The shared Open Code Review workflow cannot fully review a protocol-registry-sized PR at its pinned configuration (OCR 1.12.9, glm-5.2, --max-tokens-budget 1000000). On km43 #150 (range ba587f36..506a210) every measured configuration ended partial or failed with budget_exceeded: true, and CI still showed a green check.
Measurements are in the last comment on origin89hq/km43#151. Excluding the generated bindings did not help: crates/xtask/src/bindings.rs alone and protocol.toml plus five handwritten modules each crossed 1,000,000 tokens on their own. 82–90% of input tokens were cache reads, which suggests context is resent every tool round (inferred, not measured per round). Results varied between identical runs.
Expected: a PR of this size is reviewed with status: complete, no failed coverage, and no budget overrun. Actual: the largest source groups are never fully reviewed, so a green OCR job on km43 is not evidence of review.
Next step: decide in the shared workflow between a larger budget, a different model or effort setting, per-group budgets, or a documented size limit that asks authors to split PRs. Then rerun the #150 range fresh and non-incremental as the acceptance check, and update the km43 pin.
The shared Open Code Review workflow cannot fully review a protocol-registry-sized PR at its pinned configuration (OCR 1.12.9, glm-5.2,
--max-tokens-budget 1000000). On km43 #150 (rangeba587f36..506a210) every measured configuration endedpartialorfailedwithbudget_exceeded: true, and CI still showed a green check.Measurements are in the last comment on origin89hq/km43#151. Excluding the generated bindings did not help:
crates/xtask/src/bindings.rsalone andprotocol.tomlplus five handwritten modules each crossed 1,000,000 tokens on their own. 82–90% of input tokens were cache reads, which suggests context is resent every tool round (inferred, not measured per round). Results varied between identical runs.Expected: a PR of this size is reviewed with
status: complete, no failed coverage, and no budget overrun. Actual: the largest source groups are never fully reviewed, so a green OCR job on km43 is not evidence of review.Next step: decide in the shared workflow between a larger budget, a different model or effort setting, per-group budgets, or a documented size limit that asks authors to split PRs. Then rerun the #150 range fresh and non-incremental as the acceptance check, and update the km43 pin.