Talavine
The reviewer flagged an assignment we had not written.
Our source checked a rectangle with <= and >=. The review quoted a mangled version containing rect.right = innerWidth. We initially suspected a model error. The more useful question was: what had our router actually sent?
An HTML-cleaning regex was holding the scissors.
A reproduction without a model
The old shared router applied the pattern <[^>]+> to text prompts on its completion, streaming and typed-fallback paths. A less-than sign in one comparison and a greater-than sign later in the code looked like an HTML tag to that pattern.
This small Python reproduction uses the same pattern and horizontal-space cleanup:
import re
source = (
"if (rect.right <= 0 || rect.bottom <= 0 || "
"rect.left >= innerWidth || rect.top >= innerHeight) return null;"
)
cleaned = re.sub(r"<[^>]+>", " ", source)
print(re.sub(r"[^\S\n]+", " ", cleaned))
Output:
if (rect.right = innerWidth || rect.top >= innerHeight) return null;
The same cleaner also removed generic type arguments and literal HTML. Applied to the smoke-test source, it reproduced the two truncated loops and missing definitions reported by the panel. The original smoke source parsed successfully and passed its regression tests; the string transformation was the defect.
The reviewer was looking at a different problem
The panel selected OpenRouter's answer from two usable answer lanes. Two other lanes failed. The winning answer quoted malformed code matching the old cleaner's output; another answer made additional missing-check claims. Voting did not establish whether the original source was correct.
We did not retain the original provider-request bytes for that historical panel. We can show that the panel adapter used this router path, and that the old cleaner reproduces its distinctive corruption. The later before-and-after tests capture actual serialized requests. Those are different kinds of evidence; a reproduction cannot restore a missing historical network trace.
Fix the boundary, then test the boundary
The development patch preserves prepared prompt text verbatim. An adapter that explicitly owns an HTML document can still extract plain text. The generic router should not guess whether an angle bracket belongs to HTML, a comparison, a generic type, or code someone wants reviewed.
Five tests asserting exact provider input failed against the original implementation. After the patch, 17 focused router, streaming, typed-fallback and cache tests passed, alongside four prompt-boundary tests. The full selected portable routing lane passed 697 of 697, with zero failures or skips. The focused routing cases overlap that full lane; they are not an extra 17 independent trials.
The five prompt-preservation wire cases use the real router and OpenAI/Claude request serializers with in-memory HTTP handlers. They make no external inference calls. The portable lane excludes opt-in local-model and live-decision tests. This is prompt-transport evidence, not a model-accuracy benchmark.
For your own harness, a useful regression check is:
- Use synthetic prompts containing comparisons, generic types, fenced HTML, tabs and newlines.
- Capture the final provider request in a test HTTP handler, keeping authentication headers out of the fixture.
- Decode the request JSON and compare its prompt field with the original string exactly. A UI preview is insufficient.
- Exercise completion, streaming and typed-response fallback paths.
- Verify the requested work separately. A panel ranking and an unchanged prompt do not prove the resulting code works.
The patch is tested in development. Public-source and installer publication remain pending; existing downloads are not represented as containing this repair.
Talavine Code is the MIT-licensed coding harness where this surfaced. The current public source lets you inspect the old router boundary; its setup page explains the available source and downloads. If you try the workflow, keep the failed cases and check what actually crossed the provider boundary. If an inspectable harness is useful to you, a GitHub star helps other developers find the project.
Prepared with autonomous AI assistance for Scott Crossen, the maker of Talavine Code. Automated checks are reported as checks; no human line-by-line review or comparative quality advantage is claimed.