Decoded HandleCascadeUserInteractionRequest protobuf schema from AG's
extension.js (message #162, base64 FileDescriptor 78KB).
Working payload (variant PROTO-0):
cascadeId + interaction.{trajectoryId, stepIndex, runCommand.confirm}
Changes:
- extension.ts: Added Strategy 0-PROTO with decoded proto RPC call
- extension.ts: Fixed processResponseFile to call tryApprovalStrategies()
instead of direct clickTrigger (was bypassing all strategies)
- extension.ts: Fixed false positive Run detection (sessionStalled reset
when step_probe confirms no WAITING)
- extension.ts: Moved lastPendingStepIndex to module scope
- extension.ts: Added activeTrajectoryId tracking from session init
- bot.py: Added MERGE detection + Discord message edit for command updates
- bot.py: Added _sent_commands tracking for merge detection
Proto RE methodology:
1. Found schema exports in AG extension.js
2. Located fileDesc() with base64 protobuf descriptor
3. Decoded 58KB raw proto, found message names
4. Extracted CascadeRunCommandInteraction.confirm field
5. Tested camelCase JSON via ConnectRPC = SUCCESS
- CMD-DISCOVERY: enumerate all antigravity.* commands at activation (72) and during WAITING state (119)
- APPROVAL-CMD-CHECK: re-check commands inside tryApprovalStrategies for dynamic registration
- Confirmed: ALL 7 SDK approval commands NOT REGISTERED in current AG build
- Confirmed: sendChatActionMessage, executeCascadeAction also NOT REGISTERED
- Replaced failed keyboard simulation (Strategy 2) with renderer DOM click approach:
- Added clickTrigger variable + GET /trigger-click HTTP endpoint
- Renderer polls /trigger-click every 1s, clicks Run/Accept button via DOM
- Updated known-issues.md with comprehensive findings
- Added devlog entry 20260309-002
- Fix: processResponseFile no longer deletes response files for DOM observer
approvals, allowing renderer pollResponse to find and serve them via HTTP
- Fix: Run button regex ^Run$ → ^Run to match 'Run Alt+⏎' button text
- Fix: BTN-DUMP diagnostic added to generateApprovalObserverScript (source)
- Doc: 2 new known issues (race condition, renderer script 3-location confusion)
- Doc: devlog entry #19
- latestToolCallStep RPC 기반 즉시 감지 (30초 stall → 5초 poll)
- DOM scan 범위: findPanel() → document.body 확장
- Accept all/Reject all 리뷰 바 패턴 추가
- Stall detection을 100초 fallback으로 약화
- extractToolCommand/extractToolDescription 헬퍼 추가
- known-issues 5건 신규 추가
- start/services workflow: Python 전체 경로 + services.md 로딩
#task-258 #task-262
- Replace sync XHR tryPing() with async fetch tryPingAsync() for port discovery
- Add ag-sdk JS file to product.json checksums in updateProductChecksums()
- Revert to inline script approach for jetski HTML (vscode-file:// blocks custom .js)
- Remove old external script tag cleanup, add inline markers
- Update known-issues with 3 new findings
- Add devlog entry #16
- vscode-file:// refuses custom .js files → inline script into HTML
- Random port → deterministic port from project name hash (gravity_control=34332)
- Hardcoded port in renderer script for immediate discovery
- Auto-update product.json SHA256 checksums after HTML modification
- Bump version 0.2.0 → 0.3.5
Root cause: VS Code commands (acceptAgentStep, terminalCommand.run etc)
return undefined silently but don't actually accept WAITING steps.
LS requires HTTPS + CSRF token for RPC calls.
New approach: Phase 1 tries SDK rawRPC (has CSRF auth) with
HandleCascadeUserInteraction + ResolveOutstandingSteps.
Phase 2 tries all VS Code commands as fallback.
All results logged to bridge/extension.log for debugging.
Also removed stall detection (fundamentally broken — stepCount
keeps incrementing from other tool calls during WAITING).
GetCascadeTrajectorySteps has 775-step hard limit and cannot see
WAITING steps beyond that. No RPC exists for direct WAITING detection.
New approach: if stepCount frozen for 2+ polls (~10s) while status is
RUNNING, treat as WAITING and write pending approval to Discord.
30s cooldown prevents duplicate pending messages.
Also removes the last GetCascadeTrajectorySteps call from Extension -
now only a single GetAllCascadeTrajectories RPC per 5s poll cycle.
EventMonitor was dual-polling GetCascadeTrajectorySteps every 2s via rawRPC,
which has a 775-step hard limit and generates connection errors on port change.
Changes:
- Remove entire SDK EventMonitor (onStepCountChanged, onNewConversation, etc.)
- Keep only GetAllCascadeTrajectories POLL at 5s interval
- Remove all sdk.monitor.stop() calls
- Unleash ERR_CONNECTION_REFUSED (127.0.0.1:1080) is Antigravity's own issue
