Both installers still skip a host whose registration cannot run, and exit 0

[MongLong0214]

SSOT execution contract

Outcome

Installer exit 0 means every requested host is live-verified healthy or explicitly preserved as a healthy custom CommitLore registration.

Scope

• one shared TypeScript host-inspection/apply command used by shell and PowerShell
• exact owned wrapper + mcp args
• live initialize/serverInfo probe for custom registrations
• atomic JSON/TOML write
• non-zero on any requested host failure

Acceptance

• dead command, directory command, wrong args, malformed config, failed host CLI, interrupted write: non-zero
• healthy owned/custom registration: preserved and reported distinctly
• shell and PowerShell return the same summary schema and exit meaning

Non-goals

Do not rewrite runtime checkout/manifest activation and do not add hosts.

---

The pattern, still here after v0.8.1

An existing MCP entry is treated as working when its command merely exists as a string. The repository's own ownership predicate requires this wrapper and exactly mcp; the installers do not apply it to what they skip.

1. A config that cannot run an MCP server is skipped as "already mentions commitlore" — install.sh:718, install.ps1:851. A Cursor entry with command: "/bin/sh", args: ["mcp"], or a directory, or a non-executable file. The installer exits 0. Shared by both.

2. Codex's "already points at this install" is a substring test — install.sh:836, install.ps1:987. A codex mcp get result reading command: <dest>-old, or command: <dest> with wrong or absent mcp args, passes. When the CLI is absent, the fallback checks only for the TOML table header — not its command, not its args (install.sh:864, install.ps1:1020). Shared.

3. A host setup failure is a printed line, then overwritten by success — install.sh:650, install.ps1:788, 1157. A preserved dead target, a failed codex mcp add, malformed config, a failed plugin install, or a Hermes failure leaves that host broken and the process exits 0. Shared; PowerShell makes it explicit with exit 0.

4. Presence is inferred from directories and free text — install.sh:804, 904, install.ps1:935, 815. An orphan .cursor/.gemini directory triggers config creation with no executable present. Any claude plugin list output containing commitlore suppresses installation — including an unrelated plugin whose name contains it. Shared.

5. Wrapper ownership is still inferred from forgeable evidence — install.sh:457, 460, install.ps1:560, 565. A foreign wrapper containing the marker is overwritten. An unmarked foreign executable that prints 0.8.1 is overwritten whenever a matching checkout exists — which proves the checkout exists, not that it wrote the wrapper. Shared.

Divergences between the two installers

install.sh emits a broken wrapper for a legal POSIX path containing " — install.sh:608. Set HOME or XDG_DATA_HOME to such a path: checkout and smoke verification succeed, the generated exec … "$candidate/dist/…" has unescaped quotes and points somewhere else, and the installer reports success. Fresh JSON configs are unescaped the same way (install.sh:749). PowerShell does not share the practical case — Windows paths cannot contain " — and its JSON writer serializes values properly.

Shell's JSON detection is recursive and jq-dependent — install.sh:698, 722. With jq, {"metadata": {"commitlore": {}}} anywhere in the file suppresses wiring. Without jq, command extraction is disabled entirely, so a real entry pointing at a deleted /tmp/…/commitlore gets only "already mentions". PowerShell limits detection to mcpServers/mcp/servers and does not share either gap.

PowerShell can make an interrupted write permanently look registered — install.ps1:892, 923, 832. Set-Content writes in place; killed after "commitlore": but before the JSON closes, the next run's parse failure falls to a regex that says it is registered, and exits 0.

Smoke-exit classification differs — shell treats 126/127 as verification-unavailable; PowerShell's version and validate paths treat them as unusable.

What is genuinely clean

Both enforce Node 22.13+, run git before any write, bind the checkout to the requested tag's HEAD, verify tracked runtime assets, smoke-test the runtime before activation, retain old checkouts on upgrade, and write new MCP entries using the wrapper plus mcp. Same-version reruns reverify and reactivate; upgrades add a pinned checkout and repoint the stable wrapper. No persistent write was found outside the checkout, the wrapper, supported host config locations, and temp directories. The risk is preserved pre-existing entries, not newly generated wrong commands.

What closes it

Apply the repository's own ownership predicate to what is skipped, not only to what is written: the command must be this wrapper and the args must be exactly mcp. A host that cannot be made to work should make the installer exit non-zero, or at minimum be reported in a summary the caller can act on — CI and configuration management read the exit code, and today it says everything worked.

Follows #541, which fixed the same class for the runtime by requiring it to answer. Host wiring never got that treatment.

Found by an audit of both installers.