agentpage 0.0.38 → 0.0.39

package/README.md

@@ -1161,23 +1161,25 @@ loop 对本轮返回做以下处理：
 
 - 从当前快照直接执行，不复述任务
 - 任务按“剔除模型”推进（current + previous + this-round -> new remaining）
-- 禁止 `page_info` 作为规划手段
+- 效果检查：规划新动作前先确认上轮操作是否在快照中生效
+- 禁止 `page_info` 作为规划手段，禁止重复相同工具调用
 - 可见目标尽量同轮批量执行
 - DOM 会变化的动作执行后在下一轮继续
+- 停机前必须在快照中确认任务已完成
 - 统一输出 `REMAINING` 协议
 
 ### B. Round Messages（轮次状态层）
 
 由 `src/core/agent-loop/messages.ts` 构建，职责是把运行时状态传给模型：
 
-- `Current remaining instruction`
-- `Done steps (do NOT repeat)`
-- `Previous round planned task array`
-- `Effect verification` — 要求 AI 对比上轮操作与当前快照，确认每个操作是否生效；未生效时从同区域找其他目标而非重复
-- `Previous round model output (normalized)`
-- `Latest DOM snapshot`
+- `Remaining: <instruction>` — 当前剩余任务
+- `Done steps (do NOT repeat)` — 已完成步骤摘要
+- 关键行为强化（批量执行、禁 page_info、DOM 断轮、Effect check）
+- `Previous executed` + 效果提示（简短非阻塞）
+- `Previous model output` — 上轮模型输出摘要
+- `## Snapshot` — 最新 DOM 快照
 
-这层是“每轮变化”的动态上下文。
+这层是“每轮变化”的动态上下文，已大幅精简避免与 system prompt 规则重复。
 
 ### C. Loop Control（执行控制层）
 

package/dist/index.mjs

@@ -450,12 +450,10 @@ function buildCompactMessages(userMessage, trace, latestSnapshot, currentUrl, hi
 		const parts = [
 			userMessage,
 			"",
-			"## Progressive execution state",
-			"Current remaining instruction to execute this round:",
-			activeInstruction
+			`Remaining: ${activeInstruction}`
 		];
-		if (currentUrl) parts.push("", `URL: ${currentUrl}`);
-		if (latestSnapshot) parts.push("", "## Current page snapshot", "Apply task-reduction model directly from this snapshot. Do NOT restate the task.", "Use hash IDs (e.g. #a1b2c) from the snapshot as selector params.", "Do NOT call page_info (get_url/get_title/query_all/snapshot).", "Batch independent visible actions in one round.", "Build the minimal action array from current snapshot to finish this remaining instruction in one round whenever possible.", "For deterministic increase/decrease controls, compute delta from current visible value and issue exactly that many clicks in one round (e.g., +2 => two increase clicks). Do not overshoot then undo.", "If action changes DOM (open modal/navigate), stop that batch and continue next round.", "For dropdown/select fields, use dom with action=select_option (or fill on a select).", "If a needed list shows `... (N children omitted)` under a specific container, output `SNAPSHOT_HINT: EXPAND_CHILDREN #<containerRef>` and wait for next round snapshot.", "Stop rule: once requested state is reached, stop tool calls. If verification is needed, verify once and then output REMAINING: DONE.", allowAgentUiInteraction ? "User explicitly asked to operate AutoPilot UI. You may interact with chat input/send/dock only as requested." : "Do NOT interact with any AI chat UI elements (chat input, send button, dock). Only operate on the actual page content.", "Output one line: REMAINING: <new remaining task after this round> or REMAINING: DONE", wrapSnapshot(latestSnapshot));
+		if (currentUrl) parts.push(`URL: ${currentUrl}`);
+		if (latestSnapshot) parts.push("", "Use #hashID from snapshot. Do NOT call page_info (snapshot is auto-refreshed). Batch all visible actions in one round.", "Effect check: confirm previous actions' expected effects in current snapshot before planning new actions.", "If action changes DOM (modal/navigate), stop batch and continue next round.", "If a list shows `... (N children omitted)`, output `SNAPSHOT_HINT: EXPAND_CHILDREN #<ref>` and wait for next snapshot.", allowAgentUiInteraction ? "User explicitly asked to operate AutoPilot UI. You may interact with chat input/send/dock only as requested." : "Do NOT interact with any AI chat UI elements (chat input, send button, dock). Only operate on the actual page content.", "Output: REMAINING: <new remaining> or REMAINING: DONE", "", "## Snapshot", wrapSnapshot(latestSnapshot));
 		if (protocolViolationHint) parts.push("", protocolViolationHint);
 		messages.push({
 			role: "user",
@@ -478,42 +476,29 @@ function buildCompactMessages(userMessage, trace, latestSnapshot, currentUrl, hi
 	});
 	const hasErrors = trace.some((e) => hasToolError(e.result));
 	const contextParts = [
-		"## Execution context",
-		"Current remaining instruction:",
-		activeInstruction,
+		`Remaining: ${activeInstruction}`,
 		"",
-		"Task-reduction model:",
-		"Input: current remaining instruction + previous round executed actions + this-round actions.",
-		"Output: new remaining instruction after removing this-round actions.",
-		"Start from visible page state directly. Do NOT restate task. Do NOT output planning text.",
-		"Execute all independent visible sub-tasks in one round.",
-		"Do NOT act on elements not present in this snapshot yet.",
-		"If action changes DOM (open modal/navigate), stop after that batch and continue next round.",
-		"Do NOT call page_info (get_url/get_title/query_all/snapshot).",
-		"For dropdown/select fields, use dom with action=select_option (or fill on a select).",
-		"If a needed list shows `... (N children omitted)` under a specific container, output `SNAPSHOT_HINT: EXPAND_CHILDREN #<containerRef>` and wait for next round snapshot.",
-		"Build the minimal action array from current snapshot to finish this remaining instruction in one round whenever possible.",
-		"For deterministic increase/decrease controls, compute delta from current visible value and issue exactly that many clicks in one round (e.g., +2 => two increase clicks). Do not overshoot then undo.",
-		"Stop rule: once requested state is reached, stop tool calls. If verification is needed, verify once and then output REMAINING: DONE.",
-		allowAgentUiInteraction ? "User explicitly asked to operate AutoPilot UI. You may interact with chat input/send/dock only as requested." : "Do NOT interact with any AI chat UI elements (chat input, send button, dock). Only operate on the actual page content."
+		"Execute all independent visible sub-tasks in one round. Do NOT call page_info (snapshot is auto-refreshed).",
+		"Effect check: confirm previous actions' expected effects in snapshot before planning new actions.",
+		"Never repeat the same tool call on the same target. If no effect, try a different element.",
+		"If action changes DOM (modal/navigate), stop batch and continue next round.",
+		"If a list shows `... (N children omitted)`, output `SNAPSHOT_HINT: EXPAND_CHILDREN #<ref>` and wait.",
+		allowAgentUiInteraction ? "User explicitly asked to operate AutoPilot UI." : "Do NOT interact with AI chat UI elements.",
+		"Output: REMAINING: <new remaining> or REMAINING: DONE"
 	];
-	if (hasErrors) contextParts.push("", "The last step failed. Retry with a different approach, or skip and continue with other visible targets.");
-	else contextParts.push("", "If the goal is fully done, reply with a short summary (no tool calls).");
-	if (previousRoundTasks && previousRoundTasks.length > 0) {
-		contextParts.push("", "Previous round planned task array (already executed):", ...previousRoundTasks.map((task, index) => `${index + 1}. ${task}`));
-		contextParts.push("", "## Effect verification (MANDATORY before planning new actions)", "Compare the above executed actions against the CURRENT snapshot below:", "- Did each action produce the expected result? (e.g., click opened a dialog/navigated; fill changed the input value; select_option updated the selected item)", "- If an action had NO visible effect (page unchanged near the target area), do NOT repeat the same target.", "  Instead, look for a different actionable element nearby (sibling/ancestor in the same row/card/group) that has stronger interaction signals (click listeners, button/link semantics).", "- Only after confirming which previous actions succeeded, plan this round's new actions on the remaining unfinished parts.");
-	}
-	if (previousRoundPlannedTasks && previousRoundPlannedTasks.length > 0) contextParts.push("", "Previous round model planned task array (before execution):", ...previousRoundPlannedTasks.map((task, index) => `${index + 1}. ${task}`));
-	if (previousRoundModelOutput) contextParts.push("", "Previous round model output (normalized, for task reduction input):", previousRoundModelOutput);
-	contextParts.push("", "After this round, include one plain text line:", "REMAINING: <new remaining instruction after this-round actions>", "or REMAINING: DONE");
+	if (hasErrors) contextParts.push("", "Last step failed. Retry differently or skip to other targets.");
+	else contextParts.push("", "If fully done, reply summary only (no tools).");
+	if (previousRoundTasks && previousRoundTasks.length > 0) contextParts.push("", "Previous executed:", ...previousRoundTasks.map((task, index) => `${index + 1}. ${task}`), "If any had no visible effect, try a different nearby element instead of repeating.");
+	if (previousRoundPlannedTasks && previousRoundPlannedTasks.length > 0) contextParts.push("", "Previous planned:", ...previousRoundPlannedTasks.map((task, index) => `${index + 1}. ${task}`));
+	if (previousRoundModelOutput) contextParts.push("", "Previous model output:", previousRoundModelOutput);
 	const lastEntry = trace[trace.length - 1];
 	if (hasToolError(lastEntry.result)) {
 		const stripped = toContentString(lastEntry.result.content).replace(SNAPSHOT_REGEX, "").trim();
-		if (stripped && stripped.length < 300) contextParts.push("", "Last error: " + stripped);
+		if (stripped && stripped.length < 300) contextParts.push("", "Error: " + stripped);
 	}
 	if (currentUrl) contextParts.push("", `URL: ${currentUrl}`);
 	if (protocolViolationHint) contextParts.push("", protocolViolationHint);
-	if (latestSnapshot) contextParts.push("", "## Latest DOM snapshot", "Use hash IDs from this snapshot. Do NOT call page_info — this is already the latest.", wrapSnapshot(latestSnapshot));
+	if (latestSnapshot) contextParts.push("", "## Snapshot", wrapSnapshot(latestSnapshot));
 	messages.push({
 		role: "user",
 		content: contextParts.join("\n")
@@ -1711,52 +1696,34 @@ function buildSystemPrompt(params = {}) {
 		"You are AutoPilot, an AI agent controlling the current web page via tools.",
 		"",
 		"## Core Rules",
-		"- Work from CURRENT snapshot + CURRENT remaining task directly. Do not restate the request.",
-		"- Treat each round as task reduction:",
-		"  Input: (1) current remaining task, (2) previous round executed actions, (3) actions you execute this round.",
-		"  Output: new remaining task after removing this-round actions.",
-		"- Use only visible targets from snapshot. Use #hashID as selector. Do not guess CSS selectors.",
-		"- Only interactive elements (with events, inputs, buttons, links, etc.) carry #hashID. Elements without #hashID are context-only (labels, headings, text) and cannot be targeted.",
-		"- Snapshot tag in brackets may show ARIA role instead of HTML tag when it better describes the interaction pattern (e.g. [combobox] for input with role=\"combobox\", [slider] for div with role=\"slider\"). Treat the bracket tag as the primary interaction hint.",
-		"- listeners=\"...\" on snapshot indicates bound event handlers (see Listener Abbrevs below). Prefer targets with relevant listeners when multiple candidates look similar.",
-		"- Click targeting rule (MANDATORY): for click/navigation actions, prioritize elements with explicit click signals (listeners containing clk/pdn/mdn, onclick, native link/button semantics, or role=button/link).",
-		"- Do NOT click focus/hover-only nodes for navigation (e.g. listeners only like fcs/blr/men/mlv without click-related signals). Treat those as context labels unless no better actionable target exists.",
-		"- Correlation fallback: if a click produced no progress, in the next round choose the nearest actionable sibling/ancestor within the same semantic group (same row/card/form), such as adjacent repo path/link/button, instead of repeating the same ineffective target.",
-		"- Batch independent visible actions in one round. Do not split one form into many rounds unnecessarily.",
-		"- Strict input order (MANDATORY): before every fill/type/select_option, click or focus the SAME target immediately in the SAME round.",
-		"- Multi-field rule (MANDATORY): execute alternating pairs in one batch: focus/click field A -> fill/type A -> focus/click field B -> fill/type B.",
-		"- Build the minimal action array from CURRENT snapshot to satisfy the target in one round whenever possible.",
-		"- Do NOT run focus-only batches (e.g., focus A -> focus B). Each focused input/select target must be followed by its input/select action right away.",
-		"- Fixed sequence examples: dom.focus(#name) -> dom.fill(#name, \"new-name\") -> dom.focus(#desc) -> dom.fill(#desc, \"new-desc\"); dom.click(#select) -> dom.select_option(#select, ...).",
-		"- Deterministic delta rule: for increase/decrease steppers, compute target delta from visible current value and emit exactly |delta| clicks in one round (e.g., +2 => click increase twice). Never overshoot then undo.",
-		"- For check/uncheck, target the real input control (checkbox/radio), not nearby text/container nodes.",
-		"- Form batch rule: for one visible form, complete all independent fields in one round; do not fill one field then verify repeatedly.",
-		"- If an action will change DOM (open modal, navigate), stop after that action batch and continue next round with new snapshot.",
-		"- Do NOT call page_info (snapshot/query/get_url/get_title). Snapshot is already provided every round.",
-		"- For dropdown/select, use dom action=select_option (or fill on select).",
-		"- If a required list shows `... (N children omitted)` under a specific container, request focused expansion by outputting `SNAPSHOT_HINT: EXPAND_CHILDREN #<containerRef>`.",
-		"- After outputting snapshot expansion hint, wait for the next refreshed snapshot before further scrolling/clicking on that list.",
-		"- Verification whitelist: do NOT use get_text/get_attr to verify input/select values unless the user explicitly asks for verification.",
-		"- Stop rule: when the requested state is achieved, stop calling tools. If verification is requested, verify once and then return REMAINING: DONE (no repeated get_text/get_attr on the same target).",
-		"- Do NOT interact with AutoPilot UI unless user explicitly asks.",
+		"- Work from CURRENT snapshot + remaining task. Do not restate.",
+		"- Task reduction: (remaining, prev actions, this-round) → new remaining.",
+		"- Use #hashID from snapshot as selector. Do not guess CSS selectors.",
+		"- Only interactive elements carry #hashID; others are context-only and cannot be targeted.",
+		"- Bracket tag may show ARIA role ([combobox], [slider]) as primary interaction hint.",
+		"- listeners=\"...\" = bound event handlers (abbrevs below). Prefer targets with matching listeners.",
+		"- Click priority: clk/pdn/mdn, onclick, native link/button, role=button/link. Avoid focus/hover-only (fcs/blr/men/mlv only).",
+		"- No-effect fallback: try nearest actionable sibling/ancestor in same semantic group instead of repeating.",
+		"- Batch all independent visible actions per round. Build minimal action array. Complete all form fields together.",
+		"- Input order (MANDATORY): focus/click → fill/type/select_option per target. Multi-field: focus A→fill A→focus B→fill B.",
+		"- Do NOT run focus-only batches. Each focus must be immediately followed by its fill/type/select_option.",
+		"- Steppers: compute delta from visible value, click exactly |delta| times. Check/uncheck: target real input control.",
+		"- DOM-changing action (modal/navigate): stop batch, continue next round with new snapshot.",
+		"- Effect check: before planning new actions, confirm previous actions' expected effects are visible in current snapshot. If not, the action failed — try a different target instead of repeating.",
+		"- Do NOT call page_info — snapshot is auto-refreshed and provided every round. Do NOT use get_text/get_attr to read what is already visible in the snapshot.",
+		"- Never repeat the same tool call (same name + same args) on the same target. If it didn't work, try a different approach.",
+		"- Dropdown/select: use dom.select_option or fill.",
+		"- Omitted children: output `SNAPSHOT_HINT: EXPAND_CHILDREN #<ref>`, wait for next snapshot.",
+		"- Do NOT verify values unless user explicitly asks.",
+		"- Stop: when remaining task is fully achieved (confirmed in snapshot), output REMAINING: DONE with a summary.",
+		"- Do NOT interact with AutoPilot UI unless user asks.",
 		"",
 		"## Listener Abbrevs",
 		"clk=click dbl=dblclick mdn=mousedown mup=mouseup mmv=mousemove mov=mouseover mot=mouseout men=mouseenter mlv=mouseleave pdn=pointerdown pup=pointerup pmv=pointermove tst=touchstart ted=touchend kdn=keydown kup=keyup inp=input chg=change sub=submit fcs=focus blr=blur scl=scroll whl=wheel drg=drag drs=dragstart dre=dragend drp=drop ctx=contextmenu",
 		"",
-		"## Output Contract",
-		"- Return tool calls for this round.",
-		"- Also include one plain text line:",
-		"  REMAINING: <new remaining task after this round>",
-		"  or REMAINING: DONE",
-		"",
-		"## Minimal Example",
-		"Task: click button -> type \"abc\" in input -> send",
-		"Round1 execute: click button",
-		"Remaining: type \"abc\" in input -> send",
-		"Round2 execute: type \"abc\" in input",
-		"Remaining: send",
-		"Round3 execute: send",
-		"Remaining: DONE"
+		"## Output",
+		"Tool calls + one text line: REMAINING: <new remaining> or REMAINING: DONE",
+		"Example: Task A→B→C. Round1 do A → REMAINING: B→C. Round2 do B → REMAINING: C. Round3 do C → REMAINING: DONE"
 	].join("\n"));
 	const tools = params.tools ?? [];
 	if (tools.length > 0) {