Node / meshtastic / Meshtastic-Android / files / .agent_memory / session_context.md

Displaying Rendered • View raw • Download

• Reverted ALL google-flavor map changes to before #5684 (per user): restored MapView.kt, NodeClusterMarkers.kt, WaypointMarkers.kt, InlineMap.kt to parent commit bc9f1637; deleted MarkerBitmapRenderer.kt; re-pinned T383838play-services-maps = 20.0.0 in libs.versions.toml. The shipped #5702–#5719 series (Canvas markers + ViewTree-owner band-aids) had lost the info-window popups + interactions.

• Root cause (verified against maps-compose 8.3.0 + android-maps-utils 4.1.1 SOURCE in gradle cache): ONLY T383838Clustering(clusterItemContent=…) crashes — its T383838ComposeUiClusterRenderer builds a detached T383838InvalidatingComposeView with a fake lifecycle owner and NO SavedStateRegistryOwner. T383838MarkerComposable already bakes its icon via the safe in-scope T383838rememberComposeBitmapDescriptor; info windows render with the live marker compositionContext. So InlineMap/NodeTrack/Traceroute were left untouched.

• Fix (NodeClusterMarkers.kt ONLY): icons baked in-scope via T383838rememberComposeBitmapDescriptor(node){ PulsingNodeChip } into a snapshot stateMap; custom T383838private class NodeClusterRenderer : DefaultClusterRenderer assigns them in onBeforeClusterItemRendered/onClusterItemUpdated (bg thread, READ-only — never composes, so the crash class is gone). Native info windows (super sets title/snippet) + onClusterItemInfoWindowClick→navigateToNodeDetails; precision circles drawn from the renderer's own T383838unclusteredItems MutableState (clusterItemDecoration can't fire — T383838ClusterRendererItemState is lib-internal). Strictly better than the elegant-euler Canvas branch — keeps the REAL Compose chip.

• T383838compileGoogleDebugKotlin + T383838spotlessCheck + T383838detekt PASS. NOT committed, NOT device-verified. Next: device-test (clusters show chips + info-window popups + no FATAL), then commit/push.

• Hardened T383838DatabaseManagerWithDbRetryTest to remove CI race conditions by running the manager on a T383838StandardTestDispatcher(testScheduler) instead of real T383838Dispatchers.IO.

• Added a T383838withTimeout(10_000) guard around the test body to fail fast on coordination stalls instead of hanging/flapping.

• Kept the deterministic retry trigger (T383838error("Connection pool is closed")) and retained assertions that first attempt uses old DB and retry uses current DB.

• Made teardown resilient with T383838if (::manager.isInitialized) manager.close() so setup/early failures do not cascade into teardown crashes.

• Verified with T383838:core:database:jvmTest --tests "org.meshtastic.core.database.DatabaseManagerWithDbRetryTest*" and repeated it 5 consecutive runs without failures; T383838:core:database:detekt also passed.

• Integrated T383838NodeRepository into T383838GeminiNanoDocAssistant.kt and the Google AI Koin dependency injection module (T383838GoogleAiModule.kt).

• Developed a dynamic live-state prompt formatting block within T383838buildPrompt(...) that queries current hardware model, firmware version, connection status, GPS capability, channel utilization, airtime, battery level/voltage, user profile long/short names, and total registered mesh peer counts & active online peers directly from T383838NodeRepository's reactive flows.

• Injected this live radio diagnostics context dynamically as a system instruction metadata block on every user query. This empowers the on-device model to answer real-time, personalized diagnostic questions (e.g. "what is my battery level?", "how many active nodes are on my mesh right now?") with 100% on-device offline accuracy.

• Tuned context retrieval constraints for the modern T383838nano-v4-full (Gemini Nano v4) model: expanded the total context budget T383838MAX_CONTEXT_CHARS from 8,000 to 32,000 characters (up to ~12K tokens out of the model's native 32K window), and scaled T383838MAX_PAGE_CHARS to 16,000 characters and T383838MAX_SNIPPET_CHARS to 8,000 characters to supply vastly richer, more detailed, and complete documentation fragments.

• Upgraded the on-device inference flow inside T383838GeminiNanoDocAssistant.kt to use Firebase AI SDK's reactive T383838generateContentStream(prompt) instead of the blocking T383838generateContent invocation.

• Aggregated chunks and emitted incremental T383838AIDocAssistantResult.Partial states down the Kotlin Flow, enabling true word-by-word/chunk-by-chunk streaming in the UI for a much more responsive user experience.

• Refined the T383838SYSTEM_INSTRUCTION personality rules for Chirpy to position him as our adorable LoRa radio Node mascot instead of an avian theme, emphasizing high-enthusiasm mesh networking, signal connectivity, battery status, and radio/routing concepts while preserving technical precision.

• Overhauled system error messages inside T383838DocsNavigation.kt and the loading bubble state inside T383838ChirpyAssistantSheet.kt to align with the mascot theme.

• Added T383838firebase-config dependency to Version Catalog T383838libs.versions.toml and T383838androidApp/build.gradle.kts.

• Added the T383838AIDocAssistantResult.Partial variant to support intermediate stream updates.

• Extended the T383838AIDocAssistant interface and implemented T383838answerStream in both T383838KeywordFallbackAssistant and T383838GeminiNanoDocAssistant.

• Integrated Firebase Remote Config into T383838GeminiNanoDocAssistant to dynamically fetch the model name (T383838chirpy_model_name) and system instruction (T383838chirpy_system_instruction) with release-optimized fetch intervals.

• Refactored T383838GeminiNanoDocAssistant.answer to reuse T383838answerStream flow under the hood, eliminating duplicate prompting code.

• Verified that all unit tests (T383838:feature:docs:allTests) and static analysis checks (T383838spotlessApply spotlessCheck detekt) pass 100% green.

• Always check T383838.skills/compose-ui/strings-index.txt before reading T383838strings.xml.

• Run T383838python3 scripts/sort-strings.py after adding strings to keep the index organized.

• Always check T383838gh run list before pushing.

• Pre-commit hook T383838scripts/ai-guardrail.sh protects against binary leaks (see script for install).

<!-- Older entries archived in session_context.archive.md -->