Submission preparation — v0.5.3 candidate

Source repository: https://github.com/echelong/cobalt-cockpit. Plugin folder: repository root. Published version remains 0.5.2 until release gates close. The final submission version must match both archived manifests. Name: Cobalt Cockpit. Identifier: cobalt-cockpit. License: MIT. Use the manifest description as the listing description.

The repository includes a self-hosted marketplace named cobalt-cockpit for GitHub installs. Its root-relative ./ source installs this plugin; function commands are registered by the hooks module, not Markdown commands. Seven bundled agents select their own model and reasoning level. There are no MCP servers, bundled credentials, runtime package launchers or extra permission grants. Runtime authentication presence checks, local processes, file reads and the opt-in session router are disclosed in README, SECURITY.md and PRIVACY.md.

Supported host: Claude Code 2.1.296. Official requirements reviewed on 2026-10-12: Software Directory Policy, mod documentation, manifest reference, pre-submission checklist and submit guide. Read the current compliance report and the copied historical portal findings before opening the portal. Runtime tests, local static checks and actual external validation are separate gates.

Use the developer portal to Re-validate, without submitting, the exact pushed candidate. The portal validates additional rules and name availability; local validation cannot certify acceptance. Preserve the complete actual report and source SHA. Directory submission itself needs separate authorization. A successful Validate is not the post-submission security scan or reviewer approval.

This is a Claude Code mod. Its HUD and function commands do not run in claude.ai or Cowork. Two synthesized WAV cues are bundled binary files, so the checklist indicates a manual-review hold: they are generated by assets/sounds/make-sounds.py, read (never written) at runtime, optional, and silent when no player answers. The listing icon is in place: a 1024×1024 PNG rendered from VECTOR's own HUD portrait by scripts/make-icon.py, named by the manifest's icon field, with no metadata and no stock art. It is regenerated, not designed, so a reviewer can reproduce it byte for byte.

The current checklist also confirms that non-image/font files at least 256 KiB require manual review. hooks/register.tsx exceeds that size; this is a disclosed hold, not grounds to remove functionality or split a security-critical file without need.

For an exact committed artifact, create and hash an out-of-tree Git archive and extract it. Run the following against that extraction, using the actual source SHA and manifest version:

python3 scripts/check-submission.py /tmp/cobalt-submission --source-sha <full-sha> --version <manifest-version> > /tmp/cobalt-submission-report.json
claude plugin validate --strict --json /tmp/cobalt-submission
claude plugin validate --strict /tmp/cobalt-submission/.claude-plugin/plugin.json
claude plugin validate --strict /tmp/cobalt-submission/.claude-plugin/marketplace.json
claude plugin validate --strict /tmp/cobalt-submission/agents
python3 /tmp/cobalt-submission/scripts/audit-public.py

The artifact checker runs from the source repository, compares every extracted file and executable mode with the commit, verifies runtime import and entrypoint paths, and records a complete hidden-file-inclusive sha-256 inventory. Do not add generated host declarations, local settings or evidence reports to the artifact. TypeScript is checked separately against real supported-host declarations.

hooks/hooks.json declares ./register.tsx relative to hooks/: the actual runtime entrypoint is hooks/register.tsx. The manifest's types and icon resolve from the plugin root. No portal codeFile record has been supplied; inspect any saved form metadata against the real shipped path before re-validation.

Outstanding owner steps: open the authenticated existing submission, select the exact candidate, Re-validate without submitting, and preserve all findings. Supply the original MOD_CAPABILITY_USE_NOT_PLAIN detail if it remains (rule text, SHA, file, line and suggested correction). Confirm verified support/contact details and any reviewer access; reviewers use their own supported Claude Code installation, since Cockpit has no separate service account. Present local replay/file reads, authentication-presence checks, the opt-in JEV router and manual-review assets for policy assessment. Resolve blockers and repeat affected gates before requesting separate submission authorization.

Three review use cases: inspect a task and its gates with /cockpit status; cancel Cockpit's question for git clean -fdx in an expendable scratch project and verify it never runs; enroll an owner-created failing verifier, enable ENFORCE, and verify the task remains held within its continuation budget. These require an interactive supported host and do not authorize destructive execution.