Add codex, antigravity, and opencode enforcement-gate ports

[minhtrinh-imbue]

Summary

Adds enforcement-gate ports of imbue-code-guardian for codex, antigravity, and opencode, each via that harness's own idiomatic distribution channel:

• codex: real curated plugin marketplace — codex plugin marketplace add imbue-ai/code-guardian + codex plugin install imbue-code-guardian-codex@imbue-code-guardian. Confirmed via developers.openai.com/codex/plugins and /codex/plugins/build.
• antigravity: agy plugin install/agy plugin link. Plugin structure (plugin.json + root hooks.json) validated against a real local agy install (agy plugin validate) — the manifest shape is confirmed correct, but the exact end-to-end marketplace-publish flow is not, since the docs site didn't render usable content for this. Flagged clearly in that plugin's README for a maintainer with real publish access to confirm/fix.
• opencode: npm, referenced via the plugin config array — no marketplace exists for opencode, so npm is the actual idiomatic channel per opencode's own docs.

Scope — please read before reviewing

This is the enforcement gate only, not the reviews themselves. Each new plugin's Stop-equivalent hook correctly detects "the required reviews (autofix/architecture/conversation) haven't run yet" and blocks — using the exact same config_utils.sh/stop_hook_gates.sh as the claude plugin, copied verbatim (both were already harness-agnostic: pure bash, marker-file checks, no Claude-specific paths).

What's not included: the actual /autofix, /verify-conversation, /verify-architecture skills. Those spawn sub-agents by name via Claude's own Agent/Task tool and use Claude Code's !`command` inline-bash-in-frontmatter execution — genuinely Claude-Code-specific control flow, not portable prose like the agents/*.md review prompts (which are just plain-English instructions + YAML frontmatter, already agent-agnostic). Porting those means rewriting that orchestration in each harness's own subagent-spawning idiom — a separate, larger follow-up. Until that exists, a blocked codex/antigravity/opencode agent will correctly detect missing reviews but the block message points at claude-only slash commands.

What's verified vs. not

• All bash scripts: bash -n syntax-checked; enabled-by-default gating logic tested standalone (disabled-by-default no-op, and the enabled-with-no-diff skip path).
• All JSON manifests: schema-validated. antigravity's specifically validated against the real agy plugin validate CLI (iterated on its actual error messages to find the true schema — plugin.json + hooks.json both at plugin root, hooks.json unwrapped/bare — rather than guessing from docs).
• opencode's index.ts: valid TypeScript (esbuild, 0 errors); import.meta.dir (used to resolve bundled scripts relative to the installed package rather than the user's project) confirmed as a real, working Bun API.
• Not exercised: none of the three has been run against a live session actually hitting the Stop/idle event end-to-end. Static validation + standalone script testing only.

Each new plugin's own README has a "Known gaps / unverified" section with more detail.

Test plan

• A maintainer with real codex/antigravity/opencode installs verifies each plugin actually installs via the documented command and fires on a real session.
• Confirm antigravity's exact marketplace-registration flow (this PR's README flags this as the least-confirmed piece).
• Decide on and scope the follow-up for porting the actual review skills.

🤖 Generated with Claude Code