fix(openclaw): make the rewrite plugin work end-to-end on OpenClaw 2026.6.x

[davidbordenwi]

Summary

The OpenClaw plugin as shipped never rewrites anything on current OpenClaw (tested against 2026.6.5). This PR is the complete working set — four independent defects fixed together, verified end-to-end against live gateways. Fixes 1–3 overlap with open PRs #2202/#2220 (exit code 3), #2113 (TS build), and #1236/#1455 (extensions metadata) — credit to those authors; any one of them alone still leaves the plugin broken. Fix 4 is, as far as I can tell, not reported anywhere yet and is required even with 1–3 applied.

Closes #1810
Closes #944
Closes #1719

The four defects

1. Exit-code protocol — rtk rewrite signals a rewrite with exit code 3 (rewrite/ask) and the rewrite on stdout (per hooks/claude/rtk-rewrite.sh: 0+stdout allow-rewrite, 1 no-equivalent, 2 deny, 3+stdout ask-rewrite). execFileSync throws on non-zero exit, so tryRewrite returned null for every ask-classified rewrite — i.e. nearly always. Fixed by recovering status 3 + stdout from the thrown error, and passing through on 1/2 so the host's native deny policy still sees the original command.
2. TypeScript entry — OpenClaw managed installs reject TS entry points ("compiled runtime output required"). Ship compiled index.js (types stripped, logic identical) with main pointing at it; index.ts stays as source.
3. Missing openclaw.extensions — required in package.json for managed installs to register the entry point.
4. Gateway startup activation intent (new) — OpenClaw 2026.6.x excludes hook-only installed plugins from the Gateway startup plugin plan unless the manifest declares activation intent. Without "activation": { "onCapabilities": ["hook"] }, the agent runtime's restricted plugin load silently skips the plugin: plugins inspect reports loaded and embedded CLI runs rewrite, but daemon-served agent turns never do. A daemon can even mask this intermittently — any unrestricted plugin-load event in the process (e.g. a subagent spawn) registers the hooks for the process lifetime, which makes the failure look nondeterministic across identically-configured gateways.

Verification

A/B against a throwaway local OpenClaw 2026.6.5 gateway: pre-fix, a daemon-served exec of ls -la <dir> returns raw output; post-fix the identical turn returns the rtk-compressed form, with the rewritten command recorded in rtk's gain history and the original command preserved in the session transcript (the host applies rewritten params at execution time only — worth knowing when verifying: the transcript is not evidence of failure).

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}