Claude Code First, specification version 0.1.0. The plain-language guide
Version 0.1.0. Grade R3/S3.
F202 A gate run in a checkout that something else is writing reports on the moving tree, not on the change
| Field | Value |
|---|---|
| Status | active |
| Topic | Commits and concurrent sessions |
| Grade | R3/S3 |
| Cases | roadkeep, Shio |
| Contrary | none |
| Harness facts | none |
Claim
A test or gate run reads the live working tree, so another session's uncommitted edits, a second run of the same gate, or other work writing to the repository during the run produced reds and counts that were not about the change, measured in both roadkeep and Shio, and each project answered with a mechanism that makes the run say so.
Mechanism
A gate has no notion of whose edits it is reading. When several sessions share one checkout, one session's run sees the other's half-finished files; two runs of one gate share its log and build directory and corrupt both; a file written mid-run makes a value read at collection disagree with the disk. The output looks like any other red or green, so the session either chases a defect that is not there or, if it knows the tree is shared, has to argue in its commit message that the red is someone else's. The remedies found make the run detect the condition: a lock per gate, a snapshot taken at collection, a fingerprint that names the file that moved.
Observations
- Shio.
- Two runs of one gate overlapped, a background run and a foreground one, and the runner now
refuses the second, because the collision was measured: "reported 3 errors over 1092 tests where the tree alone reports 1894" [shio@821f18d74:scripts/run-gate.mjs#L57] green. The rule is "one run of a gate at a time" [shio@821f18d74:scripts/run-gate.mjs#L49]. The task filing it records "the issue of concurrent gate runs corrupting logs and producing false failure reports" [shio@be14dab5e].
- Another session's uncommitted work turned a gate red for a commit that did not contain
it, twice on 2026-09-14. One message says pnpm test "is red on two" [shio@632cbdac4] tests, "both caused by uncommitted work from another session" [shio@632cbdac4]; the next, nineteen minutes later, reports three reds "all three inside the chrome removal another session has uncommitted in this checkout" [shio@51511ac96].
- A measurement was postponed for the same reason: "at 06:55 another session had uncommitted" [shio@1dcd16a1e]
edits, and a census "would have been a moving target under it" [shio@1dcd16a1e].
- roadkeep.
- RK263, while shipping RK261: "one run reported six failures and the same source reported 1940" [roadkeep@91754240:tests/conftest.py#L11]
passed, and "Five of the six were not about the code." [roadkeep@91754240:tests/conftest.py#L12] Work was running beside the suite against the same repository, and "nothing in the output said so" [roadkeep@91754240:tests/conftest.py#L14]. The remedy is a fixture that skips the checkout-reading tests with a warning naming what moved [roadkeep@f7bcf12f].
- RK315, the next day: "A test asserting on this repository's own docs fails when another session writes them mid-run" [roadkeep@91754240:docs/CHANGELOG.md#L895],
answered by reading one copy taken at collection [roadkeep@ab30a6ee].
- RK1232: the suite's worker pool took every core, which "left nothing for the editor or for a second session sharing the checkout" [roadkeep@166840ee],
and the same commit discards its own wider timing sweep as "not a measurement" because "the tree moved under every run" [roadkeep@166840ee].
- RK1440, seen from an adopting repository: the gate crashed because the tool on the PATH
resolved into a checkout of roadkeep "and another session was editing it" [roadkeep@91754240:src/roadkeep/linting.py#L3219].
Independence
Two independent origins. roadkeep's and Shio's failures are each in their own suite and their own gate runner (pytest fixtures in roadkeep, run-gate.mjs in Shio), and neither remedy was copied from the other: the lock, the snapshot and the fingerprint are three different mechanisms. Within each project the failure recurred on separate occasions (roadkeep 2026-08-04, 2026-08-05 and 2026-08-17; Shio 2026-08-17 and 2026-09-14).
Counter-evidence
None found at the pins. Two limits: Shio's overlapping runs came from one session's background job, not from two sessions, and roadkeep's RK263 names worktree operations beside the suite without saying which session ran them, so the sources support "something else writing the checkout", not always "another agent session". The other three cases record no such incident, which is not evidence that it did not occur there.
Grade
R3: independent failures in two projects. S3: both projects counted the effect in their own sources, Shio's collision at 3 errors over 1092 tests against 1894 green, and roadkeep's run at six failures, five not about the code, against 1940 passed. Not S4: the lock and the fixture are mechanisms that detect the condition, and the pins record no occasion on which either was seen to refuse or skip a run.