[Blind-Striker]

What

OpenCode.Sdk.Tests' discovery guard now refuses the module when external/opencode is not checked out at the commit this repository records, and names the command that fixes it.

Fixes #469.

Why

On 2026-10-18 a gate worktree's gitlink named v2.0.32 (#465), but its submodule was still checked out at v2.0.29 (git submodule status showed +). The guard checked the checkout's presence, its install and the Bun version, but not which commit was checked out. Every pinned-server test therefore ran a 2.0.29 server, and the suite passed. The env-precedence proof (#470) was the first test that can tell the two versions apart, and it exposed this.

The change

tests/OpenCode.Sdk.Tests/Harness/PinnedServerPrerequisites.cs:

• New inputs. FindAsync takes the recorded commit and the checked-out commit.
  • The recorded commit is the index gitlink, the second field of git ls-files -s -- external/opencode. That is the same entry git submodule status compares against, so a staged gitlink counts.
  • The checked-out commit is git -C external/opencode rev-parse HEAD.
• Differing commits. The guard adds: "The pin's checkout at '' is , but this repository records . Run: git submodule update --init --depth 1 external/opencode".
• Unreadable commits. If either commit cannot be read while the source is present, the guard adds a problem naming which one and the same command. The guard fails closed here.
• Missing source. It adds only the existing missing-source problem, as before.
• Problem order. The order of the existing problems is unchanged: an invalid OPENCODE_SDK_TESTS_BUN_MISMATCH value, then missing source with its early return. The commit problems come next, followed by the Bun and install problems.
• One process helper. The Bun version read and the two git reads share one private helper. The Bun read's command, messages and cancellation handling are unchanged.

PinnedServerPrerequisitesTests:

• New rows:
  • a wrong checkout plus a missing install, asserting the full ordered list;
  • each commit unreadable;
  • missing source with no commit problem;
  • a typo'd switch value plus missing source, asserting the full ordered list.
• Existing rows reach FindAsync through a matching-commit helper, with assertions unchanged.
• Clone baseline. tools/jscpd/tests.baseline.json only removes two entries, 482c654299e74350 and 57470feca6bbb65f. They are the duplicated arranges that the helper eliminated, and nothing is added.

CI is unaffected: a fresh checkout's submodule always matches.

Proof

Each mutant is a compiling edit to the guard, run on net10.0 and on net472 and then restored:

Mutant	Effect
The commit comparison neutralized (&& string.IsNullOrEmpty(checkedOutCommit))	Only FindAsync_Should_Name_A_Wrong_Checkout_Before_A_Missing_Install is red. The && false form is rejected by S1125, a second guard.
The typo check moved after the missing-source early return	Only FindAsync_Should_Name_A_Mismatch_Typo_Before_Missing_Source is red.

Checks:

• The class: 20 of 20 on net10.0 and on net472. The guard itself accepts this matching checkout, because the module would refuse at discovery otherwise.
• The whole OpenCode.Sdk.Tools.Tests project: 2072 of 2072.
• verify-clones: exit 0, tests at 4.97%, 0 new.
• Slopwatch: 0. Format: clean.
• Local check: GATE PASS on 6d3d93eb (master df25ea59): every step including the new submodule and pinned-install steps, and 10,805 tests with 0 failed.

Provenance

Sol Fast drafted the guard and the rows, then made a fix pass. The coordinator finished the remaining items by hand, because that was cheaper than a third dispatch:

• the restored problem order, with its red proof;
• the two stale baseline removals;
• two nits.

The coordinator also verified the evidence and ran the local check.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}