fix(vision): remove sampled region coverage heuristic - #666
Conversation
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
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. Comment |
|
Exact-head validation for |
|
The baseline now records this open PR and its exact-head evidence. Documentation-only follow-up advanced the branch to |
…' into fix/no-region-coverage-heuristic
|
Revalidated concurrent exact head |
|
Revalidated concurrent documentation-only exact head |
| regions = () | ||
| full_image_region = len(regions) == 1 and regions[0] == ImageRegion(0.0, 0.0, 1.0, 1.0) | ||
| partial_regions = bool(regions) and not full_image_region and not regions_cover_image(regions) | ||
| has_subregions = bool(regions) and not full_image_region |
There was a problem hiding this comment.
📝 Info: Tiled regions now trigger an extra parent-image VISION call
With the coverage heuristic gone, has_subregions is true for any non-empty, non-full-image region set. Region sets that previously tiled the plane took the merge-only path with no parent call; they now always issue one extra vision_client.describe on the parent image. This is the ADR 0067 tradeoff, and the new test expects describe_calls == 3, but per-image VISION cost rises for tiled outputs.
Was this helpful? React with 👍 or 👎 to provide feedback.
…0' into fix/no-region-coverage-heuristic
|
Exact-head revalidation at |
f8c7895
into
feat/ontology-project-nodes-v2200
Contract
Removes the sampled
32 x 32region-coverage heuristic. ADR 0067 now makes the evidence boundary explicit: a locator response does not authoritatively prove whole-image coverage, so every proper subregion set retains one parent-image VISION request. A single explicit full-image box remains parent evidence rather than a fake decomposed region.This deletes local coverage arithmetic and introduces no replacement threshold, tolerance, weight, or fallback calculation.
Verification
uv run --extra dev --extra backend pytest -q— 1,468 passed, 17 skipped on the exact implementation tree before the frontend-only stack advancegit diff --checkThe two ontology test-double repairs align pre-existing stack tests with the current cutoff/snapshot keyword contract; they do not change production behavior.