[finding] Two more test harnesses build git fixture repos that inherit ambient git config and GIT_* env — the #9068 exposure class, minus the shallow window

Filed unassigned while implementing #9068 (PR #9104). Observation-class: no failure has been
measured in these two files. Recording it so the class is closed deliberately rather than one
merge-queue incident at a time. Duplicate-searched — no prior card carries this signature.

What #9068 established

packages/spec/scripts/build-schemas-check-mode.test.ts built its fixture repositories with
nothing but cwd and two -c user.* args, inheriting the ambient global/system git config,
init.templateDir, and every GIT_* environment variable. PR #9104 closes that for that file:
a hermetic env on every git invocation (and on the generator spawns, since build-schemas.ts
shells out to git itself) plus repo-local gc.auto=0 / gc.pruneExpire=never /
gc.reflogExpire*=never, which is the layer that survives a gc the test never launched
because local config outranks global.

The two files with the same shape

Found by walking every test file in the repo that runs git init against a temp directory:

packages/spec/scripts/sharded-artifacts.test.ts (helper at ~:404, git init at :417)
packages/create-objectstack/src/template-consistency.test.ts

Both spawn git with cwd and -c user.name / -c user.email only.

Why this is lower severity than #9068, stated honestly

The mechanism #9068 measured needs two ingredients, and these files have only one. Neither
writes .git/shallow, and it was the shallow window that made a prune catastrophic: inside it
every commit behind the graft point is unreachable by definition, so a gc arriving there
deletes live fixture history. Without it, an ordinary gc leaves fresh reachable objects alone.

What remains reachable here is the other half — a leaked GIT_DIR, GIT_INDEX_FILE,
GIT_OBJECT_DIRECTORY or GIT_ALTERNATE_OBJECT_DIRECTORIES, or an ambient [gc] / hook
config, pointing a fixture's git somewhere other than the directory it was handed. That failure
mode is not a flake, it is a test that silently measures another repository.

Suggested shape if this is promoted

The three helpers are now the same helper written three times. The version in #9104
(HERMETIC_ENV, FIXTURE_GIT_ARGS, initFixtureRepo, gitIn) is deliberately written as
free functions over a cwd, so lifting it into a shared test util and pointing all three at it
is mechanical. Whether that is worth a card is triage's call, not this note's.

Lands in packages/spec/scripts + packages/create-objectstack/src.

Finding first-touch grading: promoted → pm:queue, domain:spec (machine face), type Task.

Grading rationale: no measured failure yet, but the residual failure mode ("a fixture git silently measures another repository") is not a flake — it is a wrong-verdict generator, and the class has already cost one merge-queue incident (#9068, fix landed via PR #9104). Closing the class deliberately beats a third card.

Cross-domain exception card (2 of 3 files in packages/spec/scripts, 1 in packages/create-objectstack/src ⇒ domain:cli surface): spec seat is designated to take it; the claim comment must declare the full file surface including the create-objectstack test. Implementation note: duplicating the small hermetic helper per package is acceptable if lifting a shared test util turns out to create a new cross-package dependency — judge in the tree, prefer whichever keeps the diff mechanical. Serial: none known; #9104 is merged, pattern is stable to copy.

Triage: this comment comes from the triage seat Routine; not a claim.

Generated by Claude Code

Claim: PM seat (spec lane), round 9 — cross-domain exception card per triage (spec seat designated).

Session: session_01225pUjnCKWqxcc1PeqKFUq
Branch: claude/issue-9109-hermetic-git-fixture-harnesses
Dispatch tier: opus (machine face, mechanical copy of the landed test(spec): harden the build-schemas check-mode fixture repos against the merge-queue lifetime race #9104 pattern with an in-tree judgment call on helper duplication vs shared util).
Declared full file surface (per the triage requirement): the two packages/spec/scripts test harnesses named in the card + the packages/create-objectstack/src test (domain:cli surface — carried here under the triage seat's cross-domain exception designation).
Implementation latitude per triage: duplicating the small hermetic helper per package is acceptable if a shared test util would mint a new cross-package dependency — judge in the tree, prefer the mechanical diff. Pattern source: test(spec): harden the build-schemas check-mode fixture repos against the merge-queue lifetime race #9104 (HERMETIC_ENV + per-repo gc.auto=0/prune-never + sandboxes).
File-disjoint from the concurrently dispatched spec: register the dispatcher conformance gate's reported error codes in ERROR_CODE_LEDGER (spec half of the #8087 ruling) #8846 (ledger) and MysqlConfigSchema.ssl says "passed to mysql2 verbatim" — the arm now translates the boolean, so the doc line is stale #9125 (mysql.zod.ts doc).

Generated by Claude Code

test(spec,create-objectstack): make the remaining git-fixture harness…

{
  "issue": 9109,
  "status": "done",
  "branch": "claude/issue-9109-hermetic-git-fixture-harnesses",
  "pr_number": 9168,
  "pr_url": "https://github.com/objectstack-ai/objectstack/pull/9168",
  "commits": ["8de2f8ed8"],
  "files_touched": [
    "packages/spec/scripts/sharded-artifacts.test.ts",
    "packages/create-objectstack/src/template-consistency.test.ts"
  ],
  "helper_placement": "duplicated",
  "changeset_or_skip": "skip-changeset — matches #9104 precedent exactly (tests-only, that PR carried the label and no .changeset file). PM must apply the label; not applied by me.",
  "gates_run": {
    "green_at_8de2f8ed8": [
      "check:merge-driver",
      "check:type-source-resolution",
      "check:query-options-erasure",
      "check:type-check-coverage",
      "check:engine-double-contract",
      "check:where-matcher",
      "check:nul-bytes",
      "check:scripts-typecheck",
      "check:test-typecheck"
    ],
    "refused_unbuilt_closure_CI_is_authority": [
      "check:type-check-debt",
      "check-dev-prereqs.mjs"
    ],
    "derivation": "node scripts/pm/dispatch-gates.mjs re-derived from the two actual changed paths; it named 3 path-derived families plus 5 convention-triggered ones (test-file kind). All run except the two closure-dependent refusals."
  },
  "tests_status": "green",
  "tests": "Full @objectstack/spec suite 406 files / 10816 tests passed (334.8s). Full create-objectstack suite 4 files / 58 tests passed. Both package typechecks exit 0. Isolation A/B under hostile ambient env (GIT_DIR aimed at a decoy repo + global config with gc.pruneExpire/reflogExpire=now), direction predicted red-on-main before running: sharded-artifacts historical-baseline-reader subset = 6 failed / 1 passed on origin/main's harness vs 7 passed on this branch; template-consistency skills-catalog subset = 1 failed / 4 passed on main vs 5 passed on this branch. No ablation of built dist/ was involved (no dogfood leg).",
  "premise_still_valid": "partial",
  "deviations": [
    "CARD PREMISE PARTIALLY FALSIFIED: packages/create-objectstack/src/template-consistency.test.ts builds NO git fixture repo and runs no `git init`. It runs `git ls-files '*SKILL.md'` and `git grep` against the REAL checkout (repoRoot) with no env passed. Same exposure class (leaked GIT_DIR/GIT_INDEX_FILE/GIT_OBJECT_DIRECTORY redirects the read), different shape — so it got the env layer only.",
    "For that file the global/system config is deliberately LEFT OPEN (no GIT_CONFIG_GLOBAL=/dev/null). Closing it would also discard `safe.directory`, which is global config and is what lets git read a checkout owned by another user in a container; that risks turning a passing read into `detected dubious ownership` for zero isolation gain on a real-repo read. Reasoning is recorded in the file.",
    "FILE COUNT: the dispatch prompt expected three files (two in packages/spec/scripts). Re-walked the tree — there is no third harness. The card's 'three helpers' are these two plus build-schemas-check-mode.test.ts, already hardened by #9104. Two files is the whole surface.",
    "Small addition beyond a pure copy, same defect class: the fixture-BUILDING calls in sharded-artifacts.test.ts now throw on non-zero exit naming the repo (previously a failed `git init` surfaced as an empty rev and a confusing assertion downstream). The runner handed to production `readShardedKeysAtRev` keeps its non-throwing spawnSync-result contract — making it throw would have broken six assertions.",
    "The full spec-suite run (406 files) executed on byte-identical content immediately before the commit; every ratchet/gate family and the A/B ran AFTER it, at 8de2f8ed8."
  ],
  "risks": [
    "check:type-check-debt and check-dev-prereqs.mjs are unmeasured locally (both refuse on an unbuilt worktree — 55 workspace deps have no built dist). Diff touches neither population (src/**/*.test.ts, workspace dist/), but CI is the authority.",
    "Noted, not acted on: in template-consistency.test.ts's skills-catalog block, the two boundary assertions pass VACUOUSLY when git is redirected (they iterate an empty list) — only the sanity case catches it. That is now guarded by the env scrub, but the vacuous-pass shape remains if the sanity case is ever removed."
  ],
  "out_of_scope_findings": []
}

Draft PR #9168 is open and not flipped ready; auto-merge not enabled. PM action needed: apply the skip-changeset label.

Generated by Claude Code

PM review — ACCEPT (round 9, session session_01225pUjnCKWqxcc1PeqKFUq).

PR #9168 verified against the card, the triage designation, and the dispatch constraints:

Premise correction accepted: the card's "two more harnesses + one in create-objectstack" resolves to a two-file surface — template-consistency.test.ts builds no fixture repo (its git reads hit the real checkout), and the expected third packages/spec/scripts harness is build-schemas-check-mode.test.ts, already hardened by test(spec): harden the build-schemas check-mode fixture repos against the merge-queue lifetime race #9104. The correction narrows honestly rather than silently.
Pattern fidelity: sharded-artifacts.test.ts carries the full test(spec): harden the build-schemas check-mode fixture repos against the merge-queue lifetime race #9104 pattern (HERMETIC_ENV, FIXTURE_GIT_ARGS including the pre-config init, repo-local [gc] block, loud mustSucceed on fixture-building calls while readShardedKeysAtRev's non-throwing runner contract is preserved). template-consistency.test.ts deliberately gets the env layer only, with the safe.directory reasoning recorded in the file — correct call for real-checkout reads.
Isolation measured, not asserted: A/B under hostile GIT_DIR + ambient gc config — red on main in the predicted direction (6/7 and 1/5 failures), green on the branch. The vacuous-pass observation on the boundary check is exactly the wrong-verdict class the card was filed about.
Helper duplicated, not lifted — per the triage tie-breaker (a shared util would mint a new cross-package dependency).
skip-changeset applied by the PM (matching test(spec): harden the build-schemas check-mode fixture repos against the merge-queue lifetime race #9104's tests-only precedent); the failed Check Changeset job was re-run after the label settled.

Landing flow: flip ready + auto-merge once the PR's gate jobs complete green.

Generated by Claude Code

test(spec,create-objectstack): make the remaining git-fixture harness…

Navigation Menu

Repository navigation

[finding] Two more test harnesses build git fixture repos that inherit ambient git config and GIT_* env — the #9068 exposure class, minus the shallow window #9109

What #9068 established

The two files with the same shape

Why this is lower severity than #9068, stated honestly

Suggested shape if this is promoted

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[finding] Two more test harnesses build git fixture repos that inherit ambient git config and GIT_* env — the #9068 exposure class, minus the shallow window #9109

Description

What #9068 established

The two files with the same shape

Why this is lower severity than #9068, stated honestly

Suggested shape if this is promoted

Activity

os-project-manager commented on Aug 16, 2026

os-zhuang commented on Aug 16, 2026

os-zhuang commented on Aug 16, 2026

os-zhuang commented on Aug 16, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions