Skip to content

BACKLOG.md

Latest commit

 

History

History
73 lines (58 loc) · 6.08 KB

BACKLOG.md

File metadata and controls

73 lines (58 loc) · 6.08 KB

PokeClaw Backlog

Items go in, get prioritized, get done, get crossed out. Simple.

Priority: P0 = blocks users, fix now. P1 = next up. P2 = when we get to it. P3 = nice to have.

Bugs

  • P0 Release publishing: install stable signing secrets on agents-io/PokeClaw so tag-based GitHub releases can ship a signed release APK instead of failing closed
  • P1 Historical upgrade gap: users on the older public debug signing path still need a one-time uninstall + reinstall because the original public signing key is already lost
  • P2 K3-a: Auto-return fires on every service connect, not just user-initiated permission enable
  • P2 B2-a: No auto-return to PokeClaw after task completes in another app (e.g., stuck in YouTube)
  • P1 Investigate MediaTek/Samsung local-engine bring-up failures that still report OpenCL/LiteRT engine creation errors on some devices even after GPU→CPU fallback
  • P2 Settings screen: active model row breaks layout when the model name is long; keep the label/value aligned and truncate or wrap cleanly without shoving the left label into a narrow column

Features

  • P0 Missed-call auto follow-up: when a call is missed, let PokeClaw detect the missed-call event, send a follow-up message to that caller automatically, and keep the status visible in the same chatroom instead of as a hidden background-only action. Prefer SMS / true Android API paths first; do not make accessibility-driven WhatsApp automation the default design.
  • P1 Local model import UX: keep shared-storage .litertlm import easy and explain clearly why other apps' Android/data/... sandboxes (for example Edge Gallery) are not directly readable
  • P1 More small local model options: add 1B / 1.5B-class local models so lower-RAM phones can still run a useful on-device agent
  • P1 Custom local model sources: let users point PokeClaw at user-defined model URLs / hosted downloads instead of only the built-in catalog
  • P2 Google AI Core integration research: evaluate Android's official on-device AI / system model APIs as an optional local runtime path
  • P1 Structured monitor identifiers: let monitor setup keep a user-facing nickname while using a more stable identifier where possible (phone number / app-stable id / aliases) so WhatsApp/Telegram display-name drift stops breaking setup
  • P2 Chat keyboard dismissal polish: tapping non-button chatroom space should reliably clear focus and hide IME in both empty and non-empty conversations
  • P1 Structure-first UI matching: remove remaining language-specific text heuristics where the platform exposes a stable structural hook first (dialog positive buttons, send affordances, standard action widgets)
  • P1 Tinder automation: auto swipe + monitor matches + auto-reply using same monitor architecture as WhatsApp
  • P1 NLP Playbooks (Layer 2): 5 playbooks in system prompt (Search in App, Navigate Settings, Compose Email, Read Screen, Read Notifications) — done 2026-04-08
  • P1 In-chat task auto-return — done 2026-04-08
  • P2 Monitor stays in app — done 2026-04-08, removed GLOBAL_ACTION_HOME
  • P2 Unified task registry: monitor + agent tasks tracked in same system (top bar, floating button, etc.)
  • P3 Rename chat session (H6): pencil icon in sidebar → InputDialog → update title in DB + markdown
  • P3 Floating button: use PokeClaw icon instead of "AI" text
  • P3 ChatViewModel extraction: move business logic out of ComposeChatActivity god class

QA Gaps

  • P0 Missed-call follow-up E2E: missed-call notification / phone-state trigger reaches PokeClaw, follow-up message is sent to the caller, and the result/status is visible in the same chatroom
  • P1 C2: Auto-reply trigger E2E — needs 2nd device to send WhatsApp message to Girlfriend
  • P1 Release QA: verify locally signed 0.5.1+ public APK can upgrade in-place over the next signed public build once the stable key is installed in GitHub Actions
  • P1 M1-M12 QA: Cloud LLM complex tasks — done 2026-04-08, 10/12 PASS
  • P2 K6: Verify each Settings permission row leads to correct system settings page
  • P2 Settings layout QA: verify long local/cloud model names render cleanly on the Settings screen across Pixel/Samsung widths
  • P2 Download free space check — done 2026-04-08 (StatFs before download)
  • P1 Local vague-task UX: in Local Task mode, prompt-only behavior is correct, but vague requests like Copy that token to the clipboard currently hang instead of failing fast with a clear request for the missing content/details

Ideas / Research

  • Monetization: two-tier (dev=free open source, consumer=China APK + premium features)
  • YC application showcase
  • Layer 2 NLP Playbooks as "App Cards" like DroidRun
  • On-device LLM as competitive moat (first to ship with Gemma 4)

Done

Move completed items here with date.

  • 2026-04-08: Fix "Accessibility starting..." on every chat (A1-b)
  • 2026-04-08: Floating button IDLE→RUNNING in other apps (F3-b)
  • 2026-04-08: LiteRT-LM session conflict + GPU→CPU fallback (D1-a, D1-b)
  • 2026-04-08: Monitor permission check + auto-return after grant
  • 2026-04-08: Settings page: Notification Access row
  • 2026-04-08: Full QA pass 49/50 cases
  • 2026-04-08: Download free space check (StatFs before download)
  • 2026-04-08: Task detection keywords fix (check, compose, find, screen, notification, read my)
  • 2026-04-08: Compound task routing fix (skip Tier 1 for "and"/"then"/"after")
  • 2026-04-08: M1-M12 QA: 10/12 PASS, 2 PARTIAL (M9 camera, M12 system dialog)
  • 2026-04-08: NLP Playbooks Layer 2: 5 playbooks (Search, Settings, Email, Screen, Notifications)
  • 2026-04-08: Tinder research: UI structure documented, workflow designed, needs login
  • 2026-04-10: Chat bubble timestamps — IG-style per-message footer landed for user + assistant bubbles, with hidden timestamp metadata persisted in markdown history so relaunch/reload keeps stable times