Repository navigation
fix(compression): match legacy output text in hu, ja and zh - #15591
Merged
diegosouzapw merged 3 commits intoOct 6, 2026
Merged
diegosouzapw merged 3 commits into
diegosouzapw merged 3 commits into
Conversation
Legacy cavemanOutputMode configs reach upstream through applyOutputStyles, which maps them to the terse-prose style. Its text matched the old caveman injector in 9 of the 12 legacy languages. Hungarian got the English text, because the terse-prose i18n map had no hu entry. Japanese and Chinese got one extra space before the shared boundary sentence: the injector rejoined the stripped instruction and the boundary block with a fixed space, while every ja and zh text in the catalog runs straight into that sentence. - catalog.ts: terse-prose takes CAVEMAN_INSTRUCTION_BY_LANGUAGE.hu. - apply.ts: the boundary block follows the instruction body with the spacing the last style's own text puts before SHARED_BOUNDARIES. This also removes the space from multi-style ja and zh selections and from terse-cjk. - The catalog.ts, apply.ts and backCompat.ts comments now say the text below the marker line matches the legacy injection. The marker line itself still differs. Tests: - output-styles-legacy-parity.test.ts (new) runs every language in CAVEMAN_INSTRUCTION_BY_LANGUAGE at lite, full and ultra through both injectors and compares the system message text below each marker. On the unfixed code, 9 of its 36 cases fail (hu, ja and zh). - output-styles-boundaries.test.ts checks the character before the boundary sentence in multi-style and terse-cjk selections: none for ja and zh, one space for hu, where the last style falls back to its English text. - output-styles-i18n-matrix.test.ts adds hu to the terse-prose baseline.
…s with behavior - catalog.ts derives terse-prose levels/i18n from CAVEMAN_INSTRUCTION_BY_LANGUAGE by destructuring instead of hand-copying 11 keys, so a language added to the legacy table cannot be missed the way hu was. - apply.ts comments state the separator rule as the invariant it implements (any whitespace before SHARED_BOUNDARIES collapses to one space; a non-whitespace character suppresses it) instead of enumerating ja and zh. - The legacy-parity helper reads the instruction from every text surface (top-level system field or trailing system message), so a placement change cannot fake a failure. - The multi-style spacing matrix gains an ultra-level case; a stale regression-guard comment now points at the parity suite.
3 of 5 tasks
diegosouzapw
merged commit Oct 6, 2026
44efe61
into
diegosouzapw:release/v3.8.52
44 of 51 checks passed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What broke
Legacy
cavemanOutputModeconfigs (the pre-styles compression toggle) reach the request pipeline through the back-compat shim, which maps them to theterse-proseoutput style. Its injected text matched the old caveman injector in 9 of the 12 legacy languages. Two languages did not:huentry, so the style fell back toCAVEMAN_INSTRUCTION_BY_LANGUAGE.en.The fix
catalog.ts: terse-prose gainsCAVEMAN_INSTRUCTION_BY_LANGUAGE.hu, and the i18n map is now derived from the legacy table by destructuring, so a language added there cannot be missed here again.apply.ts: the boundary block follows the instruction body with the spacing the last style's own text puts beforeSHARED_BOUNDARIES: one space when that text has whitespace (or nothing) before the marker, none when a non-whitespace character directly precedes it. This also removes the extra space from multi-style ja/zh selections and fromterse-cjk.For existing ja/zh and terse-cjk selections the injected text loses that space, so prompt-cache prefixes for those tuples invalidate once on upgrade. The byte change is the intended fix.
Tests
tests/unit/compression/output-styles-legacy-parity.test.ts(new): every language inCAVEMAN_INSTRUCTION_BY_LANGUAGEat lite/full/ultra through both injectors, comparing the system message text below each marker line. On the unfixed code, 9 of its 36 cases fail (hu, ja, zh); all pass on this branch.tests/unit/compression/output-styles-boundaries.test.ts: pins the character before the boundary sentence for multi-style and terse-cjk selections (none for ja/zh, one space for hu via the English fallback), now at full and ultra levels.tests/unit/compression/output-styles-i18n-matrix.test.ts: adds hu to the terse-prose baseline.Verification
npm run typecheck:core: cleanNotes for reviewers
docs/compression/COMPRESSION_GUIDE.mdstill describes the old behavior (extra space, English fallback for hu). That file is owned by docs(compression): match caveman output mode notes to code #15297, so it is not touched here. Whichever PR lands second fixes the wording; also noted on docs(compression): match caveman output mode notes to code #15297.