mirror of
https://github.com/rtk-ai/rtk.git
synced 2026-09-19 07:33:17 +08:00
kylehgc
f9b107e0df
mvnd builds multi-module reactors in parallel and prefixes per-module log lines with `[module] `, interleaved line-by-line, while stack traces stay raw. The Surefire matchers keyed on `[INFO]`/`[ERROR]` at column 0, so in a parallel reactor a failing class lost its diagnostics: only the reactor summary and the generic goal error survived. - `split_lane` classifies on the prefix-stripped view and emits the original line, so module identity survives in the output. - One `SurefireLane` (block machine + keep-continuation + failures-summary cap) per module, so a passing close from one module can no longer be attributed to another module's open block. Unprefixed raw lines route to the lane that most recently opened a block or failure trail. - `filter_compile` classifies on the stripped view too. Adds a real captured fixture (`mvnd clean test` on the new multi-module-fail-skeleton, exit 1, interleaved reactor): the failing class, its assertion message, the user stack frame, and the summary entry all survive at 72.1% token savings. Adds insta snapshot regression tests locking the full filtered output of all four mvnd fixtures, per docs/contributing/CODING_PRACTICES.md. Updates the jvm README: the parallel-reactor limit it documented is fixed; only the `-q` daemon-chatter limit remains. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>