[IanFrelinger]

Summary

A durable coordination channel between Grok Bot and Claude Code: a GitHub issue used as an
append-only message log, plus the repository files that let either agent find it and use it without being
told. No product code.

The first two commits added the layer. The third corrects it: the ownership split and the message
vocabulary it shipped with were both superseded before it merged, and a CLAUDE.md is read automatically
by every future Claude Code session, so a stale one is worse than none.

What is here

File	What it is for
CLAUDE.md	Read automatically at the start of every Claude Code session: the repository facts that keep costing sessions time, and the ownership split.
_handoff/bus/PROTOCOL.md	The message format, the ownership split, and the handoff → merge → drift-audit → done loop.
_handoff/bus/README.md	One page: which issue is the channel, and how to read it.
.claude/commands/agent-bus.md	/agent-bus — read recent messages, say what is outstanding, draft a reply.
scripts/agent-bus-latest.sh	scripts/agent-bus-latest.sh [count] — the last N comments, through gh.

What the third commit changes, and why

• Ownership, in all four documents. Claude Code does all development including merging, and merges
  its own PRs once every required check is green; Grok Bot does admin, materials, and audits each merge
  for drift. Each file says it supersedes rule 5 of the issue's opening comment, which was written
  before the split was settled. The claim that Claude owns Fix packaging seam: stranger path from clone → legible receipt #643 and feat(cert-gate): add admission gate demo showing REJECT → ADMIT flow #644 is gone — those are cursor/*
  branches from 2026-09-20 that this session has never touched.
• The Kind: vocabulary now matches the live channel. status | ask | handoff | block | done, and
  To: claude | grok | both, exactly as #695
  defines them. The invented question / request / info kinds and To: all are gone, along with the
  worked examples that used them. A protocol document that disagrees with the channel it documents is the
  drift this whole layer exists to catch.
• CLAUDE.md gains the load-bearing repository facts: no host dotnet; the five required checks
  under strict + enforce_admins, which is why merges are serial; run-repo-gates.sh running its
  discovered scripts bare; the knowledge graph needing git add before it is regenerated; the readiness
  glob list that exists twice; a skipped readiness lane counting as a pass; and that a check never
  observed failing is not evidence.
• agent-bus-latest.sh: #!/usr/bin/env bash, as all 160 other scripts under scripts/ use; the
  issue number overridable through ASHLAR_BUS_ISSUE; a non-numeric count rejected instead of spliced
  into the jq expression.

Verified locally

• bash scripts/ci/run-repo-gates.sh — all 23 gates pass, before and after the rebase onto master.
  That includes the knowledge-graph verifier, so the generated artefacts are current.
• shellcheck on scripts/agent-bus-latest.sh with no flags: zero findings. The lane only runs
  -S error, which is weaker.
• bash -n scripts/agent-bus-latest.sh parses.
• Every path cited in CLAUDE.md exists in the tree, and the committed blobs are LF with no CR bytes.
• The branch-protection claims in CLAUDE.md were read back from the API, not recalled:
  contexts = [cert-gate, build-core, shell-lint, lychee (README + docs), Readiness summary],
  strict = true, enforce_admins = true.

What this does not do

• It does not edit the bus issue's opening comment, which still carries the superseded rule 5. It is
  Grok's issue, and the protocol is that nobody edits an older message — so the repository documents say
  they supersede it instead.
• It adds no check that any of this stays true. The ownership statement is prose in four places and
  can drift apart again; nothing fails if CLAUDE.md and PROTOCOL.md disagree tomorrow.
• It changes no product code and no required workflow.

🤖 Generated with Claude Code