Repository navigation
mcpx: Go tests fail on macOS (2 path tests in internal/config, plus 2 unidentified packages) #1988
Description
Relocated from dezren39/darwin-nix#366 on 2026-10-28 (placement audit): mcpx's single source lives in developing-today/code (pkgs/mcpx); re-home tracked at #309. Original body below; issue references inside it are historical (dezren39/nix numbering).
cd pkgs/mcpx && go test ./...: 10 packages ok, 3 FAIL, 5 with no tests.
internal/config:TestExecIdentityFollowsTheExecutableNotTheCommandNameandTestExecIdentityResolvesAgainstTheServersOwnPathexpect/var/folders/…but get/private/var/folders/…. On macOS/varis a symlink to/private/var, so this looks like a test that doesn't resolve symlinks, not a product bug.- The other 2 failing packages weren't recorded. Re-run
go test ./...to identify them. scripts/conformance.shwas not run.
Acceptance criteria
-
go test ./...is run on macOS against the mcpx version the nix flake pins, and every failing package is named in a comment here (the 2 unidentified packages from the original report are identified) - The two named
internal/configpath tests are confirmed fixed against the pinned mcpx (symlink-resolving temp dir) or fixed if they still fail - The full suite is green on darwin, or any remaining failures are filed as their own issues and linked here
-
scripts/conformance.shis run on darwin and its result recorded here
Metadata
Metadata
Assignees
Labels
Package/app: mcpx (pkgs/ or apps/ in this repo)Medium: expected under 1 day of workTest coverage, test infrastructure, or test-only fixesIntake review done: acceptance criteria, links, and normalized labels addedOrdinary priority (the default)Has a queued task in the Supabase agent queueFiled directly by a humanNot yet triaged; needs a first lookQueue packet kind: verify, reproduce, and route this item
Activity
developing-today commentedon Oct 28, 2026
CODE-913
dezren39 commentedon Oct 28, 2026
History carried over from dezren39/darwin-nix#366 (verbatim, oldest first):
@dezren39 · 2026-10-18T02:23:25Z
Partial fix merged via PR #385 (merge commit eff5826) — leaving this open for the still-unidentified pair.
Fixed: the two
internal/configpath tests. As suspected, test bug not product bug:ExecStampdeliberately reports the symlink-resolved path, and the tests compared it against paths built from unresolvedt.TempDir()(macOS:/var→/private/var). Both tests now usetestsupport.TempDir— the helper added for this exact trap in #281, already used bymerge_test.goin the same package.How it was verified without a Mac:
go testcanonicalises TMPDIR for the test binary, which masks the failure on Linux, so I ran the compiled test binaries directly with TMPDIR behind a symlink (outside $HOME, isolating the symlink variable):/var/folders…vs/private/var/folders…mismatch.internal/configpasses under symlinked and plain TMPDIR.Still open: the other 2 failing packages were never recorded and do not reproduce under the symlink simulation — they may hinge on a macOS-only factor beyond /var symlinking. To close this out, a macOS re-run of
cd pkgs/mcpx && go test ./...that records the failing package names is what's needed.@dezren39 · 2026-10-24T20:42:57Z
Triage re-check (issue-sweep-rotating, 2026-10-24):
The named half is fixed everywhere it needs to be. The two
internal/configtests fixed here via PR #385 carry that fix in the mcpx source's current home: indeveloping-today/codeat mainaa853c40,pkgs/mcpx/internal/config/execid_test.gousestestsupport.TempDir(t)for bothTestExecIdentityFollowsTheExecutableNotTheCommandName(line 29) andTestExecIdentityResolvesAgainstTheServersOwnPath(line 52), with the comment explaining the/var→/private/varsymlink trap.The remainder cannot be run from this repo anymore. The vendored copy this issue was filed against was deleted in #469 (commit
42b94f3: "mcpx: delete the vendored copy; build from pinned developing-today/code input"), andpkgs/mcpxno longer exists in this tree (verified:git ls-tree HEAD pkgs/). So thecd pkgs/mcpx && go test ./...re-run the last comment asked for now means thedeveloping-today/codeclone.Still owed, and blocked on hardware: one macOS run of
go test ./...indeveloping-today/code/pkgs/mcpxthat records the names of the other 2 failing packages (never recorded when this was filed). That needs Drew's Mac — it cannot be reproduced in the Linux sweep sandbox, and the earlier symlink simulation passed every other package.Leaving this open for that Mac re-run. The ROADMAP "macOS Go test failures" row already records this item as moot for this repo once the code-repo build landed (#469) — that build has landed, so this issue is now purely the unidentified-pair question in the code repo.
@dezren39 · 2026-10-27T01:36:53Z
Intake review
Partly looks solved — not closed: the two named
internal/configpath tests appear fixed in the mcpx source this repo now pins from developing-today/code:internal/config/execid_test.gothere uses a symlink-resolvingtestsupport.TempDirwith a comment explaining exactly this /var → /private/var failure. What was never done is the rest of the filing: the other 2 failing packages were never identified, and no full darwingo test ./...orconformance.shresult is recorded. This repo no longer vendors mcpx underpkgs/(it is a pinned input; seemcpx.config.nix), so the re-run must be against the pinned version.Related: developing-today/code #237 (mcpx darwin takeover tests — the sibling darwin-test item; cross-repo set), nix #362 (mcpx pin bump tracking).
Notes / risks: if the pinned mcpx passes the full suite on darwin, this closes with that evidence; if the 2 unknown packages still fail, they become the real scope. Needs a macOS machine/agent to run — that access constraint is the historical blocker for this whole family.
Suggested approach: on a darwin host, check out the pinned mcpx, run
go test ./...andscripts/conformance.sh, post the package-level results here, and close or split from that evidence.— fleet monitor intake (repo-fleet-monitor, 2026-10-26 20:19 CDT run)
22 remaining items
dezren39 commentedon Oct 31, 2026
🤖 Agent note · picked up
5b5c2b581re-verified fresh: Codex clean for head (no findings, 0 Codex review comments), lint/plugin/stress-changed-tests + docs all PASS, diff matches the packet (gate in 2 jobs, exception list in script, AGENTS.md). Onlyfinal: CI time budgetstill queued since 08:33Z (fleet infra). Merging under the standing verify-the-change rule, noting the stuck bookkeeping job.Merge pull request #2995 from developing-today/queue/3285-1988-darwin…
dezren39 commentedon Oct 31, 2026
🤖 Agent note · done
fa7e71b73— darwin/arm64 test-compile gate live in mcpx CI with theinternal/daemonexception pointing at mcpx: takeover/upgrade/handoff unsupported on darwin; Linux-only tests break the darwin build #237. Note: run 41185095612'sfinal: CI time budgetjob was still queued >2h at merge time (fleet infra); every substantive check was green and Codex was clean for the head.dezren39 commentedon Nov 3, 2026
Triage verdict (2026-11-03, main
d3476bce7): the macOS detector now exists and its red is owned by #3016 — closingThis issue's original premise — no one can see what fails on macOS — no longer holds. Two things changed since the last pass: #1973 landed a nightly two-OS full-suite stress (
mcpx.ymlstress-nightly,os: [ubuntu-latest, macos-latest],go test ./... -race -shuffle=<seed> -count=3, failure tracked in #3016 instead of gating PRs), and PR #2995 added the darwin/arm64 test-compile gate (scripts/darwin-compile-check.sh, wired atmcpx.yml:159and:451) withinternal/daemonas the documented #237 exception.Acceptance criteria, one by one:
go test ./...on macOS, every failing package named (incl. the 2 never recorded)examples,internal/cli,internal/conformance,internal/daemon,internal/defaults,internal/e2e,internal/mcpclient,internal/pool,internal/runner,internal/settings. All are recorded on #3016 ("mcpx nightly stress failure", open, updated 2026-11-02).internal/configpath tests confirmed fixedinternal/configis not in the macOS failure list, i.e. thetestsupport.TempDirfix (PR #385) holds under a real darwin run, not just the symlinked-TMPDIR proxy.scripts/conformance.shon darwin, result recordedmodelcontextprotocol/conformanceclone and currently runs only inmcpx-conformance.yml(ubuntu). Scoped as queue task 8163 so the AC survives this close.Why close rather than keep open: with the detector live, #1988 is a duplicate of #3016 for the failure half and of the follow-up for the conformance half. Leaving it open is the "audit without a verdict" shape the intake review warned about.
🤖 Agent note · state: done (closed, superseded by the nightly detector)
bf419306)ses_e5e5e1353ceffeJh7tpGId1sUg· queue task 7822dezren39 commentedon Nov 3, 2026
agent-note:v1
state: triaged (coverage pass)
Coverage pass 2026-11-03: reviewed in the fleet-wide per-issue pass (drift / overlap / already-handled / stale / blockers / ordering). Verdict: still live; queued in repo-health-code-pilot as triage at priority 40 (priority/normal). If this is already handled at current main, the triage closes it with evidence; if it needs splitting, the triage files the split.
Found via: coverage pass 2026-11-03
developing-today commentedon Nov 3, 2026
agent-note:v1
state: triaged (coverage pass)
Coverage pass 2026-11-03: reviewed in the fleet-wide per-issue pass (drift / overlap / already-handled / stale / blockers / ordering). Verdict: still live; queued in repo-health-code-pilot as triage at priority 40 (priority/normal). If this is already handled at current main, the triage closes it with evidence; if it needs splitting, the triage files the split.
Found via: coverage pass 2026-11-03