fix(runner): capture parallel worker stderr instead of discarding it by Chemaclass · Pull Request #891 · TypedDevs/bashunit

Chemaclass · 2026-07-27T08:47:53Z

🤔 Background

Related #864

Each parallel file worker was spawned as call_test_functions … 2>/dev/null & (added in #358 to keep worker noise off the progress line), so the same run could report differently depending on --parallel, with nothing indicating output had been suppressed.

💡 Changes

Worker stderr is captured per file under the run output dir and rendered after aggregation as a Stderr from <file> block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Cleanup rides the existing run-dir EXIT trap, so no per-file rm fork is added.
The capture deliberately does not live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and a stray file makes it report "No tests found".
Scope correction on the issue: a test body's stderr was never lost. execute_test_body runs the function as "$fn_name" "$@" 2>&1, so it is merged into the captured stdout and already appeared in that test's failure block. What the redirect dropped is stderr with no owning test, which is why the new block is attributed to the file rather than to a test.
The regression test compares occurrence counts between the two modes instead of grepping for the message — the fixture's provider also runs in the main-shell counting pass, so a bare assert_contains passes against the unfixed code.

Each file worker was spawned as `call_test_functions … 2>/dev/null &` since #358, to keep worker noise off the progress line. That also made the same run report differently depending on --parallel. The worker's stderr is now redirected to a per-file capture under the run output dir and rendered after aggregation as a `Stderr from <file>` block. The capture cannot live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and would report "No tests found" for a stray file. Cleanup is the existing run-dir EXIT trap, so no per-file `rm` fork is added. Scope note: a test *body*'s stderr was never lost. execute_test_body runs the function as `"$fn_name" "$@" 2>&1`, so it is merged into the captured stdout and already surfaced in that test's failure block. What the redirect dropped is stderr with no owning test — data providers, hook plumbing, scratch-dir errors — which is why the new block is attributed to the file rather than to a test. The regression test compares occurrence counts between modes rather than just grepping for the message: the fixture's provider also runs in the main-shell counting pass, so the diagnostic leaks in either way and a bare assert_contains passes against the unfixed code. Closes#864

Chemaclass · 2026-07-27T08:47:53Z

🤔 Background

Related #864

Each parallel file worker was spawned as call_test_functions … 2>/dev/null & (added in #358 to keep worker noise off the progress line), so the same run could report differently depending on --parallel, with nothing indicating output had been suppressed.

💡 Changes

Worker stderr is captured per file under the run output dir and rendered after aggregation as a Stderr from <file> block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Cleanup rides the existing run-dir EXIT trap, so no per-file rm fork is added.
The capture deliberately does not live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and a stray file makes it report "No tests found".
Scope correction on the issue: a test body's stderr was never lost. execute_test_body runs the function as "$fn_name" "$@" 2>&1, so it is merged into the captured stdout and already appeared in that test's failure block. What the redirect dropped is stderr with no owning test, which is why the new block is attributed to the file rather than to a test.
The regression test compares occurrence counts between the two modes instead of grepping for the message — the fixture's provider also runs in the main-shell counting pass, so a bare assert_contains passes against the unfixed code.

Each file worker was spawned as `call_test_functions … 2>/dev/null &` since #358, to keep worker noise off the progress line. That also made the same run report differently depending on --parallel. The worker's stderr is now redirected to a per-file capture under the run output dir and rendered after aggregation as a `Stderr from <file>` block. The capture cannot live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and would report "No tests found" for a stray file. Cleanup is the existing run-dir EXIT trap, so no per-file `rm` fork is added. Scope note: a test *body*'s stderr was never lost. execute_test_body runs the function as `"$fn_name" "$@" 2>&1`, so it is merged into the captured stdout and already surfaced in that test's failure block. What the redirect dropped is stderr with no owning test — data providers, hook plumbing, scratch-dir errors — which is why the new block is attributed to the file rather than to a test. The regression test compares occurrence counts between modes rather than just grepping for the message: the fixture's provider also runs in the main-shell counting pass, so the diagnostic leaks in either way and a bare assert_contains passes against the unfixed code. Closes#864

Chemaclass · 2026-07-27T08:47:53Z

🤔 Background

Related #864

Each parallel file worker was spawned as call_test_functions … 2>/dev/null & (added in #358 to keep worker noise off the progress line), so the same run could report differently depending on --parallel, with nothing indicating output had been suppressed.

💡 Changes

Worker stderr is captured per file under the run output dir and rendered after aggregation as a Stderr from <file> block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Cleanup rides the existing run-dir EXIT trap, so no per-file rm fork is added.
The capture deliberately does not live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and a stray file makes it report "No tests found".
Scope correction on the issue: a test body's stderr was never lost. execute_test_body runs the function as "$fn_name" "$@" 2>&1, so it is merged into the captured stdout and already appeared in that test's failure block. What the redirect dropped is stderr with no owning test, which is why the new block is attributed to the file rather than to a test.
The regression test compares occurrence counts between the two modes instead of grepping for the message — the fixture's provider also runs in the main-shell counting pass, so a bare assert_contains passes against the unfixed code.

Each file worker was spawned as `call_test_functions … 2>/dev/null &` since #358, to keep worker noise off the progress line. That also made the same run report differently depending on --parallel. The worker's stderr is now redirected to a per-file capture under the run output dir and rendered after aggregation as a `Stderr from <file>` block. The capture cannot live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and would report "No tests found" for a stray file. Cleanup is the existing run-dir EXIT trap, so no per-file `rm` fork is added. Scope note: a test *body*'s stderr was never lost. execute_test_body runs the function as `"$fn_name" "$@" 2>&1`, so it is merged into the captured stdout and already surfaced in that test's failure block. What the redirect dropped is stderr with no owning test — data providers, hook plumbing, scratch-dir errors — which is why the new block is attributed to the file rather than to a test. The regression test compares occurrence counts between modes rather than just grepping for the message: the fixture's provider also runs in the main-shell counting pass, so the diagnostic leaks in either way and a bare assert_contains passes against the unfixed code. Closes#864

Chemaclass · 2026-07-27T08:47:53Z

🤔 Background

Related #864

Each parallel file worker was spawned as call_test_functions … 2>/dev/null & (added in #358 to keep worker noise off the progress line), so the same run could report differently depending on --parallel, with nothing indicating output had been suppressed.

💡 Changes

Worker stderr is captured per file under the run output dir and rendered after aggregation as a Stderr from <file> block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Cleanup rides the existing run-dir EXIT trap, so no per-file rm fork is added.
The capture deliberately does not live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and a stray file makes it report "No tests found".
Scope correction on the issue: a test body's stderr was never lost. execute_test_body runs the function as "$fn_name" "$@" 2>&1, so it is merged into the captured stdout and already appeared in that test's failure block. What the redirect dropped is stderr with no owning test, which is why the new block is attributed to the file rather than to a test.
The regression test compares occurrence counts between the two modes instead of grepping for the message — the fixture's provider also runs in the main-shell counting pass, so a bare assert_contains passes against the unfixed code.

Each file worker was spawned as `call_test_functions … 2>/dev/null &` since #358, to keep worker noise off the progress line. That also made the same run report differently depending on --parallel. The worker's stderr is now redirected to a per-file capture under the run output dir and rendered after aggregation as a `Stderr from <file>` block. The capture cannot live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and would report "No tests found" for a stray file. Cleanup is the existing run-dir EXIT trap, so no per-file `rm` fork is added. Scope note: a test *body*'s stderr was never lost. execute_test_body runs the function as `"$fn_name" "$@" 2>&1`, so it is merged into the captured stdout and already surfaced in that test's failure block. What the redirect dropped is stderr with no owning test — data providers, hook plumbing, scratch-dir errors — which is why the new block is attributed to the file rather than to a test. The regression test compares occurrence counts between modes rather than just grepping for the message: the fixture's provider also runs in the main-shell counting pass, so the diagnostic leaks in either way and a bare assert_contains passes against the unfixed code. Closes#864

Chemaclass · 2026-07-27T08:47:53Z

🤔 Background

Related #864

Each parallel file worker was spawned as call_test_functions … 2>/dev/null & (added in #358 to keep worker noise off the progress line), so the same run could report differently depending on --parallel, with nothing indicating output had been suppressed.

💡 Changes

Worker stderr is captured per file under the run output dir and rendered after aggregation as a Stderr from <file> block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Cleanup rides the existing run-dir EXIT trap, so no per-file rm fork is added.
The capture deliberately does not live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and a stray file makes it report "No tests found".
Scope correction on the issue: a test body's stderr was never lost. execute_test_body runs the function as "$fn_name" "$@" 2>&1, so it is merged into the captured stdout and already appeared in that test's failure block. What the redirect dropped is stderr with no owning test, which is why the new block is attributed to the file rather than to a test.
The regression test compares occurrence counts between the two modes instead of grepping for the message — the fixture's provider also runs in the main-shell counting pass, so a bare assert_contains passes against the unfixed code.

Each file worker was spawned as `call_test_functions … 2>/dev/null &` since #358, to keep worker noise off the progress line. That also made the same run report differently depending on --parallel. The worker's stderr is now redirected to a per-file capture under the run output dir and rendered after aggregation as a `Stderr from <file>` block. The capture cannot live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and would report "No tests found" for a stray file. Cleanup is the existing run-dir EXIT trap, so no per-file `rm` fork is added. Scope note: a test *body*'s stderr was never lost. execute_test_body runs the function as `"$fn_name" "$@" 2>&1`, so it is merged into the captured stdout and already surfaced in that test's failure block. What the redirect dropped is stderr with no owning test — data providers, hook plumbing, scratch-dir errors — which is why the new block is attributed to the file rather than to a test. The regression test compares occurrence counts between modes rather than just grepping for the message: the fixture's provider also runs in the main-shell counting pass, so the diagnostic leaks in either way and a bare assert_contains passes against the unfixed code. Closes#864

Chemaclass · 2026-07-27T08:47:53Z

🤔 Background

Related #864

Each parallel file worker was spawned as call_test_functions … 2>/dev/null & (added in #358 to keep worker noise off the progress line), so the same run could report differently depending on --parallel, with nothing indicating output had been suppressed.

💡 Changes

Worker stderr is captured per file under the run output dir and rendered after aggregation as a Stderr from <file> block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Cleanup rides the existing run-dir EXIT trap, so no per-file rm fork is added.
The capture deliberately does not live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and a stray file makes it report "No tests found".
Scope correction on the issue: a test body's stderr was never lost. execute_test_body runs the function as "$fn_name" "$@" 2>&1, so it is merged into the captured stdout and already appeared in that test's failure block. What the redirect dropped is stderr with no owning test, which is why the new block is attributed to the file rather than to a test.
The regression test compares occurrence counts between the two modes instead of grepping for the message — the fixture's provider also runs in the main-shell counting pass, so a bare assert_contains passes against the unfixed code.

Each file worker was spawned as `call_test_functions … 2>/dev/null &` since #358, to keep worker noise off the progress line. That also made the same run report differently depending on --parallel. The worker's stderr is now redirected to a per-file capture under the run output dir and rendered after aggregation as a `Stderr from <file>` block. The capture cannot live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and would report "No tests found" for a stray file. Cleanup is the existing run-dir EXIT trap, so no per-file `rm` fork is added. Scope note: a test *body*'s stderr was never lost. execute_test_body runs the function as `"$fn_name" "$@" 2>&1`, so it is merged into the captured stdout and already surfaced in that test's failure block. What the redirect dropped is stderr with no owning test — data providers, hook plumbing, scratch-dir errors — which is why the new block is attributed to the file rather than to a test. The regression test compares occurrence counts between modes rather than just grepping for the message: the fixture's provider also runs in the main-shell counting pass, so the diagnostic leaks in either way and a bare assert_contains passes against the unfixed code. Closes#864

Chemaclass · 2026-07-27T08:47:53Z

🤔 Background

Related #864

Each parallel file worker was spawned as call_test_functions … 2>/dev/null & (added in #358 to keep worker noise off the progress line), so the same run could report differently depending on --parallel, with nothing indicating output had been suppressed.

💡 Changes

Worker stderr is captured per file under the run output dir and rendered after aggregation as a Stderr from <file> block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Cleanup rides the existing run-dir EXIT trap, so no per-file rm fork is added.
The capture deliberately does not live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and a stray file makes it report "No tests found".
Scope correction on the issue: a test body's stderr was never lost. execute_test_body runs the function as "$fn_name" "$@" 2>&1, so it is merged into the captured stdout and already appeared in that test's failure block. What the redirect dropped is stderr with no owning test, which is why the new block is attributed to the file rather than to a test.
The regression test compares occurrence counts between the two modes instead of grepping for the message — the fixture's provider also runs in the main-shell counting pass, so a bare assert_contains passes against the unfixed code.

Each file worker was spawned as `call_test_functions … 2>/dev/null &` since #358, to keep worker noise off the progress line. That also made the same run report differently depending on --parallel. The worker's stderr is now redirected to a per-file capture under the run output dir and rendered after aggregation as a `Stderr from <file>` block. The capture cannot live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and would report "No tests found" for a stray file. Cleanup is the existing run-dir EXIT trap, so no per-file `rm` fork is added. Scope note: a test *body*'s stderr was never lost. execute_test_body runs the function as `"$fn_name" "$@" 2>&1`, so it is merged into the captured stdout and already surfaced in that test's failure block. What the redirect dropped is stderr with no owning test — data providers, hook plumbing, scratch-dir errors — which is why the new block is attributed to the file rather than to a test. The regression test compares occurrence counts between modes rather than just grepping for the message: the fixture's provider also runs in the main-shell counting pass, so the diagnostic leaks in either way and a bare assert_contains passes against the unfixed code. Closes#864

Chemaclass · 2026-07-27T08:47:53Z

🤔 Background

Related #864

Each parallel file worker was spawned as call_test_functions … 2>/dev/null & (added in #358 to keep worker noise off the progress line), so the same run could report differently depending on --parallel, with nothing indicating output had been suppressed.

💡 Changes

Worker stderr is captured per file under the run output dir and rendered after aggregation as a Stderr from <file> block, so data-provider diagnostics, hook-plumbing messages and scratch-dir errors survive. Cleanup rides the existing run-dir EXIT trap, so no per-file rm fork is added.
The capture deliberately does not live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and a stray file makes it report "No tests found".
Scope correction on the issue: a test body's stderr was never lost. execute_test_body runs the function as "$fn_name" "$@" 2>&1, so it is merged into the captured stdout and already appeared in that test's failure block. What the redirect dropped is stderr with no owning test, which is why the new block is attributed to the file rather than to a test.
The regression test compares occurrence counts between the two modes instead of grepping for the message — the fixture's provider also runs in the main-shell counting pass, so a bare assert_contains passes against the unfixed code.

Each file worker was spawned as `call_test_functions … 2>/dev/null &` since #358, to keep worker noise off the progress line. That also made the same run report differently depending on --parallel. The worker's stderr is now redirected to a per-file capture under the run output dir and rendered after aggregation as a `Stderr from <file>` block. The capture cannot live under TEMP_DIR_PARALLEL_TEST_SUITE: state::aggregate_parallel_results walks every entry there and would report "No tests found" for a stray file. Cleanup is the existing run-dir EXIT trap, so no per-file `rm` fork is added. Scope note: a test *body*'s stderr was never lost. execute_test_body runs the function as `"$fn_name" "$@" 2>&1`, so it is merged into the captured stdout and already surfaced in that test's failure block. What the redirect dropped is stderr with no owning test — data providers, hook plumbing, scratch-dir errors — which is why the new block is attributed to the file rather than to a test. The regression test compares occurrence counts between modes rather than just grepping for the message: the fixture's provider also runs in the main-shell counting pass, so the diagnostic leaks in either way and a bare assert_contains passes against the unfixed code. Closes#864

Chemaclass added the bug Something isn't working label Jul 27, 2026

Chemaclass self-assigned this Jul 27, 2026

Chemaclass merged commit 5aadd6b into mainJul 27, 2026
37 checks passed

Chemaclass deleted the fix/864-parallel-stderr branch July 27, 2026 08:53