[stanleytejakusuma]

Bug

Default compression could silently clip historical tool results to about 2 KB even when a request had not opted into lossy compression. Disabling tool-result clipping alone was insufficient: Lite still removed trailing spaces and folded blank lines. The Responses WebSocket and offline eval callers also discarded the selected request-scoped pipeline and could execute the stored pipeline instead.

Fix

• Give safe/default Lite steps request-local settings that disable clipping and preserve unique tool-result whitespace, including RTK fallback, standalone Lite and retained stacked Lite.
• Pass the full selected plan through the Responses WebSocket and eval callers. Do not mutate shared or stored settings.
• Retain existing explicit lossy opt-in behavior, including whitespace normalization when compressToolResults is false; leave non-tool processing, caps, dependencies and session deduplication unchanged.
• Extract the existing per-case eval operation into one private helper to meet the standalone function-length gate. A before-and-after control pins call ordering, gold grading and metering, including exclusion of judge self-test costs.

Regression evidence and fresh verification

Actual failing-then-passing tests were captured during implementation: a 10,800-byte historical result was clipped to 1,937 bytes; after the clipping fix, a whitespace-sensitive 10,800-byte result still became 10,791 bytes. The whitespace regression set failed 74 of 79 tests before its fix and passed the same 79 afterward. Coverage includes 10,800 / 16,000 / 100,000-byte unique results, Chat/Responses layouts, sync/async engines, opt-in controls and unchanged stored settings.

Fresh results after rebasing onto 493101c3f1c82091e316ac3162a8511e17f43f7c:

Check	Result
All 12 touched native test files plus 3 adjacent Lite/eval files	213 passed, 0 failed
Vitest	56 files, 588 passed, 0 failed
Changed-file configured ESLint, Prettier and core typecheck	Passed
Standalone complexity/length findings over the same 6 production files	Base: 3; head: 2; no new length violation
Eval behavior control plus existing eval tests, before and after helper extraction	7 passed in each run

Commands used for focused and static verification:

BASE=493101c3f1c82091e316ac3162a8511e17f43f7c
mapfile -t TESTS < <(git diff --name-only "$BASE"...HEAD -- tests/)
mapfile -t CHANGED < <(git diff --name-only "$BASE"...HEAD -- '*.ts')
mapfile -t SOURCE < <(git diff --name-only "$BASE"...HEAD -- open-sse/ src/)
env -u DATA_DIR DISABLE_SQLITE_AUTO_BACKUP=true node \
  --import tsx/esm --import ./open-sse/utils/setupPolyfill.ts \
  --import ./tests/_setup/isolateDataDir.ts \
  --test --test-force-exit --test-concurrency=4 "${TESTS[@]}" \
  tests/unit/compression/eval-runner.test.ts \
  tests/unit/compression/lite.test.ts tests/unit/compression/lite-current-turn.test.ts
./node_modules/.bin/eslint --suppressions-location config/quality/eslint-suppressions.json "${CHANGED[@]}"
./node_modules/.bin/prettier --check "${CHANGED[@]}"
npm run typecheck:core
node_modules/.bin/eslint --no-config-lookup --config eslint.complexity.config.mjs --format json "${SOURCE[@]}"

The standalone complexity command still exits 1 for two existing cyclomatic violations; the measured finding count decreases from 3 to 2. The same config was also run against each base blob via --stdin --stdin-filename to compare identical paths without changing the checkout. Vitest ran with a separate temporary DATA_DIR and plugin directory, with SQLite automatic backups disabled.

Normal commit hooks ran without bypass or stashing. The private-reference UI-kit snapshot gate self-skipped in this fork environment. At initial draft publication, the full native unit suite, production build and coverage gate had not been rerun locally. Ready-for-review CI has since passed on 383217d4bd179f9cc676be3b2099a7f1c6f5cece (2026-11-03): all 8 unit-test shards, Vitest, Build, Coverage, Quality Ratchet, Lint, Security Tests and both aggregate gates.

• CI — success
• Quality Gates — success
• API Route Typecheck — success

The earlier draft-triggered aggregate failures are superseded by those successful runs. Unselected package-artifact and live E2E jobs remain skipped, not certified passes.

Known base status

⚠️ base-red inherited: #18791

The base tracker reports a package-artifact gate timeout and a skipped tarball boot-smoke. This PR does not attempt to fix those base failures.

[coderabbitai]

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

Run configuration

• Source mode: Auto
• Repository: cheaper-inference/OmniRoute
• Branch: release/v3.8.53
• File: .coderabbit.yaml
• Commit: f000c11
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Advanced
• Run ID: 05b01b2d-80c1-45e1-8f7c-45241484a36c

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• Trigger review

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts