A crew run runs the tests it touched, never the gate its check already runs #885

Crew runs ran the full sharded gate inside their pods, while the check on the same push ran the same gate again. That is how the droplet reached load 41 tonight: two crew runs each running tsc plus the four-shard gate, beside a check doing the same.

The check is the gate (landIfGreen needs it green), so running it inside the run as well is paid twice.

• The brief (src/crew/brief.ts, every crew run's system prompt). It used to say "Run the same command yourself before you push." It now says to run the type check and the tests that cover what changed, plus any the repository's CLAUDE.md or AGENTS.md names, and not the whole check, because the forge runs that once.
• CLAUDE.md, "How to work". A crew run runs bun x tsc --noEmit and one bun test call over the files it touched plus the five cross files: reserved-names, studio, shards (the gate list), check-list and hollow. It never runs bun scripts/test-shards.ts --gate.
• Not in this pull: FORGE_CHECK_SHARDS=2 in forge.json's check command. The release agent is opening that one.

The saving, from the run table (trevato/forge, the last 24 hours):

• Crew runs: 74 finished, averaging 1234 s each.
• Crew runs that ran the sharded gate: 32 of them, or 43%, by their logs (test-shards).
• Gate duration: the crew-pull checks that ran it (check passed/failed in Ns) took 821 s on average (median 774 s).
• What that adds up to: about 32 × 13.7 min ≈ 7.3 h of gate run a second time inside crew runs in a day. That time is spent on CPU the crew and the checks share, and also on the run's own wall clock: the gate was roughly two thirds of an average crew run.

Test: test/crew-brief-gate.test.ts, added to the gate list.

• The brief asks for the touched tests and no longer says to run the whole check.
• CLAUDE.md forbids the sharded gate and names the five files.

🤖 Generated with Claude Code