[Bug] Backend did not answer on port 3900 within 300 s (OMNIVOICE_STARTUP_BUDGET_S). C

[luomo1]

Describe the bug

Error

Backend did not answer on port 3900 within 300 s (OMNIVOICE_STARTUP_BUDGET_S). Check the log above.

Environment

App: VoiceStudio 0.5.6
Shell: Electron

Backend reachability

Connection: managed local
Last backend response: none this session

Current backend failure

Stage: failed
Managed: true
Backend did not answer on port 3900 within 300 s (OMNIVOICE_STARTUP_BUDGET_S). Check the log above.

Recent actions

09:42:56 view:home

What I was doing

[nirmalchatur]

Fixed in #2461.

Short version: OMNIVOICE_STARTUP_BUDGET_S is supposed to be the backend's time budget, but we were starting the clock before the backend process even existed.

waitUntilReady() computed its deadline as startedAt + budgetMs, and startedAt gets stamped at the top of start() — long before spawn(). So everything in between got billed to the backend. That in-between time isn't small on a slow machine: resolveRuntimeProject() calls runtimeDependenciesReady(), which imports torch in a child interpreter (roughly 30s per candidate project, and it happens again in resolveSpawnPlan()), then we copy the bundled sources into place, then port selection walks up to 17 candidates. When that total crosses 300s, the readiness loop fails on its very first probe and kills a backend that has been alive for about a second. Hence "did not answer within 300 s" for a process that never got 300s, and Last backend response: none with no crash to point at.

Two commits fix it:

• 616e2f0 moves the deadline to Date.now() + budgetMs at the top of the poll loop, which begins right after the spawn. The launch stays as slow as it is, it just no longer eats the backend's window.
• 8b576a5 resets the captured child output on every launch (childLog is now separate from the session log). Before this, a failure could quote the previous run's output, or the installer's, as though it came from the backend that just died.

I also made the error message say what actually happened, because "did not answer within 300 s" was covering three quite different failures: Last output: ... when we owned a process and it printed something, It printed no output. when we owned one that died silently, and Nothing was spawned for this attempt. for attach-only waits. The ownership check has to be captured before killChild() nulls the field — otherwise the silent-death case, which is the one worth seeing, gets suppressed.

electron/CONTRACT.md and electron/README.md both described the budget as starting at the beginning of the launch, so I updated them to match the actual behaviour. Changelog entry is in as well.

Tests: added backend-startup-budget.test.ts, using fake timers with a controllable pre-spawn cost so the regression is covered without waiting 300s for real. It covers the slow-launch case, a genuinely slow backend, an attach-only wait not blaming output it never produced, installer output not leaking into a later failure, and a managed backend that exits without printing.

On the CodeRabbit point about localizing status.message — I deliberately left that out of this PR. That field carries raw backend output, so localizing it means a real main-to-renderer contract: stable reason codes, interpolation parameters, a fallback to the raw text, and new keys in all 21 locale bundles (tests/locale-coverage.mjs gates the build). Happy to open that separately if you want it tracked.

The PR carries Fixes #2445, so merging it will close this.