Repository navigation
fix: register issue2752 suite in CI LIVE_FILES tripwire - #2900
ranxianglei wants to merge 1 commit into
Conversation
tests/issue2752-forward-partial-outcome.test.ts (c81be81, #2752) matches SPAWN_RE (it spawns the MCP shim over stdio) but was never registered in LIVE_FILES when the two-phase scheduler landed (c2914c3, #2816), so the tripwire dies every test leg before any test runs on master c60c2d1 (#2899). Concurrency review per the tripwire contract (#1248): all HTTP servers bind listen(0) ephemeral ports; the single child spawn receives its own env copy with BILI_MCP_PROXY pinned to its own canned loopback origin; src/mcp.ts is stdio-only (no listen port, no state/cache/config writes); parent-env save/ restore stays inside this file's process. Safe at c=4.
📦 Built Package ArtifactBranch: Option A — Install this exact build (recommended)npm install -g billion-context@0.0.3221-pr.2900.d64d60eThis exact version is immutable and reproducible. The rolling Option B — Download artifact
tar xzf billion-context-pr2900.tgz
npm install -g packageInstall instructions are refreshed on each push. |
|
Duplicate of #2898 — byte-identical one-line registration ( Per repo policy an external contribution is not replaced by an agent-owned PR — #2898 stays the merge vehicle; this one closes as its duplicate. Independently verified before closing:
Issue lifecycle: #2896 auto-closes via #2898's reference; |
What
One-line registration:
tests/issue2752-forward-partial-outcome.test.tsadded toLIVE_FILESinscripts/run-ci-tests.mjs.Why
The tripwire introduced with the two-phase CI scheduler (c2914c3, #2816) requires every process-spawning suite to be registered in
LIVE_FILES. The suite from #2752 (c81be81) spawns the MCP shim over stdio (spawn(process.execPath, ["--import", "tsx", "src/mcp.ts"], …)) so it matchesSPAWN_RE, but was never registered — on master c60c2d1 the guard therefore exits 1 before any test runs, killing all four test legs (ubuntu/windows × node 22/24). Every PR merged onto that tree goes red for no reason of its own (instance: #2893).Concurrency review (required by the tripwire contract, #1248)
listen(0)→ unique ephemeral ports per process; no fixed-port contention (the fix: emit /acp panel as persistent custom message and strip it from the proxy model context (#359) #360 discipline).BILI_MCP_PROXYpinned to its own canned loopback origin — no cross-test or cross-file env leakage (each test file is already a separate process undernode --test).src/mcp.tsis stdio-only: no listen port, no state/cache/config-dir writes, and withBILI_MCP_PROXYexplicitly set it never touches the shared live-instance record.afterEach; node:test runs tests within a file serially by default.Conclusion: the suite legitimately needs its spawn (it is the wire-level lane — the real shim over stdio JSON-RPC), and it is safe at c=4. Register, don't de-spawn.
Verification
node --import tsx --test tests/issue2752-forward-partial-outcome.test.ts→ 22 tests / 4 suites, all pass (~8s), after the samenpm run pretestbuild CI performs first.No behavior change beyond re-enabling the CI gate: same tests, same assertions, same scheduler — the guard now passes instead of dying. Existing red PRs need no code change; re-run/rebase turns them green.
Closes #2899