Compass forks reversal — Open-Question resolutions (RIG-2336)

Status: Active

Amends design.md (the frozen forks-reversal record, PR #434, merged as 9d48f9a9). That record froze with its two load-bearing Open Questions unresolved — the pre-freeze ruling gate (skill://design § “No merge with open questions”) did not run before merge. A frozen record is never edited in place (skill://design: “there is no folding-in after merge — a later change ADDS a new record”), so the binding rulings are recorded here. The frozen record’s ## Open Questions §1 and §2 are superseded by this file and read as history; the L1/L2/L3 executors build against the decisions below.

Decision 1 — oh-my-pi: consume a Rigel-tagged npm release (supersedes OQ1)

Ruled (Matt, 2026-08-21): consume — not drop, and not a github: repo pin. Cut a Rigel-owned tagged npm release of the oh-my-pi coding-agent package and consume that from compass; repoint the credential-denylist generator to read the schema from the Rigel npm release rather than the vendored subtree.

This is a third shape distinct from both options the frozen record framed (its recommended “drop + repoint to the upstream-npm-installed schema” and its fallback “pin github:RigelBuild/oh-my-pi”): the schema source becomes a Rigel-controlled npm release, so the denylist tracks a release Rigel cuts, not upstream’s latest resolved version and not a raw git rev.

Impact on L3 (design.md:437-490):

• The generator repoint target (go/internal/store/gen_credential_keys.go:42 schemaRelPath) resolves to the Rigel npm package’s installed settings-schema.ts (pin the Rigel release in packages/compass-agent/package.json:19 in place of the current @oh-my-pi/pi-coding-agent pin), not the vendored forks/oh-my-pi/... path and not upstream npm.
• The hardcoded generated-header literal (gen_credential_keys.go:74, which emits credential_keys_gen.go:3) and the provenance at agent_config.go:656-659 + agent_config_test.go:505-518 update to cite the Rigel npm release, not forks/oh-my-pi/....
• Expect a denylist diff, re-review TestCredentialKeysMatchSchema’s want set (agent_config_test.go:508-516) in the same PR — the exact schema the Rigel release ships is what the denylist must track. Version-gap handling is unchanged from the frozen record: a diff is the expected, correct outcome, not drift to paper over.
• The deletion set (forks/oh-my-pi/ incl. moon.yml; the oh-my-pi-fork entry and its dedicated comment at .moon/workspace.yml:86-89) and the prose sweep from the frozen record’s L3 still apply unchanged. One line-range correction to the frozen record’s L3: it cites this entry as .moon/workspace.yml:80-83, but at the current tree that range is the tail of the shared VENDORING A NEW FORK comment block (:70-83, not oh-my-pi specific — it must NOT be deleted by L3; it belongs to L4’s shared-machinery teardown, design.md:500-503). The oh-my-pi-specific lines are the comment at :86-88 and the oh-my-pi-fork: 'forks/oh-my-pi' entry at :89. L3 deletes :86-89 only.
• Prerequisite: the Rigel npm release must exist and be published before L3 repoints. Coordinate the release cut with forge (shared RigelBuild fork convention) the same way L0a/L1 coordinate the devenv/nix2container repos.

Decision 2 — raw nix run sites: install the fork tools, pinned (supersedes OQ2)

Ruled (Matt, 2026-08-21): install the fork tools as pinned packages; do not keep raw nix run literals. The flake-input half of OQ2 was already settled (converge on orion’s lockfile-pinned github:RigelBuild/<fork> default). For the raw-CLI half — the six sites that invoke a fork CLI with nix run and so bypass devenv.lock — the resolution is to install the two fork tools (devenv’s patched CLI carrying compass’s container module, and skopeo-nix2container) from the same lockfile-pinned github:RigelBuild/<fork> input and invoke them by explicit name, rather than the frozen record’s recommended compass-side wrapper flake or the alternative literals-plus-CI-assert.

Why this is strictly cleaner than either framed option: the pin rev then lives in one place — the lockfile input, exactly like every other dependency — so there is nothing to drift, no wrapper flake to maintain, and no CI rev-equality assert to keep. The “six raw sites bypass the lockfile” problem dissolves because no raw nix run path:... / github:.../<rev> literal remains.

The reason the sites use nix run today is a PATH collision, not a hard requirement: the dev shell is itself booted by an upstream devenv, so a bare devenv container build would run upstream’s modules; nix run path:../forks/devenv#devenv was the escape hatch to force the fork’s modules (agent-image/devenv.yaml comment). Installing the fork CLI explicitly (a distinctly-named binary from the pinned input) removes the collision without a raw invocation.

Impact on L1 (design.md:348-386) and L2 (design.md:388-435):

• The six raw-nix run sites — agent-image/publish.sh:32,50-51, agent-image/moon.yml:44, devenv.nix:461, .github/workflows/ci.yml:812, .github/workflows/publish-agent-image.yml:139,160, tools/agent-image-env-gate/index.ts:100,118 — install and invoke the fork tools by explicit name from the github:RigelBuild/{devenv,nix2container} input pinned in agent-image/devenv.lock; no path:... or /<rev> literal remains at any of them.
• The frozen L1/L2 gate lines “the OQ2 pin shape applied (wrapper flake, or a lock-vs-URL rev-consistency assert)” resolve to: no raw literal present; the installed tool resolves the lockfile-pinned rev; single source of truth is agent-image/devenv.lock.
• The executor validates the devenv-CLI-vs-modules mechanics (how to force the fork’s container module through the installed CLI) at implementation time — that is an execution detail, not a further design fork.
• Everything else in L1/L2 (the devenv.yaml/devenv.lock repoints, the deletion sets, the mirrored env-gate globs, the comment sweeps) is unchanged from the frozen record.

Not changed

• OQ3 (secret-scanning re-widening) was non-load-bearing; the merge ratified its deferral. Unchanged — handled in L4 as the frozen record describes.
• L0a / L0b / L4 carry no OQ dependency and are unchanged from the frozen record. L0b is already satisfied (RigelBuild/nix2container main = 8f4a6fd7).