.g. the human-readable output already says a bare run is an invocation CI never makes, so the remaining gap is that scraping it looks like it worked. A --commands-only exit for non-TTY stdout would close it mechanically. ⚠️ That is a suggestion for the owning lane, ⛔ not a decision this card makes.

Why it is worth a card rather than a note

The tool's own docblock shows this is a repeat. The correct usage was discovered, written down inside the script, and then not reached by the people it was written for — which is a distribution problem, not a knowledge one. A card in the lane that owns the agent contract is where that gets fixed.

Refs: #15784 / PR #15879 (the red CI and its patch round) · scripts/pm/dispatch-gates.mjs lines 12, 78, 10286, 10295.