Skip to content

FAST_FOLLOW.md

Latest commit

 

History

History
36 lines (28 loc) · 3.86 KB

FAST_FOLLOW.md

File metadata and controls

36 lines (28 loc) · 3.86 KB
Raw

Fast-Follow Backlog (post v0.0.101)

Small polish items deferred from the v0.0.101 stability release. None are release blockers — v0.0.101 shipped with all of them as known, low-impact rough edges. Each is grounded in file:line and should land with a test.

1. Re-base image URIs through resolveDocumentPath (iOS reinstall → old images blank)

  • Symptom: After an iOS reinstall (new sandbox container UUID), previously-generated images render blank. Newly-generated images are fine.
  • Root cause: Audio paths are re-based onto the current container via resolveDocumentPath() (src/utils/resolveDocumentPath.ts), but image attachment URIs are NOT. src/services/imageGenerationService.ts:334 stores uri: file://${result.imagePath} (absolute, old-container), and src/components/ChatMessage/components/MessageAttachments.tsx:40 renders source={{ uri }} raw.
  • Fix: Pipe image attachment URIs (chat render + Gallery) through resolveDocumentPath() at render time, mirroring the audio fix. + test.
  • Size: S

2. Reload-banner memory-bail feedback (silent no-op under memory pressure)

  • Symptom: Tapping "Settings changed — tap to reload model" appears to do nothing when the device is memory-starved (observed on Android mid-triple-download; reload unloaded then couldn't refit the model).
  • Root cause: handleReloadTextModel (src/screens/ChatScreen/useChatScreen.ts:328) routes through initiateModelLoad; on insufficient-memory it bails silently. The generate path alerts ([GEN-SM] ... alerted=true); the reload path does not.
  • Fix: Surface the same "not enough memory" alert on the reload memory-bail path. + test.
  • Size: S

3. Queue a send that arrives while the model is still loading

  • Symptom: Sending a message while a text model is loading dispatches immediately instead of queueing (works out via ensureModelLoaded await, but shows no queued indicator).
  • Root cause: The queue gate (src/screens/ChatScreen/useChatGenerationActions.ts:402) only enqueues when isGenerating; during a model LOAD nothing is generating.
  • Fix: (a) also enqueue when activeModelService.getActiveModels().text.isLoading; (b) drain the queue on model-load-complete (the queue currently only drains on generation-complete — without this, an enqueued message would be stranded). + test that a send-during-load is queued AND drained when load finishes.
  • Size: S-M (the drain-on-load-complete hook is the real work)

4. Merge the text-engine capability seam

  • Branch fix/text-engine-capability-seam, commit 0c1fd790 — done + full suite green, sitting in a worktree, not yet merged.
  • Builds TextEngineService + activeModelService.supportsVision/ToolCalling/Thinking() as the single source of truth (kills 5+ duplicated engine==='litert'? capability computations — audit B3), engine-agnostic unloadTextModel with residency bookkeeping for litert too (B2), real interface in engines.ts (H7).
  • Fix: Merge into main. Size: S (already built/tested)

Deferred from the platform-divergence audit (verify, not necessarily fix)

  • iOS memory gate: clamp GGUF fit to min(totalBudget, os_proc_available_memory) + wire real lowMemory flag (ties to the iOS OOM crash cluster). memoryBudget.ts:48-53, DeviceMemoryModule.swift:46.
  • Multi-turn tool calls on iOS (stateless re-prompt) vs Android (native stateful) — verify context holds across a 2nd tool call.
  • Background download on iOS restarts from byte 0 on app-kill (Android resumes) — confirm it fails gracefully, not phantom "downloading".

Hexagon/HTP NPU text support (parked, not scoped for a near-term release)

  • Backend is fully wired + unit-tested; gated behind HTP_ENABLED + HTP_UI_ENABLED (both false). Turning on = ~0.5 day; ship-quality (capability-as-data so the pill can't silently fall back to CPU + on-device tok/s verification) = ~2-4 days. No native blockers.