fix(plugin): restore the no-../src deployment contract — lazy widget-helper import (#995)

[aarontrowbridge]

Important

Problem — The opencode plugin that serves the entire amicode_* tool pack statically imported the widget-authoring helper from the extension service source tree. The plugin directory's own deployment contract (carried as comments by calib_chain, regime_priors, rehearsal) is "never anything from ../src/": trimmed deployments — the hub service bundle — ship only the plugin directory. The static import therefore killed module init on the hub, silently (the service runner captures engine stderr in memory and emits it only on health failure), and every amicode_* tool vanished from hub-served sessions with this morning's 04:31 bundle deployment — including amicode_session, the "spool a fresh session" verb that surfaced this.

Approach — Replace the static import with a lazy, fail-soft dynamic import inside the widget tool's execute: module init survives everywhere; where the helper is absent, that ONE tool returns an honest refusal (nothing written — the engine's built-in widget tool and the MCP transport carry authoring on those surfaces) and the rest of the pack stays registered. A floor test now scans the plugin directory for static parent-relative imports, so the contract is enforced by the suite instead of comments. Tool names/descriptions/args are untouched (core/plugin parity stays byte-pinned).

Acceptance Criteria

• Zero static parent-relative imports across the plugin dir — floor test in the suite
• The widget tool still authors widgets end-to-end where the helper is present (the MCP floor: carry the widget-authoring tool (fork-registry delta) #799 e2e now exercises the lazy path — green)
• Targeted suites green locally (tools core + parity, MCP round-trip, session spawn, warrant parity: 50/50)
• CI green

Verification honesty

pnpm --filter amicode test locally on this box carries pre-existing red in service/auth/overlay-seam tests (they bind ports that collide with the live hub on this machine) — identical failures on a clean main worktree, and CI's fast lane passes for a clean tree (PR #990's current run is all-green), so CI is the arbiter.

Notes

Two follow-up seams discovered during diagnosis, deliberately out of scope: (1) the service runner swallows engine stderr on healthy boots (engineLog is in-memory, emitted only on health failure) — that silence is why a whole tool pack could vanish unnoticed; (2) the MCP transport's session tool refuses by design (no in-process engine client) — an HTTP fallback would make the spawn verb work on every surface. Also observed: the fast suite mutates packages/app-bundle/overlay/** in the working tree as a side effect of running it (reproduced in a fresh worktree) — worth its own issue.

Closes #995

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}