Improve real Go behavior coverage, not the number of test files. The legacy workflow name/path is retained for continuity. Aider has no MCP client; use the safeoutputs MCP CLI. Use one shell command per line, no heredocs, and SEARCH/REPLACE blocks for edits. Never create skipped tests, placeholders, empty tests, or tests that merely call code without checking its result. A quiet run is better than a low-value PR.

No files need to be attached. Your first reply must run this exact command. It inspects at most three small, non-generated library packages, prefers recent changes, records a fresh baseline, and prints the selected production source, coverage by function, and existing tests. It does not generate tests. An existing open PR from this workflow suppresses further work until reviewed.

python3 scripts/aider_go_test_coverage.py discover

Consume its output before editing. A discovery/test/tool failure is incomplete, not evidence that coverage is complete. Do not retry with insecure TLS settings.

Choose an uncovered function whose behavior is clear from the printed source. Prefer a branch returning a literal string, boolean, or integer: the publication gate must prove the new tests detect an incorrect return value. In exactly the printed TARGET file, write real tests with explicit expected results, including a normal case and an error/boundary case where applicable. Use the actual package declaration and existing assertion style. Tests should catch an incorrect return value or branch condition, not just execute a function. Do not change production code, existing tests, modules, workflow files, skills, or the helper. Do not rerun discovery or modify its baseline.

After adding tests, run this command. It formats only the target file, runs the whole selected package's tests without the test cache, and requires strictly more covered production statements with no lost baseline coverage. It rejects skips, placeholders, missing assertions, existing-file edits, and unrelated files. It also checks up to three mutations of newly covered literal returns using temporary Go compiler overlays. At least one mutation must fail an assertion in the new tests alone, not just an existing test or compilation. Production files are never edited by these probes. Coverage without this behavioral witness is insufficient. Only the verified new test file is staged, never Aider history or skills. The PR includes measured before/after coverage; no speculative coverage claims.

python3 scripts/aider_go_test_coverage.py publish || safeoutputs report_incomplete --reason "Could not verify a real coverage gain or publish the focused test improvement; no pull request confirmed."

Only if discovery succeeded with NO_CANDIDATE, or source inspection establishes no useful bounded assertion-based test, run instead (do not claim every package is covered):

safeoutputs noop --message "No useful bounded coverage improvement found; no placeholder PR created."

Always invoke exactly one safeoutputs CLI command before finishing. Use report_incomplete for failed discovery, failed tests, unchanged coverage, or unfinished work. Never publish a partial improvement or fix a gate by weakening it.