Skip to content

HANDOFF.md

Latest commit

 

History

History
103 lines (88 loc) · 7.21 KB

HANDOFF.md

File metadata and controls

103 lines (88 loc) · 7.21 KB
Raw

Handoff — 2026-10-01

State of play for the next session. Conventions, how-tos and release history live in README.md; the agent-facing subset is CLAUDE.md.

Where things stand

2026-10-01, later: v1.18.0 is released and waved (22 targets, converged): the validator fix (#73 — every clean Dependabot PR had carried a red required check since v1.15.0), the implementer prompt's comments item (#74), the subagents rule (#72), plus the kit-only changes since v1.17.0. Canary and audit: README.md. driver-onboarding is a wave target for claude-standards.md only.

2026-10-01: v1.17.0 is released and waved. claude.yml is a reusable + thin stub (#60), the canary on vite-plugin-shopify-clean passed every assertion in claude-yml-wave-plan.md "Running it", all 21 targets are on the tag, and the audit reads converged (driver-agents' hand-installed lint.yml, outside the wave, was re-copied by hand the same night, ed2ad6a). Release record and what is not yet proven: README.md.

To-dos are the open todo issues on this repo since 2026-10-01 (#61–#64 carry the former CLAUDE.local.md items); that file is retired across Driver repos.

2026-09-30 (#59) settled the decisions the release was built on:

  • The two retired review rails are deleted from .github/workflows/ — this repo was their last home (fleet callers went at v1.12.0); any tag through v1.16.0 still holds them.
  • Identity unification is dropped, its scope doc deleted: the claude[bot] / driver-digital-agents split is load-bearing for the driver-agents review loop. The two invariants worth keeping moved into CLAUDE.md.
  • claude.yml becomes a reusable + stub, and the next claude.yml wave is that conversion carrying every ride-along: claude-yml-wave-plan.md, reusable-conversion-scope.md.
  • The model stays fable; Fable billing is no longer tracked. Marcella-NYC-Main review coverage: not doing it.
  • The round-marker arm is dead, not a re-entry point. The Macroscope revise loop shipped in the driver-agents dispatcher (2026-09-18) using a plain tag-mode @claude, and the Macroscope → Bonsai receiver was retired unbuilt (2026-09-29). macroscope-integration-scope.md now describes the loop.
  • House standards ship as a kit file, templates/github/claude-standards.md, imported by CLAUDE.md here; fleet repos copy it once and the wave keeps it current (added to FULL_FILES and the audit).

Open decisions

  • dependabot-report's future. It runs Claude automatically on every Dependabot PR (verdict over the inert artifact, never the diff). Macroscope reviews Dependabot PRs too since 2026-09-10, so it is the one place two bots still review automatically. Keep, or retire like the review rails.

Watch-items

  • Dependabot PRs opened before v1.18.0 still show the old red validate on their current head; a @dependabot rebase or any push gets a fresh run. Nothing to do fleet-wide; they clear as they move.
  • The first real ticket through v1.18.0. The guard steps should stay quiet; if one fires, the failure note lands on the issue and, on a private repo, the transcript is in the run log.
  • Major bumps of the kit's floated actions. actions/checkout@v7, actions/upload-artifact@v7 and anthropics/claude-code-action@v1 float in the kit; when a new major ships, a fleet repo's Dependabot moves ahead and the audit reads the kit as behind until the kit's major is bumped. Below a major boundary the foundrae-blackridge #174 drift-and-rollback cannot recur.
  • The cooldown exemption is still unverified live: at v1.17.0 the wave repinned within minutes of the tag, so Dependabot had nothing to bump. vite-plugin-shopify-clean remains the one to watch.
  • WebSearch/WebFetch stay off until #690 ships a fix (still open at 2026-07-28); the action floats on v1, so the fix arrives on its own and the caveat comment in claude.yml is what gets removed.
  • claude-standards.md loading in CI — confirmed on the issues path (Avara run 36789868486 carried both CLAUDE.md and .github/claude-standards.md as attachments; Avara #226, 2026-10-01). The comment path is still unverified: the @claude must come from a collaborator or the dispatcher (bot comments are author_association NONE), and the transcript never prints loaded instructions, so have the implementer quote the attachment header. The PR-run gap (the import resolves to the PR head's copy) is accepted: the file is style-only, now a CLAUDE.md invariant (Maria, 2026-10-01).
  • A human @claude (tag mode) still gets the action's own co-author text; the nine-item quality standard is global to --append-system-prompt — both unchanged.

Recommended next steps

  1. Make bonsai-link a required check per repo — dropped (Maria, 2026-10-01): the check skips PRs with no Bonsai mention, since most human PRs have no ticket and a red X read as a failing build; a required check that passes when skipped would enforce nothing. The wave also installs claude-standards.md beside claude.yml since 2026-10-01; the @.github/claude-standards.md import is a todo issue in each repo that lacks it. driver-agents joins the fleet as an implementer target (driver-agents #50, Maria 2026-10-01); the next wave after its kit install reaches 22 targets.
  2. One real ticket end to end on the new rail, transcript read (private repos log it) — after driver-engineering-app's security hardening pass (Maria, 2026-10-01: a couple of weeks; she will not run tickets through the app before it). On Avara, the first store run's log must read Provisioned store 'avara' and the "Mint the store token as a log mask" step must pass — never print the token cache to prove the mask.
  3. fleet-wave.sh gains .macroscope/check-run-agents/ (#61) — after the first real Avara design ticket tunes the rubric (driver-agents #32 holds the prompt; Avara's copy merged 2026-09-29). The dest() helper is where a second root goes.
  4. Fleet dependabot.yml standard (#62): the kit block, cooldown included, is the candidate; the gaps are in fleet-operations.md.

Pointers

  • README.md — release + repin order, what's in the kit, dated release history.
  • fleet-operations.md — wave mechanics, the fleet counts, what the pin audit cannot see, branch protection.
  • claude-yml-wave-plan.md — the next implementer wave and its gates.
  • macroscope-integration-scope.md — the 2026-09-12 decision that Macroscope owns automatic review, and the driver-agents loop that replaced the review rails. Observed 2026-09-10: it re-reviews every push, resolves its own threads once a push addresses them, and its verdict reads Approved at <sha> once nothing is left; a fleet-changing kit release gets "not approved" on risk with zero findings.
  • driver-agents — the dispatcher (since 2026-09-10) and the box-retirement spec.