fix(omp): resolve host modules and tokenizer from plugin runtime

[randomvariable]

Refs #657

Fix compiled OMP host-module imports and tokenizer runtime fallback resolution. Preserve package-local tokenizer precedence.

Independent of PR #656. Initial standalone verification: typecheck, build, 29 focused tests. Review regression improvements are in progress.

Confidence Score: 5/5

The PR appears safe to merge, with no new actionable issues found.

Summary

The PR keeps host-package imports available for OMP to resolve at runtime and adds tokenizer lookup paths for plugin installs.

• OMP can resolve its session API from either host package.
• Token counts can find the plugin’s tokenizer under compiled hosts.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A["preloadTokenizer()"] --> B["Load relative to import.meta.url"]
  B -->|Succeeds| C["Use tokenizer"]
  B -->|Fails| D["Probe project and OpenCode cache"]
  D --> E["Probe argv and module ancestors"]
  E --> F["Probe OMP plugin tree"]
  F --> G["Import first matching package"]
  G -->|Succeeds| C
  G -->|Fails| H["Use approximate counts"]

Loading

_{Reviews (2) · Last reviewed commit: "fix(tokenizer): prove fallback loading a..." · Reviewed by Greptile}

[cortexkit-ci]

Design gate skipped: the trivial label is applied.

[cubic-dev-ai]

All reported issues were addressed across 6 files

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Re-trigger cubic</sub>

[magic-alfonso]

Thanks for this, and for splitting it from the discovery change.

Before review, please address the open bot finding. Greptile notes that the new tokenizer-roots test may not exercise the fallback: preloadTokenizer() first calls loadTokenizer(), which resolves from import.meta.url regardless of cwd or process.argv[1]. So the test can pass even if the new probe roots are never used. A test that fails without your new roots (for example with the module-relative lookup made to miss) would show the fix does what it says.

Once the finding is resolved, we'll take it through review.

[randomvariable]

Addressed all three threads in 3b046f4. The runtime fixture forces the primary loader to fail under bun --no-install and proves fallback estimator identity. A second fixture proves the plugin dependency beats a conflicting host-wide copy. Three focused tests, plugin typecheck, and Pi build passed. Negative control with the old probe ordering fails the precedence regression. Refs #657 for maintainer design approval.

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.

[magic-alfonso]

Thanks for resolving the findings. I've labelled this trivial and ran it through our merge checks against current master. Two of the new tests fail there, because of a change that landed on master after your branch point, not a problem with your fix.

read-session-formatting.ts now imports a new sibling module, ./token-count-exact (added today for issue 653). Your tokenizer-roots tests copy read-session-formatting.ts into a temporary module tree, so the copy can't resolve that import:

ResolveMessage: Cannot find module './token-count-exact' imported from …/mod/src/hooks/magic-context/read-session-formatting.ts

Affected: "the fallback loads from the plugin tree when the primary loader fails" and "the plugin's own tree outranks the host-wide ~/.omp/plugins copy". Could you rebase onto master and have the fixture copy token-count-exact.ts too, or better, every relative import of the copied file, so the next new import doesn't break it? Once that's green, we'll merge it.