[Feat]: Configurable per-server network request timeout for remote models

[a-ghorbani]

Summary

Adds a user-configurable network request timeout for remote / OpenAI-compatible model servers. Some self-hosted or remote LLMs have slow cold starts (>180s) or slow generation; the previously hard-coded 30s connection / 60s idle guards would abort those conversations prematurely. Users can now set a per-server timeout (e.g. 600s).

Closes #776.

What changed

• ServerConfig.requestTimeoutMs — new optional per-server field. Absent on older persisted configs → falls back to existing defaults; no migration.
• src/api/openai.ts — all four request helpers (streamChatCompletion, fetchModels, fetchModelsWithHeaders, testConnection) accept an optional timeoutMs. A single resolveTimeout helper is the only normalizer: undefined / ≤0 / NaN / non-finite → the existing default (30s connection, 60s idle). Stores, engine, and UI forward the raw stored value untouched.
• Two-phase structure preserved — a supplied value overrides both the connection-phase and idle-phase guards; the idle timer still resets per chunk, so a healthy stream is never killed by a total wall-clock cap.
• UI — a "Request timeout (seconds)" input on both the edit path (ServerDetailsSheet) and the add path (RemoteModelSheet); seconds↔ms conversion at the save boundary. Live connection probes honor the timeout: the in-edit field on add/edit, and the saved value on a known-server chip press.
• ServerStore remains the sole writer of the field (addServer / updateServer).

Tests

• Per-server timeout normalization (≤0/NaN/non-finite/undefined → default; positive passes through), raised-timeout success, default-on-unset, idle abort, and idle-reset-per-chunk (fake timers).
• Engine forwards raw timeoutMs and rebuilds on reselect; ServerStore forwards undefined without crashing.
• Both sheets: input render + s↔ms conversion; the two distinct probe feeds (in-edit field vs saved value) verified not cross-wired.
• 301 affected tests pass; tsc --noEmit and lint clean. No native changes (no package.json / ios/ / android/).

Visual evidence

Both new "Request timeout (seconds)" inputs were captured on the iOS simulator (iPhone 17 Pro) and attached as a PR comment:

• Add path (RemoteModelSheet, remote-timeout-input) — field appears after a probe attempt, with helper text.
• Edit path (ServerDetailsSheet, server-details-timeout-input) — empty (default) and populated (600) states.

A local OpenAI-compatible mock server (GET /v1/models) was used so the probe succeeds and a server can be saved to reach the edit sheet. Screenshots are hosted as PR-comment images (no screenshots are committed to the source tree).

The "slow server succeeds with a raised timeout" runtime outcome still needs a real reachable slow remote server and remains a manual check.

Generated by PocketPal Dev Team

[a-ghorbani]

Visual evidence — Request timeout (seconds) fields

Captured on iOS simulator (iPhone 17 Pro, iOS 26.0) for issue #776. Both new timeout inputs render with label and helper text.

Add path — RemoteModelSheet (testID remote-timeout-input)

Shown in the server-fields block after a connection probe.

Edit path — ServerDetailsSheet (testID server-details-timeout-input)

Empty / default state:

Populated (600 seconds):

The "slow server succeeds with a raised timeout" runtime behaviour still needs a real reachable slow server and remains a manual check.

Generated by PocketPal Dev Team

[Copilot]

Pull request overview

This PR adds a configurable, per-server network request timeout for remote/OpenAI-compatible model servers, allowing slow cold-start or slow-streaming servers to complete requests without being cut off by the existing fixed timeouts.

Changes:

• Introduces ServerConfig.requestTimeoutMs and forwards it through stores/engine into OpenAI request helpers.
• Adds a single normalization point (resolveTimeout) in src/api/openai.ts and applies the configured timeout to both connection and idle guards (while keeping per-chunk idle resets).
• Adds UI inputs (add/edit flows) for “Request timeout (seconds)” with seconds↔ms conversion, plus unit/e2e test coverage for wiring and behavior.

Reviewed changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/utils/types.ts	Adds requestTimeoutMs to persisted server config shape.
src/store/ServerStore.ts	Forwards per-server timeout to model fetch and connection test calls.
src/store/ModelStore.ts	Passes server timeout into the remote OpenAI engine construction.
src/store/tests/ServerStore.test.ts	Verifies timeout forwarding (and undefined passthrough) for store calls.
src/store/tests/ModelStore.test.ts	Verifies engine rebuild/selection carries updated timeout values.
src/locales/en.json	Adds i18n strings for the new timeout inputs.
src/components/ServerDetailsSheet/ServerDetailsSheet.tsx	Adds edit-flow timeout input, prefill, save, and probe wiring.
src/components/ServerDetailsSheet/tests/ServerDetailsSheet.test.tsx	Tests edit-sheet rendering, ms↔s conversion, and probe wiring.
src/components/RemoteModelSheet/RemoteModelSheet.tsx	Adds add-flow timeout input, probe wiring, and persistence via addServer.
src/components/RemoteModelSheet/tests/RemoteModelSheet.test.tsx	Tests add-flow probe feed vs chip feed and persistence conversions.
src/api/openai.ts	Adds resolveTimeout and threads timeout through fetch + streaming helpers.
src/api/completionEngines.ts	Extends OpenAICompletionEngine to carry/forward timeout to streaming.
src/api/tests/openai.test.ts	Adds coverage for timeout normalization and stream/fetch timeout behavior.
src/api/tests/completionEngines.test.ts	Ensures engine forwards constructed timeout argument.
e2e/helpers/selectors.ts	Adds e2e selectors for the new timeout inputs.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[a-ghorbani]

E2E run — all feature specs + quick-smoke (iOS sim + Android emulator)

Ran the 12 feature/smoke specs on iPhone 17 Pro simulator (iOS 26.0) and Android emulator-5554 (API 16) against this branch. App built fresh for both (iOS Release-iphonesimulator .app, Android app-e2e-releaseE2e.apk).

Spec	iOS	Android
quick-smoke	✅	✅
remote-server (this PR's surface)	✅	✅
draft-autosave	✅	✅
context-banner	✅	✅
download-cancel	✅	✅
hub-run	✅	✅
language	✅	✅
pal-greeting	⚠️ env	✅
talent-tool-use	⚠️ env	✅
thinking	⚠️ env	⚠️ env
thinking-pal-override	⚠️ env	⚠️ env
purchase-flow	⚠️ env	⚠️ env

The spec this PR actually touches — remote-server — passes on both platforms, as does quick-smoke. None of the ⚠️ failures are in code this PR changes.

Why the ⚠️ specs failed (environmental, not regressions)

• thinking / thinking-pal-override / pal-greeting / talent-tool-use — these download a model (Qwen3-0.6B / 1.7B) on a fresh app install and assert on inference output (thinking bubble, greeting chip, tool-call HTML). The failures are NoSuchElement waiting on model download/load/inference to settle in time. pal-greeting and talent-tool-use pass on Android, confirming the feature works — the iOS misses are model-load/inference timing on a cold install.
• purchase-flow — fails in the beforeEach hook with TypeError: fetch failed: it needs a reachable PalsHub backend + authenticated session, neither of which is wired in this local run.

Screenshots

Generated screenshots are hosted as release assets (same pattern as the timeout-field captures) and embedded below.

language — 11 locales, iOS (passing)

language — 11 locales, Android (passing)

pal-greeting / talent-tool-use — Android passing result + iOS failure state

pal-greeting (Android, passing):

pal-greeting (iOS, failure):

talent-tool-use (Android, passing):

talent-tool-use (iOS, failure):

thinking — failure state (iOS + Android)

Generated by PocketPal Dev Team

[a-ghorbani]

Update — earlier ⚠️ failures root-caused and fixed; suite now green

Following up on the E2E run above. That comment chalked the ⚠️ specs up to "model-load timing / environment" — that diagnosis was wrong. They were two E2E-harness tap bugs (not product or feature issues), now fixed on this branch, plus one backend that was simply down at the time.

Root causes

1. Drawer tap no-op on iOS (pal-greeting, talent-tool-use, thinking-pal-override). The palsTab selector had been switched to the drawer-item-pals testID so the drawer open/close indicator survives a language switch — but navigateToPals also taps that selector, and a Paper Drawer.Item doesn't reliably respond to a testID tap on iOS, so the tap silently missed and the drawer never closed. Fix: tap the visible label (like the other tabs); keep the testID for the indicator. → fix(e2e): tap Pals drawer label so navigateToPals works on iOS
2. Thinking-toggle tap miss (thinking, both platforms). The toggle tap offsets to avoid the adjacent TTS VoiceChip, but the offset could compute an x just past the toggle's edge, so the tap missed and the toggle never flipped. Fix: clamp the tap inside the button bounds. → fix(e2e): clamp thinking-toggle tap inside the button bounds
3. purchase-flow — was failing only because the PalsHub e2e test backend was unreachable at the time. With it up, the authenticated checkout flow passes on iOS. (This spec is iOS-only — it uses -ios predicate string locators + ASWebAuthenticationSession.)

Re-run — all feature specs + quick-smoke (iPhone 17 Pro sim + emulator-5554)

Spec	iOS	Android
quick-smoke	✅	✅
remote-server (this PR's surface)	✅	✅
pal-greeting	✅	✅
talent-tool-use	✅	✅
thinking	✅	✅
thinking-pal-override	✅	✅
draft-autosave	✅	✅
context-banner	✅	✅
download-cancel	✅	✅
hub-run	✅	✅
language	✅	✅
purchase-flow	✅	n/a (iOS-only spec)

Every spec now passes on every platform it applies to. The two fixes are test-harness only (no src/ change), so the timeout feature itself is unchanged from the approved review.

Generated by PocketPal Dev Team

[a-ghorbani]

Thanks @copilot — addressed in 4979d336:

• parseTimeoutMs duplication → extracted into src/utils/timeout.ts, now imported by both ServerDetailsSheet and RemoteModelSheet (byte-equivalent, no behavior change).
• requestTimeoutPlaceholder "default 30s" → reworded to "Optional - leave empty for default", since an empty field keeps two defaults (30s connection / 60s idle), so a single number was misleading.

On the uncontrolled defaultValue timeout input in ServerDetailsSheet: left as-is intentionally. It mirrors the existing defaultValue={url} field directly above it, so it's consistent with the sheet's established pattern. The save-desync concern doesn't apply here — onChangeText={setTimeoutSeconds} keeps state in sync on every keystroke and handleSave reads that state. Switching only the timeout field to a controlled value would make it inconsistent with the URL field in the same sheet; converting both is out of scope for this change.

Generated by PocketPal Dev Team