box/CONTRIBUTING.md
dan-claude-bot f63fa952d4 refactor: one drill record per version, in drills/
Drill records move from sections inside drill/RUNS.md to one file per
version: drills/<version>.md. drill/RUNS.md is untouched — it stays the
harness's own run log, traps table and lore. drills/ is release evidence
only, and the docs now say which is which.

The old guard parsed headings: em-dash field matching, an optional
' — DATE' tail, whole-version comparison so 0.9.0-rc1 could not satisfy
0.9.0, avoiding \x escapes because CI runs mawk not gawk, and a non-blank
body rule. Every one of those existed only because records shared one
file, and two separate defects were found in review because of that
complexity — the sed '/./,$!d' whitespace bypass, and heading-grammar
drift from the sibling repos.

One file per version makes almost all of it unrepresentable: 0.9.0.md and
0.9.0-rc1.md are simply different files, so whole-version matching is
free rather than a trap. The guard keeps only the rule that was never
about headings — a record of pure whitespace is not a record — and keeps
the failure message that names the version, the path, the unblock, and
the recorded-waiver escape hatch.

Plain drills/, not .drills/ — a dot-directory is invisible to globs
without dotglob, which is what caused #116 and #118.

Also corrects the release-drill framing: the three repos' drills are
INDEPENDENT and run in any order. What dissolves the box<->rig recursion
is that every drill pins the same fixed candidate refs, not sequencing.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-21 16:35:28 +00:00

18 KiB

Contributing

How change lands in this repo. The short version: PRs are born as drafts, three reviewer bots take the first rounds, a human takes the last word — and labels tell you where everything is without opening anything.

The PR loop

  1. Fork and branch. Contributors work from forks; upstream branches are for maintainers. Title the PR conventionally (feat:, fix:, docs:), and include a CHANGELOG.md entry under ## Unreleased when the change deserves one.

  2. Open as a draft while you build. Drafts are invisible to the reviewer bots on purpose.

  3. When it's ready: mark ready-for-review and request all three bots — claude-bot-andresmgsl, codex-bot-andresmgsl, grok-bot-andresmgsl. They poll roughly every 15 minutes.

  4. Rounds are answered whole. Wait until all three have reviewed, then answer the entire round in a single reply, push the fixes, and re-request the bots that didn't approve. Prefer verification over argument: a test settles what a comment thread can't.

  5. Reviews end in a verdict. A reviewer — bot or human — either approves or requests changes, never a bare comment. A comment-only review is a non-verdict: it doesn't say whether the round passed, and the state machine (and anyone scanning the board) has to guess. The verdict carries blockingness only, the body carries the feedback: non-blocking nits ride an approval and the author addresses them at their discretion; anything blocking — including a question that gates the verdict — is request changes, saying what unblocks it. The reconciler treats a comment-only review as not-approved, so commenting without a verdict only stalls the PR. The machine never reads review bodies: when a comment-only reviewer's line is really an agreement, that judgment belongs to the author — escalate by requesting the maintainer's review (step 6), and the reconciler flips the label on that request, because an explicit request is a fact it can trust.

  6. When the round passes, the author hands the PR to the maintainer in three acts, in this order: post the tagged round summary, request the maintainer's review, then set state:needs-human yourself — removing the state label it replaces. The review request is what earns the label, provided the PR carries no blocker:* label. A blocker means the work is still yours whatever the round said, so on a conflicted or red PR neither the request nor your own label write will stick — the sweep takes it straight back off. With three formal head-current approvals the labels workflow requests the maintainer automatically; when part of the panel is comment-only, reading their agreement is the author's judgment, so the author makes the request.

    Writing the label by hand is an optimistic write, not a transfer of ownership. The machine stays the authority — but because the workflow wakes on labeled, the author's own write fires the sweep that validates it, and a handoff that had not earned the label is corrected seconds later. Forgetting the write is not a failure either; it only means the label waits for the cron, which is the lag this replaced.

  7. Checks must be green: shellcheck and bash test/cli.sh locally mirror what CI runs; the multi-user rehearsal runs in CI on a real Incus.

Changelog entries

Every PR that changes behaviour adds one line to ## Unreleased. One line is the whole rule — if it wraps more than twice in your editor, cut it down.

  • Say what changed, and stop. Why it was wrong, how it was found, what it cost, what it implies — that belongs in the PR body and the commit message, which is where anyone chasing the reasoning already goes. This file answers one question: what is different in this version.
  • Any word that can be removed, is removed.
  • Lead with the surface, not the mechanism. "state:needs-human is set at handoff" beats "the labels workflow now also wakes on labeled".
  • Cite the issue or PR(#141) — and let the reader follow it for the rest.
  • Mark a breaking change with a leading BREAKING:.
  • Group under ### Added / ### Changed / ### Fixed / ### Removed.
  • No bold run-in headings, no sub-paragraphs, no code blocks, no prose essays.

Good:

- `state:needs-human` is set at handoff, not by the cron (#141)
- An unreadable check rollup no longer reads as "nothing is failing" (#136)
- BREAKING: `--class human|server` is now `--root-door closed|open` (#77)

Not an entry — that is a PR body:

- **`state:needs-human` no longer waits on the cron to become true** (#141) —
  the labels workflow now also wakes on `pull_request_target: labeled` and
  `unlabeled`, and the author sets it themselves when handing a PR over. A
  review landing was never a trigger. There is no `pull_request_review_target`,
  and on fork PRs — which is all of them here — ...

Releases

A release is a PR, and merging it ships it (#96, building on #83):

  1. The release PRrelease: X.Y.Z, labeled release — bumps VERSION from X.Y.Z-dev and stamps the ## Unreleased section with version + date (feature PRs land their changelog entry as part of the PR, so the section is already written).

    Stamping is two edits, not one — the second is re-arming. After rewriting ## Unreleased into ## X.Y.Z — DATE, put an empty ## Unreleased back at the top, immediately above the section you just stamped:

    ## Unreleased
    
    ## 0.7.1 — 2026-07-19
    
    ### Fixed
    ...
    

    Not cosmetic, and not deferrable to the next PR that happens to need it. Between the stamp and the next re-creation of that heading, main has no ## Unreleased. A PR authored before the release wrote its entry under that heading; with the heading gone, git lands the entry under whatever now occupies the position — the section that just shipped — and it merges cleanly, no conflict, no signal. The changelog then credits a released version with a change it does not contain, and nothing but a human reading the file will ever say so (#108; confirmed in the sibling repo as heavy-duty/rig#66).

    CI enforces the arming rule with .github/scripts/changelog-armed.sh, keyed on VERSION: a -dev tree must carry ## Unreleased on top; a bare-VERSION tree (the ceremony PR, and the merge that publishes it) may carry either ## Unreleased or its own stamped section. That is why the guard cannot simply demand ## Unreleased unconditionally — the unconditional form is false on the ceremony PR's own tree and makes the release unshippable, which is why rig and cast both reverted it. The practical consequence: forgetting to re-arm does not block the release PR, it turns main red on the very next push — the automatic -dev bump the release itself makes. Do it in the ceremony PR and main is never disarmed at all.

    Release headings are append-only. When you add your entry under ## Unreleased, insert above the heading below it — never type over that line. Replacing ## 0.8.0 — 2026-07-19 with your own ## Unreleased block deletes a shipped section: its prose is absorbed into Unreleased, release-notes.sh can no longer find the version it extracts by heading, and the next release republishes the absorbed prose as if it were new. git merges that edit cleanly and changelog-armed.sh stays green on it — the top section is still the right one — so .github/scripts/changelog-monotonic.sh asserts the other half on every PR: the set of ## X.Y.Z headings on your branch must be a superset of the set at the merge base (#122, caught in review of #118). The ceremony's own stamp passes it by construction — rewriting ## Unreleased into ## X.Y.Z — DATE adds a heading and removes none.

    The drill gates the release, and CI enforces it. The release PR's flow is: draft → ready → bot round → drillstate:needs-human → maintainer merge (which is the release). CI proves the tier's semantics on every PR; the drill is what proves the VM trust boundary, and it wants real hardware and the better part of an hour, so it runs on the release PR's branch and nowhere else.

    Record it in its own file, named exactly for the version:

    drills/X.Y.Z.md
    

    ...with the run in it — what it measured, what it found, what it cost. See drills/README.md for the shape. .github/scripts/drill-recorded.sh asserts exactly that on every tree with a bare VERSION, which is every release PR and the merge that publishes it; a -dev tree passes with nothing to assert. It is no longer a thing a reviewer has to remember.

    drills/ is release evidence; drill/RUNS.md is the harness's own log. Two artifacts, two purposes. drill/RUNS.md keeps being what it has always been — every run of the harness, the traps table, the lore — and is not release-scoped. drills/X.Y.Z.md is the per-release record the gate reads. Appending to drill/RUNS.md does not satisfy the gate, and is not meant to.

    One file per version is also why the guard is short. When records shared a file it had to match em-dash heading fields, tolerate an optional date tail, and compare versions whole so 0.9.0-rc1 could not satisfy 0.9.0 — and two separate defects were found in review because of that complexity. Now 0.9.0.md and 0.9.0-rc1.md are simply different files and none of it is representable.

    It became CI's job because remembering did not work. The sentence this paragraph replaces described a step no release had ever performed — there was no drill record anywhere in the repo — and #95, #114 and #148 all shipped through the gap as a VERSION bump plus a CHANGELOG.md stamp. The one time it was caught was the one time somebody happened to look, which is not a gate.

    The three drills are INDEPENDENT — there is no fixed order. box, rig and cast each drill separately, in any order, on any schedule, in separate sittings if you like. What makes that safe is not sequencing, it is that every drill pins the same fixed set of candidate refs: rig's drill runs --host yes with BOX_REF=release/<box-version>, so it exercises the box that will actually ship; box's drill mints with RIG_REF=release/<rig-version>, so it exercises the rig that will actually ship. Both measure the same pair.

    That is what dissolves the box↔rig recursion. box and rig really are mutually recursive — rig sits below box as the host-builder (rig bootstrap --host yes installs box and runs setup-host) and above it as the guest-converger (a box new seed's cloud-init curls rig's installer and runs rig bootstrap <tenant>-box); box's own source says as much, the seeds "invert the rig→box install edge (rig#28: rig installs box on hosts; now box guests install rig)". A cycle at runtime becomes two independent tests against one fixed pair, because the refs are static identifiers that exist as soon as the release branches do, long before any drill runs. Nothing has to be released, or drilled, first.

    Within a single drill you of course bring the substrate up before probing it — a host before a guest. That is how you run a drill; it is not an ordering rule between repos.

    Each repo drills in a different way and asserts a different thing: box asserts the isolation contract (the 85-probe VM trust boundary), rig asserts convergence (a machine reaches its role, idempotently), cast asserts promotion (A→B reproduces, the diff is idempotent). Three different exercises sharing a substrate — not three phases of one script, which is exactly why the records are per-repo.

    It drills candidate refs, not released artifacts. RIG_REPO and RIG_REF are mint-time environment variables (defaults heavy-duty/rig and main, bin/box), so a run pins the exact commits under test. That dissolves the chicken-and-egg: no repo has to be released before another can be drilled. And drilling the candidate is drilling the release — a release PR's diff is VERSION plus CHANGELOG.md and nothing else, so no executable byte differs between the tree that was drilled and the tree that ships.

    The drills of a release set share one run ID. Each repo records its own legs in its own drills/X.Y.Z.md, citing that run ID and the other two repos' commit SHAs, so the three records reconcile afterwards. The guard reads only this repo's file — it asserts box's record exists, never the other two.

    If a defect shows up only in the combination: patch, re-drill, re-record. The three releases converge on a set that holds together; they are not required to be right in one pass.

    A known gap, and box is where it belongs. A released box still defaults RIG_REF to main, so what a user mints a week after a drill is not the combination that was drilled — the guest converges against whatever rig's main has become since. Pinning RIG_REF to a released rig tag in the templates is the outstanding step from #81 (rig#32 step 5), and it is what would make a drilled combination reproducible for users. This PR does not fix it; box's source already says the two directions "track main unpinned today, said honestly ... until a release flow exists".

    A maintainer waiver is possible, and it is still written down. The guard requires a record, not a passing result, so a release that must ship without a full drill creates drills/X.Y.Z.md anyway and says plainly that the drill was waived and why. Skipping then costs a deliberate, reviewable file in the diff — which is precisely what the three silent skips above did not.

  2. The maintainer's merge IS the release. release.yml fires on the merged, release-labeled PR and asserts before creating anything: VERSION at the merge commit is non--dev and changed in this PR (the -dev interlock — a mislabeled ordinary PR fails loudly and creates nothing), the version's CHANGELOG.md section extracts non-empty, and no tag or release exists for it yet. Then, in the same job, it tags the merge commit bare X.Y.Z (no v prefix, the 0.6.0 precedent) and publishes the GitHub release with that section as the body. No assets — the source tarball for the tag is the package, and install.sh downloads exactly that.

    Manual fallback/backfill: the tag-push path stays. Tagging the merge commit bare X.Y.Z by hand and pushing the tag still publishes the same way (release.yml asserts the tag names the tree's own VERSION) — for backfills, or the day the merge path is red.

  3. The release re-arms main itself: the same workflow run bumps VERSION to X.Y.(Z+1)-dev and pushes the commit straight to main — no follow-up PR (it opens one only if branch protection refuses the direct push, and says so loudly). Not cosmetic — the versioned layout names install trees after VERSION, so a main install without the bump would land in versions/X.Y.Z and impersonate the release just cut. On the manual tag path the bump stays yours: open the one-line PR after publishing.

Labels — who sets what

The full taxonomy lives in LABELS.md. What matters day to day is who sets each kind — most of it is machinery, and hand-moving a machine-owned label just gets corrected on the next pass:

Labels Set by
state:* the labels workflow (.github/workflows/labels.yml) — recomputed from GitHub's own facts on PR events (label changes included) and every 15 minutes. Machine-owned, with one exception: the author sets state:needs-human at handoff (step 6) and the workflow reconciles it. Otherwise never by hand. Exactly one per PR: whose ball is it.
blocker:* the same workflow, from the same facts — what is in the way. Any number per PR, or none. Never by hand: applying one does not stop a merge, and removing one does not unblock anything. Fix the thing and the next sweep drops the label.
stale the same workflow — 48h without commits, comments, or reviews. blocked PRs are exempt: they are quiet legitimately.
scope:* on PRs actions/labeler, from the changed paths (.github/labeler.yml). Additive — you may add more, the machine won't remove them.
scope:* on issues you, when opening or triaging — issues have no paths to derive from.
blocked, release you — automation never guesses intent.
merge-next you or the agent owning the queue. Which PR lands first is a judgement about how they conflict, so the workflow never sets it — it only clears it, the moment the PR stops being something a human could merge.
bug / enhancement / documentation you, on issues only — a PR's type already lives in its title.

Issues

Give issues the same care as PR titles: say the surface in the title, apply a scope: label and a type label (bug / enhancement / documentation) when you open one, and blocked when it waits on something — that is what keeps the board navigable as the issue count grows.