Bisect Driver

bisect-driver — find the commit that introduced a change

The engine is bisect-driver.js (portable Node, zero deps). You hand it a commit where the behavior was still correct (--good), a commit where it is broken (--bad, default HEAD), and a repro command; it drives git bisect run over the range, parses the culprit, prints its sha + subject + author date, and always runs git bisect reset so the repo ends where it started. This is the catalog's regression-blame — the same operation under a different name.

Commands

node bisect-driver.js --good <ref> [--bad <ref>] --cmd "<repro command>" [--dir <repo>]
node bisect-driver.js --canary
node bisect-driver.js --help

• --good <ref> a commit known to still behave correctly (required).
• --bad <ref> a commit where the behavior is broken (default HEAD).
• --cmd "<...>" the repro, run per candidate commit via git bisect run sh -c.
• --dir <repo> repo to bisect (default: current directory).

The repro command's exit code IS the verdict

git bisect run classifies each commit by the repro's exit status:

So write the repro to exit 0 while the behavior is still correct and non-zero once it is broken. Examples: a test runner that already exits non-zero on failure works as-is; to blame an output change, run.py ... && diff -q got.txt want.txt (0 when it still matches). Wrap a build that must pass first as make || exit 125; ./repro so uncompilable commits are skipped, not blamed.

Examples

• A frozen regression started drifting (a report pinned to hold byte-exact): find the good commit before the drift, then node bisect-driver.js --good <sha-before> --cmd "python run_backtest.py --report && diff -q report.txt frozen/report.txt" --dir /path/to/repo. The culprit is the commit that first broke the pin.
• A frozen regression test started failing: point --cmd straight at the test (python -m pytest tests/test_frozen.py -q) — pytest exits non-zero on failure, which is exactly "bad".
• An API endpoint changed behavior: node bisect-driver.js --good <last-known-good> --cmd "pytest tests/test_api.py::test_signup -q" --dir /path/to/backend.

Safety

• Refuses (exit 2) on a dirty working tree. Bisect checks out historic commits during the run and would clobber uncommitted work — commit, stash, or clean first. "Dirty" includes untracked files.
• Refuses (exit 2) on an in-progress bisect — run git bisect reset first.
• Always resets. The bisect state is torn down in a finally, and best-effort on Ctrl-C (SIGINT), so an interrupted run still restores your HEAD. The tool verifies HEAD is back where it started and warns if not.
  • SIGTERM on Windows is the exception, and it cannot be fixed in code. Windows delivers it via TerminateProcess, which gives no JS handler a chance to run, so kill -TERM can leave bisect state active and HEAD detached. The next run refuses on that state (exit 2) and tells you to run git bisect reset. Use Ctrl-C, not kill -TERM.
• Re-tests its own answer before reporting it (exit 1 if it does not hold). git bisect never re-checks the classifications it made, so ONE bad classification anywhere in the search produces a confident, wrong culprit — and git bisect run still exits 0 with no warning. Measured 2026-08-20: a Ctrl-C mid-run killed one repro invocation, that commit was misclassified, and the driver reported a commit two places past the real defect with no hedge. The driver now re-runs the repro on the reported culprit and its parent: the culprit must fail, the parent must pass. If either half does not hold, no culprit is reported. This also catches a flaky or nondeterministic repro, and a commit where the build itself was broken. Cost: two extra repro runs per bisect.
• Read-only toward the world: it only mutates the target repo's transient bisect state, which it restores.
• Refuses (exit 2) when the repro misclassifies the endpoints. git bisect trusts the marked --good / --bad refs and never re-tests them — it only tests commits between them. So a repro that never actually flips to "bad" on historic checkouts (wrong path, missing script, an always-exit 0 command, or the opposite — an always-non-zero command) would otherwise yield a confident but false culprit. Guard: before starting, the tool checks out each endpoint and runs the repro itself — the --good ref must exit 0 and the --bad ref must exit 1–124. If not, it refuses with an actionable message instead of handing back a silent false positive. (This costs two extra repro runs up front; on a slow repro that is the price of not trusting a broken one.)

Windows notes

• The repro is executed as git bisect run sh -c "<cmd>"; Git for Windows ships the sh it needs (bisect has always depended on it), so this works under both PowerShell 5.1 and Git Bash.
• PowerShell 5.1: wrap the whole --cmd value in double quotes — --cmd "pytest -q && diff -q a b". PS 5.1 has no && at the pipeline level, but here && lives inside the quoted string and is interpreted by sh, so it is fine.
• Do not pass node -e "<quoted code>" as the repro on this machine — PS 5.1 mangles quoted -e and leaves 0-byte junk files. Point the repro at a real script file instead.

Exit codes

0 culprit found · 1 no culprit parsed / bisect error · 2 usage error or preflight refusal (dirty tree, in-progress bisect, unresolvable ref, or a repro that misclassifies the --good/--bad endpoints).

Verification (the done-check)

node bisect-driver.js --canary

Builds a throwaway git repo of ~8 commits with a behavior change planted at a known middle commit, then proves both directions: the culprit is identified as exactly the planted commit and the repo is restored + bisect-state clean (good direction), AND a dirty tree, an in-progress bisect, an unresolvable ref, an always-good repro (bad endpoint passes), and an always-bad repro (good endpoint fails) are each refused with exit 2 (bad direction). Cleans up after itself. MUST print CANARY PASS 16/16 before you trust a result. Never bisects a real repo.