Summary

• Fixes the Android chat context meter so it renders live session context usage from gateway totalTokens / contextTokens metadata instead of mapping the thinking preset to a fixed percentage.
• Keeps the compact thinking control visible by showing the current thinking level next to the context usage label.
• Wires Android operator sessions into gateway session events (sessions.subscribe) and consumes sessions.changed, session.message, and chat.history.sessionInfo session metadata so the meter can update after normal chat activity.
• Treats missing, stale, or explicitly absent token usage as Context -- instead of displaying a stale or misleading percentage.
• Follow-up for ClawSweeper's session-state risk: newer session snapshots that omit context-usage fields now clear cached usage instead of silently preserving an older chat's token state.
• Out of scope: changing gateway token accounting semantics or redesigning the chat composer control.

Linked context

Closes #

Related #

Was this requested by a maintainer or owner?

No public maintainer issue. This was found while testing the Android chat UI: the composer showed the same Context 82% value across different selected chats.

Real behavior proof (required for external PRs)

• Behavior or issue addressed: Android chat composer context usage no longer stays at a fixed thinking-derived percentage across chats.
• Real environment tested: Android emulator running the third-party debug APK from this branch, paired to a throwaway local OpenClaw proof gateway with a synthetic Proof Main session.
• Exact steps or command run after this patch:
  • Rebased branch onto current upstream/main.
  • git diff --check
  • ANDROID_HOME=/home/nabu/Android/Sdk ANDROID_SDK_ROOT=/home/nabu/Android/Sdk ./gradlew :app:clean :app:testThirdPartyDebugUnitTest :app:assembleThirdPartyDebug --no-daemon --stacktrace
  • ANDROID_HOME=/home/nabu/Android/Sdk ANDROID_SDK_ROOT=/home/nabu/Android/Sdk ./gradlew :app:testPlayDebugUnitTest :app:assemblePlayDebug --no-daemon --stacktrace
  • Earlier connected proof on this branch: ./gradlew :app:installThirdPartyDebug --no-daemon --stacktrace, Android emulator pairing to a throwaway local proof gateway, opening the synthetic Proof Main chat, and capturing adb screenshots plus uiautomator dumps.
  • Earlier lint probe: ./gradlew :app:lintThirdPartyDebug --no-daemon --stacktrace.
• Gateway metadata used for live proof, redacted to only the fields relevant to this UI contract:

session=Proof Main
totalTokens=20500
totalTokensFresh=true
contextTokens=50000
thinkingLevel=high

• Evidence after fix:
  • Current rebased head: BUILD SUCCESSFUL for :app:clean :app:testThirdPartyDebugUnitTest :app:assembleThirdPartyDebug.
  • Current rebased head: BUILD SUCCESSFUL for :app:testPlayDebugUnitTest :app:assemblePlayDebug.
  • Focused regression coverage now verifies session metadata merge behavior clears old token usage when a newer session snapshot omits usage fields.
  • Focused regression coverage also verifies explicit stale metadata (totalTokensFresh=false) replaces previous fresh metadata and renders as unknown usage.
  • Emulator screenshot with fresh gateway metadata: Android composer shows Context 41% · high, matching 20500 / 50000 = 41%.
  • Emulator screenshot with stale gateway metadata: Android composer shows Context -- · high when totalTokensFresh=false.

Fresh metadata screenshot:

Stale metadata screenshot:

• Observed result after fix: context label is computed from fresh token usage and displays Context -- for stale or absent usage. Thinking level remains visible in the same control.
• What was not tested: full manual exploration against a private real user gateway/session was not tested; proof uses a throwaway local gateway and synthetic session to avoid exposing private data.
• Proof limitations or environment constraints: screenshots use a synthetic local proof session and a throwaway local gateway, not a private real user session. No personal chats, account data, tokens, local endpoints, or non-public session identifiers are included. The follow-up for missing usage metadata is covered by focused merge-policy tests because the visible UI outcome is the same safe Context -- state shown in the stale-metadata screenshot.
• Before evidence: source-reproducible on current main: the Android ChatContextMeter derives the percentage from thinkingLevel, with high mapped to 82%, so the displayed context value is independent of gateway totalTokens / contextTokens usage.

Tests and validation

Commands run on the current rebased head:

• git diff --check
• ANDROID_HOME=/home/nabu/Android/Sdk ANDROID_SDK_ROOT=/home/nabu/Android/Sdk ./gradlew :app:clean :app:testThirdPartyDebugUnitTest :app:assembleThirdPartyDebug --no-daemon --stacktrace
• ANDROID_HOME=/home/nabu/Android/Sdk ANDROID_SDK_ROOT=/home/nabu/Android/Sdk ./gradlew :app:testPlayDebugUnitTest :app:assemblePlayDebug --no-daemon --stacktrace

Connected Android proof also run earlier on this branch:

• ./gradlew :app:installThirdPartyDebug --no-daemon --stacktrace
• Android emulator + adb connected-runtime screenshot proof against a throwaway local proof gateway.

Regression coverage added:

• Active session token usage resolution.
• Canonical main alias resolution.
• Missing context budget fallback.
• Stale token usage fallback.
• Overfull usage clamping.
• Session merge clears old usage when a newer snapshot omits context-usage metadata.
• Session merge applies explicit stale usage metadata instead of retaining older fresh usage.

Known validation note:

• :app:lintThirdPartyDebug still fails on existing unrelated lint errors outside this PR's touched code:
  • OnboardingFlow.kt: POST_NOTIFICATIONS API-level lint.
  • DeviceHandler.kt: package visibility lint.
  • GatewayDiscovery.kt: two @RequiresApi lint findings.
  • app/src/debug/AndroidManifest.xml: exported receiver permission lint.

Risk checklist

Did user-visible behavior change? (Yes/No)

Yes.

Did config, environment, or migration behavior change? (Yes/No)

No.

Did security, auth, secrets, network, or tool execution behavior change? (Yes/No)

No security/auth change. The Android operator connection now best-effort subscribes to existing gateway session events after connect.

What is the highest-risk area?

Session metadata freshness in the Android chat UI.

How is that risk mitigated?

The controller now distinguishes snapshots that explicitly include context-usage metadata from snapshots that omit usage fields. Explicit fresh usage renders the percentage, explicit stale usage renders Context --, and newer snapshots with no usage fields clear previously cached token usage so the active chat cannot inherit stale context usage from an earlier event. Focused tests cover the meter decision logic and the session merge policy. The debug build installs on an emulator and the redacted emulator proof shows both fresh and stale gateway metadata states.

Current review state

What is the next action?

Android maintainer review and merge judgment on the clean rebased head.

What is still waiting on author, maintainer, CI, or external proof?

Waiting on maintainer review/merge judgment. CI is green on the pushed head, and ClawSweeper re-review is complete with status: ready for maintainer look, proof: sufficient, and no contributor-facing automated repair. ClawSweeper still flags the session-state risk for maintainer awareness because this path depends on session event freshness.

Which bot or reviewer comments were addressed?

Addressed ClawSweeper's proof request with redacted Android emulator proof showing the composer context meter updating from connected gateway totalTokens / contextTokens metadata and showing Context -- for stale usage. Also addressed ClawSweeper's session-state risk by clearing cached usage when newer session snapshots omit usage metadata, so Android prefers unknown usage over stale or wrong usage.

AI-assisted: implemented with Codex, then manually inspected and tested in a local Android/OpenClaw setup.