Harness coverage: the offline floor + ledger exist only for Claude Code — survey of 11 plugins and the three ways to close it

[davidcrowe]

The gap

Phase 1 (#29, shipped 0.16.0) gave the Claude Code hook three things when the gateway cannot see a call: the hardline floor, the new ask-level destructive floor, and a local ledger that flushes on connect. The gateway ships the same floor (gatewaystack-connect#1097, live) and the ingest endpoint (#1115, live).

Every other harness still allows those calls with a warning and records nothing. Per the harness-coverage doctrine, a capability that exists for one harness is owed to all of them.

Survey (2026-09-08)

Five different shapes, which is why this needs a design decision rather than eleven hand-ports:

Plugin	Governance entry point	Language
claude-code	bin/govern.mjs + bin/decide.mjs	mjs — has phase 1
codex	bin/govern.mjs (455 lines, v0.6.6, no decide.mjs)	mjs
antigravity, grok-build, muse-code	hook.mjs	mjs
dsh	index.js	js
fx	lib/bridge.js	js
opencode	src/index.ts	ts
openclaw	src/govern.ts	ts
pi, prime-agent	index.ts	ts
hermes	src/acp_hermes/__init__.py	python

Nine of these are JS/TS and could share a module. Hermes needs a Python port of the same rules. None of the eleven has a generator; every copy is manual today, which is how install.sh ended up shipping a pre-floor engine for a full release cycle (acp-install#33).

The decision this needs

Option A — a published shared module. Extract decide.mjs (classifier + both floors) and the ledger into @agenticcontrolplane/decide, depend on it from the nine JS/TS plugins, port once to Python for Hermes. Fixes drift permanently and makes the next floor a version bump. Costs: a publish pipeline per plugin, and the plugins stop being single-file drop-ins, which is a property several of them were designed around (govern.mjs says so in its header).

Option B — one generator, per-harness params. Keep single-file plugins, generate them from one source (the doctrine's stated preference). No runtime dependency, but a build step in eleven repos.

Option C — floors only, ledger later. Port just the two floors now (small, pure, no I/O) and leave the ledger to the harnesses that have a config directory. Cheapest correct step; leaves the on-ramp story Claude-only.

Recommendation: C now, A or B as the follow-up. The floors are the safety half and they are pure functions with a fixture suite that already exists; the ledger needs a per-harness home for ~/.acp equivalents and is worth doing once the flush data from #29 says the on-ramp matters.

Acceptance (whichever option)

• Same fixtures pass for every harness: the four floor rules, plus the negatives (prose in a heredoc, a grep pattern, an issue title, --force-with-lease, rm -rf ./build).
• No harness gains a floor that the gateway does not enforce, and none is stricter than the gateway.
• Real-sandbox install test per harness (standing rule).
• Codex first: it is the only sibling with a govern.mjs of the same shape, and the one with a live external user.

Related: #29 (phase 1), acp-install#33 (installer sync), gatewaystack-connect#1097 (gateway floor), #4 (stale embedded govern.mjs fallback, still v0.5.0 — separate).