refactor(observer): v7 step-aware AG Native DOM parser with data-testid/data-step-index based content extraction

- Replace CSS class-based scanning with [data-testid='conversation-view'] + [data-step-index] traversal
- New extractCleanStepText(): clone-and-strip buttons/SVG/icons before text extraction
- New extractStepContext(): step-container-aware context with header + code block
- NOISE_RE: block Material icon names, button labels, UI artifacts
- Auto DOM structure dump on first conversation-view detection
- Enhanced deep-inspect with step element + button inventory
- known-issues: document AG Native SDK API incompatibility
This commit is contained in:
Variet Worker
2026-04-12 06:14:46 +09:00
parent 70dc301dca
commit a4d7286bce
7 changed files with 766 additions and 472 deletions

View File

@@ -5,8 +5,13 @@
* Handles:
* - Response file watching (file-based bridge fallback)
* - Response processing (diff_review, DOM observer, step_probe paths)
* - Multi-strategy approval execution (RPC, VS Code commands, DOM click)
* - Multi-strategy approval execution (VS Code commands, RPC, DOM click)
* - Diff review Accept/Reject via VS Code commands
*
* STRATEGY ORDER (most reliable first):
* 0. antigravity.acceptAgentStep / rejectAgentStep — AG's own commands, always works
* 1. HandleCascadeUserInteraction RPC — cross-platform, needs stepIndex
* 2. DOM click trigger via HTTP bridge — fallback
*/
import * as vscode from 'vscode';
@@ -256,7 +261,7 @@ async function processResponseFile(filePath: string) {
} catch { }
}
// ═══ MULTI-STRATEGY APPROVAL (v2.1) ═══
// ═══ MULTI-STRATEGY APPROVAL (v3.0) ═══
const approved = resp.approved;
// ── diff_review: Accept all / Reject all ──
@@ -268,16 +273,10 @@ async function processResponseFile(filePath: string) {
button_index: resp.button_index,
step_type: pendingStepType,
});
} else if (isDomObserver) {
// DOM observer path: ALSO try RPC strategies (renderer click is unreliable)
const targetSession = sessionId || ctx.activeSessionId;
ctx.logToFile(`[RESPONSE] dom_observer → tryApprovalStrategies(${approved}, ${targetSession.substring(0, 8)}, type=${pendingStepType}, step=${pendingStepIndex})`);
const strategyResult = await tryApprovalStrategies(approved, targetSession, pendingStepType, pendingStepIndex);
ctx.logToFile(`[RESPONSE] dom strategy result: ${strategyResult}`);
} else {
// Step probe path: run ALL approval strategies
// ALL paths (dom_observer + step_probe) use same strategy pipeline
const targetSession = sessionId || ctx.activeSessionId;
ctx.logToFile(`[RESPONSE] step_probe → tryApprovalStrategies(${approved}, ${targetSession.substring(0, 8)}, type=${pendingStepType}, step=${pendingStepIndex})`);
ctx.logToFile(`[RESPONSE] → tryApprovalStrategies(${approved}, ${targetSession.substring(0, 8)}, type=${pendingStepType}, step=${pendingStepIndex})`);
const strategyResult = await tryApprovalStrategies(approved, targetSession, pendingStepType, pendingStepIndex);
ctx.logToFile(`[RESPONSE] strategy result: ${strategyResult}`);
}
@@ -307,9 +306,9 @@ async function processResponseFile(filePath: string) {
* Returns a string describing which method succeeded (or all failed).
*
* Strategy order (most reliable first):
* 1. HandleCascadeUserInteraction RPC (cross-platform, no focus)
* 2. VS Code accept/reject commands (focus-dependent)
* 3. Log failure for manual intervention
* 0. antigravity.acceptAgentStep / rejectAgentStep (AG VS Code commands — always works)
* 1. HandleCascadeUserInteraction RPC (cross-platform, needs stepIndex)
* 2. Renderer DOM Click via HTTP Bridge (fallback)
*/
export async function tryApprovalStrategies(approved: boolean, sessionId: string, stepType: string = '', stepIndex: number = -1): Promise<string> {
const action = approved ? 'APPROVE' : 'REJECT';
@@ -317,90 +316,153 @@ export async function tryApprovalStrategies(approved: boolean, sessionId: string
: (ctx.lastPendingStepIndex >= 0 ? ctx.lastPendingStepIndex : -1);
ctx.logToFile(`[APPROVAL] Starting ${action} strategies for session ${sessionId.substring(0, 8)} stepType=${stepType} stepIndex=${effectiveStepIndex}`);
// ── Dynamic Command Discovery (log what's available during WAITING state) ──
let approvalCmdList: string[] = [];
try {
const allCmds = await vscode.commands.getCommands(true);
const agCmds = allCmds.filter((c: string) => c.startsWith('antigravity.'));
approvalCmdList = agCmds.filter((c: string) => {
const lower = c.toLowerCase();
return lower.includes('accept') || lower.includes('reject') || lower.includes('approve')
|| lower.includes('terminal') || lower.includes('run') || lower.includes('step')
|| lower.includes('cascade') || lower.includes('action');
});
ctx.logToFile(`[APPROVAL-CMD-CHECK] ${agCmds.length} total, ${approvalCmdList.length} approval-related:`);
for (const c of approvalCmdList) {
ctx.logToFile(`[APPROVAL-CMD-CHECK] → ${c}`);
// ══════════════════════════════════════════════════════════
// STRATEGY 0: SDK-verified AG commands (step_type-aware dispatch)
//
// From SDK index.js (verified command mapping):
// antigravity.agent.acceptAgentStep — code edits, file writes
// antigravity.agent.rejectAgentStep — reject code edits
// antigravity.command.accept — non-terminal commands (Run, Allow, etc.)
// antigravity.command.reject — reject non-terminal commands
// antigravity.terminalCommand.accept — terminal commands
// antigravity.terminalCommand.reject — reject terminal commands
// antigravity.terminalCommand.run — run terminal commands
//
// These operate on the currently focused/active step — no stepIndex needed!
// ══════════════════════════════════════════════════════════
{
const typeLower = stepType.toLowerCase().replace('cortex_step_type_', '');
// Determine which SDK command pair to use based on step_type
let acceptCmd: string;
let rejectCmd: string;
if (typeLower.includes('code_edit') || typeLower.includes('write_to_file')
|| typeLower.includes('propose_code') || typeLower.includes('write_cascade_edit')
|| typeLower === 'diff_review') {
// Code edits → agent step commands
acceptCmd = 'antigravity.agent.acceptAgentStep';
rejectCmd = 'antigravity.agent.rejectAgentStep';
} else if (typeLower.includes('run_command') || typeLower.includes('shell_exec')
|| typeLower.includes('send_command_input')) {
// Terminal commands → terminal command pair
acceptCmd = 'antigravity.terminalCommand.accept';
rejectCmd = 'antigravity.terminalCommand.reject';
} else if (typeLower === 'command' || typeLower.includes('permission')
|| typeLower.includes('browser') || typeLower.includes('mcp')
|| typeLower.includes('extension_code') || typeLower.includes('subagent')
|| typeLower.includes('open_browser') || typeLower.includes('read_url')
|| typeLower.includes('invoke_subagent')) {
// Non-terminal commands (Run, Allow, etc.) → command pair
acceptCmd = 'antigravity.command.accept';
rejectCmd = 'antigravity.command.reject';
} else {
// Unknown type — try all three in order
acceptCmd = 'antigravity.command.accept';
rejectCmd = 'antigravity.command.reject';
}
const primaryCmd = approved ? acceptCmd : rejectCmd;
ctx.logToFile(`[APPROVAL-0] stepType="${stepType}" → ${primaryCmd}`);
try {
await vscode.commands.executeCommand(primaryCmd);
ctx.logToFile(`[APPROVAL-0] ✅ ${primaryCmd} SUCCESS`);
return `SDK:${primaryCmd}`;
} catch (e: any) {
ctx.logToFile(`[APPROVAL-0] ❌ ${primaryCmd} failed: ${e.message?.substring(0, 200)}`);
}
// Fallback: if the primary type-specific command failed, try the other pairs
const fallbackPairs = [
approved ? 'antigravity.command.accept' : 'antigravity.command.reject',
approved ? 'antigravity.agent.acceptAgentStep' : 'antigravity.agent.rejectAgentStep',
approved ? 'antigravity.terminalCommand.accept' : 'antigravity.terminalCommand.reject',
].filter(cmd => cmd !== primaryCmd); // skip already-tried
for (const fallbackCmd of fallbackPairs) {
try {
ctx.logToFile(`[APPROVAL-0-FB] Trying ${fallbackCmd}...`);
await vscode.commands.executeCommand(fallbackCmd);
ctx.logToFile(`[APPROVAL-0-FB] ✅ ${fallbackCmd} SUCCESS`);
return `SDK-FB:${fallbackCmd}`;
} catch (e: any) {
ctx.logToFile(`[APPROVAL-0-FB] ❌ ${fallbackCmd}: ${e.message?.substring(0, 100)}`);
}
}
} catch (e: any) {
ctx.logToFile(`[APPROVAL-CMD-CHECK] error: ${e.message}`);
}
// ══════════════════════════════════════════════════════════
// STRATEGY 0-PROTO: Correct proto-based RPC (decoded from AG source)
// STRATEGY 1: HandleCascadeUserInteraction RPC
// Now supports BOTH approve AND reject.
// Requires valid stepIndex for most step types.
// ══════════════════════════════════════════════════════════
if (ctx.sdk && approved && effectiveStepIndex >= 0) {
// Build interaction sub-message based on step_type
if (ctx.sdk && effectiveStepIndex >= 0) {
const typeLower = stepType.toLowerCase().replace('cortex_step_type_', '');
let interactionPayload: Record<string, any> = {};
// Code edit steps — use dedicated RPC
if (typeLower.includes('code_edit') || typeLower.includes('write_to_file') || typeLower.includes('propose_code') || typeLower.includes('write_cascade_edit')) {
// CODE EDIT: Uses acknowledgeCodeActionStep RPC (correct AG LS method)
try {
ctx.logToFile(`[APPROVAL-CODE-EDIT] trying submitCodeAcknowledgement command`);
ctx.logToFile(`[APPROVAL-1-CODE] trying submitCodeAcknowledgement command`);
await vscode.commands.executeCommand('antigravity.prioritized.submitCodeAcknowledgement');
ctx.logToFile(`[APPROVAL-CODE-EDIT] ✅ submitCodeAcknowledgement OK`);
ctx.logToFile(`[APPROVAL-1-CODE] ✅ submitCodeAcknowledgement OK`);
return `CMD:submitCodeAcknowledgement(accept=${approved})`;
} catch {
ctx.logToFile(`[APPROVAL-CODE-EDIT] submitCodeAcknowledgement not available, trying RPC`);
ctx.logToFile(`[APPROVAL-1-CODE] submitCodeAcknowledgement not available, trying RPC`);
}
// Direct LS RPC with correct method name
try {
ctx.logToFile(`[APPROVAL-CODE-EDIT] acknowledgeCodeActionStep(cascadeId=${sessionId.substring(0, 8)}, accept=${approved}, stepIndices=[${effectiveStepIndex}])`);
ctx.logToFile(`[APPROVAL-1-CODE] acknowledgeCodeActionStep(cascadeId=${sessionId.substring(0, 8)}, accept=${approved}, stepIndices=[${effectiveStepIndex}])`);
const ackResult = await ctx.sdk.ls.rawRPC('acknowledgeCodeActionStep', {
cascadeId: sessionId,
accept: approved,
stepIndices: [effectiveStepIndex],
});
ctx.logToFile(`[APPROVAL-CODE-EDIT] ✅ SUCCESS: ${JSON.stringify(ackResult).substring(0, 200)}`);
ctx.logToFile(`[APPROVAL-1-CODE] ✅ SUCCESS: ${JSON.stringify(ackResult).substring(0, 200)}`);
return `RPC:acknowledgeCodeActionStep(accept=${approved})`;
} catch (e: any) {
ctx.logToFile(`[APPROVAL-CODE-EDIT] ❌ ${e.message.substring(0, 200)}`);
ctx.logToFile(`[APPROVAL-CODE-EDIT] falling back to HandleCascadeUserInteraction`);
interactionPayload = { runCommand: { confirm: true } };
ctx.logToFile(`[APPROVAL-1-CODE] ❌ ${e.message.substring(0, 200)}`);
// Fall through to generic HandleCascadeUserInteraction
interactionPayload = { runCommand: { confirm: approved } };
}
}
// Map step_type to interaction sub-message field
// CRITICAL FIX: Use `confirm: approved` (not always true) to support REJECT
if (typeLower.includes('run_command') || typeLower.includes('shell_exec')) {
interactionPayload = { runCommand: { confirm: true } };
interactionPayload = { runCommand: { confirm: approved } };
} else if (typeLower.includes('open_browser')) {
interactionPayload = { openBrowserUrl: { confirm: true } };
interactionPayload = { openBrowserUrl: { confirm: approved } };
} else if (typeLower.includes('send_command_input')) {
interactionPayload = { sendCommandInput: { confirm: true } };
interactionPayload = { sendCommandInput: { confirm: approved } };
} else if (typeLower.includes('read_url')) {
interactionPayload = { readUrlContent: { confirm: true } };
interactionPayload = { readUrlContent: { confirm: approved } };
} else if (typeLower.includes('mcp')) {
interactionPayload = { mcpTool: { confirm: true } };
interactionPayload = { mcpTool: { confirm: approved } };
} else if (typeLower.includes('invoke_subagent') || typeLower.includes('extension_code') || typeLower.includes('browser_subagent')) {
interactionPayload = { runExtensionCode: { confirm: true } };
interactionPayload = { runExtensionCode: { confirm: approved } };
} else if (typeLower.includes('file_permission')) {
const scope = typeLower.includes('conversation') ? 2 : 1;
interactionPayload = { filePermission: { allow: true, scope } };
if (typeLower.includes('deny')) {
interactionPayload = { filePermission: { allow: false, scope: 1 } };
} else {
const scope = typeLower.includes('conversation') ? 2 : 1;
interactionPayload = { filePermission: { allow: approved, scope } };
}
} else if (typeLower.includes('elicitation')) {
interactionPayload = { elicitation: {} };
} else if (typeLower === 'permission' || typeLower.includes('permission')) {
// DOM observer 'permission' type: browser_subagent Allow/Deny dialog
// Try runExtensionCode first (most common for JS execution permission)
interactionPayload = { runExtensionCode: { confirm: true } };
interactionPayload = { runExtensionCode: { confirm: approved } };
} else if (typeLower === 'command' || typeLower === '') {
// Generic command — most common case from DOM observer
interactionPayload = { runCommand: { confirm: approved } };
} else {
// Default: try run_command (most common)
interactionPayload = { runCommand: { confirm: true } };
// Default: try run_command
interactionPayload = { runCommand: { confirm: approved } };
}
const activeTrajectoryId = getTrajectoryId();
const protoVariants = [
// Variant A: camelCase with trajectoryId (proven working for run_command)
// Variant A: camelCase with trajectoryId
{
cascadeId: sessionId,
interaction: {
@@ -431,20 +493,17 @@ export async function tryApprovalStrategies(approved: boolean, sessionId: string
for (let i = 0; i < protoVariants.length; i++) {
try {
const payload = protoVariants[i];
ctx.logToFile(`[APPROVAL-PROTO-${i}] HandleCascadeUserInteraction(${JSON.stringify(payload).substring(0, 250)})`);
ctx.logToFile(`[APPROVAL-1-${i}] HandleCascadeUserInteraction(${JSON.stringify(payload).substring(0, 250)})`);
const rpcResult = await ctx.sdk.ls.rawRPC('HandleCascadeUserInteraction', payload);
ctx.logToFile(`[APPROVAL-PROTO-${i}] ✅ SUCCESS: ${JSON.stringify(rpcResult).substring(0, 200)}`);
return `RPC-PROTO-${i}:HandleCascadeUserInteraction(${typeLower})`;
ctx.logToFile(`[APPROVAL-1-${i}] ✅ SUCCESS: ${JSON.stringify(rpcResult).substring(0, 200)}`);
return `RPC-${i}:HandleCascadeUserInteraction(${typeLower},${action})`;
} catch (e: any) {
lastRpcError = e.message || '';
ctx.logToFile(`[APPROVAL-PROTO-${i}] ❌ ${lastRpcError.substring(0, 300)}`);
ctx.logToFile(`[APPROVAL-1-${i}] ❌ ${lastRpcError.substring(0, 300)}`);
}
}
// ── Auto-recovery: wrong-LS detection ──────────────────────
// All 3 proto variants failed. If the error is "input not registered",
// SDK is likely connected to wrong LS process. Attempt fixLSConnection
// and retry ONE time to avoid permanent failure.
if (ctx.fixLSConnection && lastRpcError.includes('input not registered')) {
ctx.logToFile('[APPROVAL] ⚠️ wrong-LS detected ("input not registered"), attempting LS fix...');
try {
@@ -453,10 +512,9 @@ export async function tryApprovalStrategies(approved: boolean, sessionId: string
ctx.logToFile('[APPROVAL] LS reconnected — retrying first proto variant...');
try {
const retryPayload = protoVariants[0];
ctx.logToFile(`[APPROVAL-RETRY] HandleCascadeUserInteraction(${JSON.stringify(retryPayload).substring(0, 250)})`);
const retryResult = await ctx.sdk.ls.rawRPC('HandleCascadeUserInteraction', retryPayload);
ctx.logToFile(`[APPROVAL-RETRY] ✅ SUCCESS: ${JSON.stringify(retryResult).substring(0, 200)}`);
return `RPC-RETRY:HandleCascadeUserInteraction(${typeLower})`;
return `RPC-RETRY:HandleCascadeUserInteraction(${typeLower},${action})`;
} catch (retryErr: any) {
ctx.logToFile(`[APPROVAL-RETRY] ❌ ${retryErr.message?.substring(0, 200)}`);
}
@@ -467,9 +525,14 @@ export async function tryApprovalStrategies(approved: boolean, sessionId: string
ctx.logToFile(`[APPROVAL] fixLSConnection error: ${fixErr.message?.substring(0, 200)}`);
}
}
} else if (ctx.sdk && effectiveStepIndex < 0) {
ctx.logToFile(`[APPROVAL-1] SKIPPED RPC: stepIndex=${effectiveStepIndex} (unknown) — Strategy 0 (VS Code command) was the primary attempt`);
}
// ── Strategy 2: Renderer DOM Click via HTTP Bridge (primary fallback) ──
// ══════════════════════════════════════════════════════════
// STRATEGY 2: Renderer DOM Click via HTTP Bridge (fallback)
// Sets a click trigger that the observer script polls and executes.
// ══════════════════════════════════════════════════════════
try {
const triggerAction = approved ? 'approve' : 'reject';
ctx.logToFile(`[APPROVAL-2] Setting clickTrigger=${triggerAction} for renderer DOM click`);
@@ -479,6 +542,6 @@ export async function tryApprovalStrategies(approved: boolean, sessionId: string
ctx.logToFile(`[APPROVAL-2] ❌ FAIL: ${e.message}`);
}
ctx.logToFile(`[APPROVAL] strategies complete — check logs for results`);
ctx.logToFile(`[APPROVAL] All strategies complete for ${action}`);
return `STRATEGIES_DONE:${action}`;
}