5dive CLAUDE.md

5dive CLI — contributor rules

The test corpus is TIERED and BUDGETED IN WALL-CLOCK (DIVE-2525)

tests/*.sh went 96 → 267 harnesses in 13 days against four harness deletions ever. A guard is bought once and paid for on every future change, forever, while its benefit stays fixed — so the ledger tilts by construction and no single decision to add one is ever wrong on its own merits. See community/wiki/guards-compound-in-cost-and-nothing-ever-retires-one.md.

• The PR trigger is ADVISORY. It does not block a merge, and it is deliberately not a required context: it is paths-filtered, so it is absent from ~28% of PRs, and GitHub leaves a required check that never reports pending forever. A red there is a signal to read, not a gate. It also covers ONE axis — a regression in the harness corpus outside the core tier — and not the release path (release-cut.yml never runs on a PR) or the selfcheck probe class, which the corpus never reaches. Do not read "full-sweep runs on branches" as "a tree-level regression can no longer land behind a green PR".

• The nightly tier is not a cheap subset: 25% of the corpus by file count and 81% of it by wall-clock (1210s of 1481s, measured DIVE-2789), because membership is selected on cost. Cost any plan that leans on the core/nightly split in seconds, not in files — community/wiki/a-demoted-tier-is-not-a-cheap-subset-it-holds-almost-all-the-cost.md.

• A new harness is core unless you say otherwise. Nothing to write.

• Over budget, CI FAILS (exit 4, distinct from a test failure's exit 1) and names the slowest files in the tier. Past the cap a new guard replaces or merges an existing one — that is the reverse gear, and it is the point.

• A budget red confirms itself first (DIVE-2592). PR #395 went red at 356s and green at 289s on the same commit, no rebase — a 67s swing, all of it in one network-priced harness. So when the sum comes in over, the runner re-times the 3 slowest files and keeps the smaller sample per file: noise is one-sided (contention and a slow remote only add time), so the low sample is the least contaminated estimate, while real growth appears in both samples and survives. Paid only on the red path; a green run re-times nothing. --confirm-top=0 disables it, which can only make the gate stricter — there is no flag that confirms more, and no CI job passes one.

• A budget red says it is a budget red. exit 4 beside "0 failed" reads as systemic to the author and as flake to the next reader; it is neither. The message now states that no test failed, whether the number was confirmed twice, and the smallest set of harnesses that covers the overage — the actionable set, not the top-10 leaderboard.

• Three ways out, in order: merge by subject (hundreds of harness files for one CLI means the unit of organisation is the incident, not the subject — folding two files about one subject reclaims their setup cost and drops no assertion); retire a guard whose class can no longer occur; or demote: