Fixes

• Feishu GUI runner now drives step-by-step. The chat bubble updates with the current action on every step instead of staying frozen on "正在执行任务...". Each step has a 5-minute timeout so a wedged LLM call can't hang the loop forever.
• Clean error surfacing for Feishu replies. SDK errors are decoded to code: message instead of the previous Error@<hash>. Pre-flight failures (missing Vision creds, no device-control auth) now reply to Feishu with a human-readable reason instead of silently bailing. Image-reply failures land in Settings → 外部通道 → 调试日志 so users can see them without logcat.
• @_user_N mention stripping. Was hardcoded to @_user_1 ; now regex-matches @_(user|chat)_N so @bot 发朋友圈 reaches PhoneAgent as just 发朋友圈.
• Live screenshot fallback when image-reply is enabled but trace recording is off — users still get a final-state image instead of silent nothing.
• IME enabled check via the Android Framework API instead of ime list -s. Users who'd already enabled "ClawGUI Input" via system settings would see "未启用" in our panel because the shell command needed Shizuku/wadb auth to succeed. Settings → 输入法 also dropped the manual switcher button — agent handles the switching itself.

Chore

• Repo cleanup: retired the v1 clawgui-app/ (Java client) and promoted clawgui-app-ng/ into the natural clawgui-app/ path. Android applicationId stays com.clawgui.ng so existing v0.3.0 installs upgrade in place.

Structured Plan + execution trace

• PhoneAgent now maintains a user-visible task plan through a 6-state machine — PENDING / IN_PROGRESS / DONE / SKIPPED / FAILED / BLOCKED.
• The model emits incremental <plan>{"ops":[...]}</plan> blocks each step (init / update / insert_after / remove).
• A PlanCard renders inside the assistant bubble; status transitions fade + scale, the IN_PROGRESS row pulses.
• New ActionTraceList: every step shown as a chip (index / name / args / status badge), with a spinner placeholder while waiting on the next LLM round-trip.

Floating Plan + Trace overlay

• While a task runs, a semi-transparent draggable card floats above the host app so the user can keep an eye on progress without flipping back to ClawGUI.
• 3 collapse stages: 20dp blue dot → 1-line chip → full panel. Tap cycles, long-press drags.
• Auto-hides on finish / cancel / error. ×ばつ dismisses for the current run.
• Settings → Floating Panel: enable toggle, alpha slider (40–100%), permission prompt.

Mid-task Ask

• The agent now proactively asks the user when uncertain: missing parameters, multiple candidates, sensitive operations, no clear next step, fuzzy task scope.
• The input field is embedded directly in the floating panel — no need to switch back to ClawGUI to answer. IME pops automatically.
• The answer is injected as a system-role context turn; PhoneAgent resumes immediately.

Image attachments + multimodal

• + button in the input bar opens a sheet: pick from gallery or take a photo. Up to 3 attachments.
• Thumbnails render inside the user bubble; tap for a full-screen Lightbox (pinch-zoom + pan).
• Images are treated as task inputs (not just reference): auto-exported to Pictures/ClawGUI/ so any host app's picker can find them.
• Brain falls back to the Vision provider when it can't see images. ImageInsight runs a generic content-summary + extractable-text pre-pass.

Polish

• Follow-up suggestions: 3 one-tap chips under the latest assistant reply that drop a draft into the input box.
• Thinking separation: Brain reasoning_content and inline <think> tags are split off into a collapsible panel. PhoneAgent bubbles show per-step thinking when you peek a trace row.
• Appearance: Light / Dark / System in settings now actually drives the theme (the UI was there but unwired).
• Return-to-ClawGUI: task finish brings the chat back to the foreground via NEW_TASK + REORDER_TO_FRONT + a PendingIntent fallback that works on Honor / MIUI.

Robustness

• AutoGLM parser hardening: tolerates stray <plan> blocks, natural-language preambles, markdown fences, full-width punctuation (【】, ,), single/double quotes, whitespace variants (do ( / do(action='), JSON-shape fallbacks, and nested quotes.
• Step-1 action guard: if the model violates the "step 1 may only Launch / Home / Ask / finish" rule, the runtime intercepts and injects a system feedback turn — no more poking ClawGUI's own UI.
• Plan protocol fault-tolerance: any malformed op is dropped silently; the main task still moves forward on <answer>.
• Parse-failure retry: first parse error gets one retry with a system correction; only the second consecutive failure bails.
• Overlay lifecycle fix: LifecycleRegistry is single-shot after ON_DESTROY, which was why the floating panel never came back on the second task — each show() now allocates a fresh owner.

Prompt overhaul

• System prompt compressed from 200 → ~80 lines, format contract pulled to the top so the model's attention lands on it.
• Step 1 is pure planning (no screenshot; the prompt no longer hints there's a screen), restricted to Launch / Home / Ask / finish.
• Explicit "pick the app first" framing with concrete mappings (WeChat / Moments / Xiaohongshu / Weibo / Douyin / Meituan / ...).
• Tells the model to ignore ClawGUI's own floating panel — it's UI for the user, not part of the task.
• Ask reframed as "ask when uncertain, not only when something breaks", with 5 trigger scenarios + 4 counter-examples + a 3-per-task cap.

What's Changed

• feat(clawgui-eval): add Kimi K2.5 evaluation support by @ZichenWen1 in #1
• fix: handle reasoning_content in streaming for reasoning models by @ekkoitac in #8
• fix(agent): fix issues in webui.py by @jucve in #9
• feat: add clawgui-app on-device deployment module by @gta886 in #14
• docs cleanup, remove diagnostic export, refine Roadmap wording by @lhppppp in #15
• fix(clawgui-app): restore source files swallowed by root .gitignore by @lhppppp in #16
• Remove some configs, as it causes the system to freeze. by @HangFang6 in #17

New Contributors

• @ZichenWen1 made their first contribution in #1
• @ekkoitac made their first contribution in #8
• @jucve made their first contribution in #9
• @gta886 made their first contribution in #14
• @lhppppp made their first contribution in #15
• @HangFang6 made their first contribution in #17

Full Changelog: https://github.com/ZJU-REAL/ClawGUI/commits/clawgui-app-v0.2.0

Contributors

• @gta886
• @HangFang6
• @lhppppp
• @ZichenWen1
• @jucve
• @ekkoitac