git-worktree-verification-false-signal
Done. I empirically reproduced both false signals and their fixes, then wrote the solution.
Deliverables
- ~/git-worktree-verification-false-signal.md — the full solution (root cause, exact fix, verification, checklist)
- ~/worktree-verify-harness.sh — the red/green harness used for verification
What was actually reproduced and verified
(A) Silent git checkout refusal → false negative. Confirmed: when the worktree has a local modification to a file that differs between the current and target commit, git checkout -q <new-sha> exits 1, writes only to stderr, and leaves HEAD at the old SHA. With 2>/dev/null and an unpropagated rc, the probe measured old source and reported a phantom regression. Key subtlety found: it's intermittent — if the modified file is unchanged between the two commits, git carries it over and checkout succeeds. Fix verified: git checkout -- . + git clean -fdq, plus a hard assertion of rev-parse --short HEAD and the fix-marker count (0 before, 1 after) that aborts loudly.
(B) vdev guard failure. Confirmed environmental. Go's isVCSRoot (cmd/go/internal/vcs/vcs.go) requires .git to be a directory, but a linked worktree's .git is a gitdir-pointer file, so Go omits all vcs.* build settings and the binary reports vdev. Verified with the same test at the same commit: worktree fails, normal clone passes, and applying an unrelated diff doesn't change the worktree failure. Detached HEAD alone is not the trigger. go version -m shows 0 vs 3 vcs.* keys.
The document closes with the triage rule (re-run at pristine base in the same environment) and the recommendation to let the pre-commit hook judge the staged diff or run the guard in a normal clone.
# Evidence - Problem class: git-worktree-verification-false-signal - Model: openrouter/deepseek/deepseek-v4.1-flash - Solved: 2026-09-16T21:43:18.857Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "Symptom pair seen while verifying a one-line Go fix in an isolated `git worktree` (a foreman-style 'prove it is red at the old commit, green at the new one' check). (A) `git checkout -q <new-sha>` inside the worktree failed SILENTLY (stderr redirected to /dev/null, exit status swallowed by a `&&` chain), so the subsequent build/probe measured the OLD pre-fix source and reported 'the fix does not work' - a false negative that looks exactly like a real regression. Root cause: the worktree had LOCAL modifications to tracked files (test files copied in for the falsification run), and `git checkout <sha>` refuses to clobber modified tracked files with a non-zero exit and no stdout. (B) The same worktree then produced a FALSE guard failure: `gitreins guard` reported `go_tests FAIL` on a test asserting `--version` prints `v<version>-<commit>` because inside a linked worktree the built binary reported `vdev`; the identical test FAILS at the pristine base commit in that worktree, i.e. the failure is environmental (build identity / VCS stamping does not resolve the same way in a linked worktree) and NOT attributable to the change under test. Fixes that actually worked: for (A) assert and PRINT the source identity before building anything - `git rev-parse --short HEAD` compared against the expected sha, plus a grep for the fix's own marker line in the file (e.g. `grep -c '<fixed expression>' path/to/file.go`) - and abort loudly on mismatch; also clean the worktree first (`git checkout -- .` / `git clean -fd` with the probe sources stashed outside the worktree) so the checkout cannot refuse. For (B) do not conclude 'the diff broke the tests': re-run the SAME failing test at the PRISTINE base commit in the same environment - if it fails there too, it is environmental. For a guard run that must judge a real diff while the live tree is clean, either accept the pre-commit hook's run (it judged the staged diff) or run the guard in a normal clone; a linked worktree is not a faithful build environment for version-stamping tests. Verification: the probe at a hard-asserted post-fix sha returned `Get(keyless-agent) PASS, List() 2 agents PASS` while the identical probe at the pre-fix sha returned `invalid key length: 0, want 32` and an empty registry, and the live tree's full `go test ./... -count=1` was 14/14 ok where the worktree's `go_tests` failed only on the build-identity test.", "environment": "Linux, Go 1.26.6, git linked worktree created with `git worktree add --detach <dir> <sha>`, gitreins guard tier-1 (secrets/go_build/go_lint/go_tests)", "language": "bash", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "git-worktree-verification-false-signal", "provider": "openrouter", "solved_at": "2026-09-16T21:43:18.857Z", "version": ""}Done. I empirically reproduced both false signals and their fixes, then wrote the solution.
Deliverables
- ~/git-worktree-verification-false-signal.md — the full solution (root cause, exact fix, verification, checklist)
- ~/worktree-verify-harness.sh — the red/green harness used for verification
What was actually reproduced and verified
(A) Silent git checkout refusal → false negative. Confirmed: when the worktree has a local modification to a file that differs between the current and target commit, git checkout -q <new-sha> exits 1, writes only to stderr, and leaves HEAD at the old SHA. With 2>/dev/null and an unpropagated rc, the probe measured old source and reported a phantom regression. Key subtlety found: it's intermittent — if the modified file is unchanged between the two commits, git carries it over and checkout succeeds. Fix verified: git checkout -- . + git clean -fdq, plus a hard assertion of rev-parse --short HEAD and the fix-marker count (0 before, 1 after) that aborts loudly.
(B) vdev guard failure. Confirmed environmental. Go's isVCSRoot (cmd/go/internal/vcs/vcs.go) requires .git to be a directory, but a linked worktree's .git is a gitdir-pointer file, so Go omits all vcs.* build settings and the binary reports vdev. Verified with the same test at the same commit: worktree fails, normal clone passes, and applying an unrelated diff doesn't change the worktree failure. Detached HEAD alone is not the trigger. go version -m shows 0 vs 3 vcs.* keys.
The document closes with the triage rule (re-run at pristine base in the same environment) and the recommendation to let the pre-commit hook judge the staged diff or run the guard in a normal clone.
# Evidence - Problem class: git-worktree-verification-false-signal - Model: openrouter/deepseek/deepseek-v4.1-flash - Solved: 2026-09-16T21:43:18.857Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "Symptom pair seen while verifying a one-line Go fix in an isolated `git worktree` (a foreman-style 'prove it is red at the old commit, green at the new one' check). (A) `git checkout -q <new-sha>` inside the worktree failed SILENTLY (stderr redirected to /dev/null, exit status swallowed by a `&&` chain), so the subsequent build/probe measured the OLD pre-fix source and reported 'the fix does not work' - a false negative that looks exactly like a real regression. Root cause: the worktree had LOCAL modifications to tracked files (test files copied in for the falsification run), and `git checkout <sha>` refuses to clobber modified tracked files with a non-zero exit and no stdout. (B) The same worktree then produced a FALSE guard failure: `gitreins guard` reported `go_tests FAIL` on a test asserting `--version` prints `v<version>-<commit>` because inside a linked worktree the built binary reported `vdev`; the identical test FAILS at the pristine base commit in that worktree, i.e. the failure is environmental (build identity / VCS stamping does not resolve the same way in a linked worktree) and NOT attributable to the change under test. Fixes that actually worked: for (A) assert and PRINT the source identity before building anything - `git rev-parse --short HEAD` compared against the expected sha, plus a grep for the fix's own marker line in the file (e.g. `grep -c '<fixed expression>' path/to/file.go`) - and abort loudly on mismatch; also clean the worktree first (`git checkout -- .` / `git clean -fd` with the probe sources stashed outside the worktree) so the checkout cannot refuse. For (B) do not conclude 'the diff broke the tests': re-run the SAME failing test at the PRISTINE base commit in the same environment - if it fails there too, it is environmental. For a guard run that must judge a real diff while the live tree is clean, either accept the pre-commit hook's run (it judged the staged diff) or run the guard in a normal clone; a linked worktree is not a faithful build environment for version-stamping tests. Verification: the probe at a hard-asserted post-fix sha returned `Get(keyless-agent) PASS, List() 2 agents PASS` while the identical probe at the pre-fix sha returned `invalid key length: 0, want 32` and an empty registry, and the live tree's full `go test ./... -count=1` was 14/14 ok where the worktree's `go_tests` failed only on the build-identity test.", "environment": "Linux, Go 1.26.6, git linked worktree created with `git worktree add --detach <dir> <sha>`, gitreins guard tier-1 (secrets/go_build/go_lint/go_tests)", "language": "bash", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "git-worktree-verification-false-signal", "provider": "openrouter", "solved_at": "2026-09-16T21:43:18.857Z", "version": ""}