Simplify Todo workflow and refresh cards

src/config.ts

@@ -22,15 +22,18 @@ export const DEFAULT_LANGEXTRACT_MODEL = "deepseek/deepseek-v4-flash";
 export const DEFAULT_LANGEXTRACT_PROVIDER = "openai";
 export const DEFAULT_LANGEXTRACT_BASE_URL = "https://api.novita.ai/openai/v1";
 export const DEFAULT_TODO_EXTRACT_TIMEOUT_MS = 120_000;
-export const DEFAULT_TODO_EXTRACT_MAX_LLM_SESSIONS = 12;
 // STEP-11: extraction scope. sinceDays = only sessions whose endedAt/startedAt
 // falls within the last N days are eligible (the primary scope control). Max
 // interactions = per session, keep at most M most-recent interaction records
 // (a "turn": one user message through everything before the next). maxSessions
-// has no default cap; REST callers can still pass an explicit positive cap.
+// stays a backend safety cap (not surfaced as a setting).
 export const DEFAULT_TODO_EXTRACT_SINCE_DAYS = 7;
 export const DEFAULT_TODO_EXTRACT_MAX_INTERACTIONS = 10;
-export const DEFAULT_TODO_EXTRACT_MAX_SESSIONS = Number.POSITIVE_INFINITY;
+// Interactive safety cap on how many sessions one extraction pass touches. The
+// day window is the primary control, but with the LLM extractor each session is
+// a serial sidecar call (up to the per-call timeout), so a single "organize"
+// click must stay bounded. REST callers can still pass a larger maxSessions.
+export const DEFAULT_TODO_EXTRACT_MAX_SESSIONS = 8;
 const LEGACY_LANGEXTRACT_MODELS = new Set(["pa/gpt-5.5"]);
 export const WRITABLE_TODO_EXTRACT_KEYS = new Set([
   "AGENTMEMORY_TODO_EXTRACTOR",
@@ -43,7 +46,6 @@ export const WRITABLE_TODO_EXTRACT_KEYS = new Set([
   "AGENTMEMORY_TODO_EXTRACT_TIMEOUT_MS",
   "AGENTMEMORY_TODO_EXTRACT_SINCE_DAYS",
   "AGENTMEMORY_TODO_EXTRACT_MAX_INTERACTIONS_PER_SESSION",
-  "AGENTMEMORY_TODO_EXTRACT_MAX_LLM_SESSIONS",
 ]);
 
 let warnPremiumModelShown = false;
@@ -94,8 +96,6 @@ export function getTodoExtractorUserConfig(): Record<string, string | boolean> {
     LANGEXTRACT_THINKING_DEPTH: env["LANGEXTRACT_THINKING_DEPTH"] || "medium",
     AGENTMEMORY_TODO_EXTRACT_TIMEOUT_MS:
       env["AGENTMEMORY_TODO_EXTRACT_TIMEOUT_MS"] || String(DEFAULT_TODO_EXTRACT_TIMEOUT_MS),
-    AGENTMEMORY_TODO_EXTRACT_MAX_LLM_SESSIONS:
-      env["AGENTMEMORY_TODO_EXTRACT_MAX_LLM_SESSIONS"] || String(DEFAULT_TODO_EXTRACT_MAX_LLM_SESSIONS),
     AGENTMEMORY_TODO_EXTRACT_SINCE_DAYS:
       env["AGENTMEMORY_TODO_EXTRACT_SINCE_DAYS"] || String(DEFAULT_TODO_EXTRACT_SINCE_DAYS),
     AGENTMEMORY_TODO_EXTRACT_MAX_INTERACTIONS_PER_SESSION:

src/functions/action-candidates.ts

@@ -23,10 +23,13 @@ export interface ActionCandidateOptions {
 type BrowserTurn = { role?: string; text?: string };
 
 const READ_ONLY_TYPES = new Set(["file_read", "search", "web_fetch"]);
-const ACTION_VERB_PATTERN = "(修复|补充|实现|调整|验证|排查|定位|跟进|提交|创建|更新|移除|处理|合并|重试|重新(?:运行|跑))";
+const ACTION_VERB_PATTERN = "(修复|修正|补充|实现|调整|验证|排查|定位|跟进|整理|生成|上传|提交|创建|更新|移除|删除|处理|审查|合并|推送|构建|重试|重新(?:运行|跑))";
 const ENGLISH_ACTION_VERB = "(fix|add|update|create|remove|validate|retry|rerun|re-run|submit|handle|implement|investigate|debug|resolve)";
 const ENGLISH_FOLLOW_UP_PATTERN = new RegExp(`\\b(follow up|follow-up|need to|must)\\b.{0,80}\\b${ENGLISH_ACTION_VERB}\\b`, "i");
 const ACTION_DESCRIPTION_MAX_LENGTH = 180;
+const PROCESS_CHECK_PHRASES = /(?:做最后一次状态确认|最后一次状态确认|启动后做健康检查|做健康检查|健康检查已完成|确认工作区干净|确认当前分支|确认 PR 链接|服务可用|重启 Codex desktop app 后再测一次|重启 Codex desktop app 后再测|重启后再测一次|重启后再测)/i;
+const PROCESS_CHECK_REWRITEABLE = /(?:重启 Codex desktop app 后再测一次|重启 Codex desktop app 后再测|重启后再测一次|重启后再测)/i;
+const DURABLE_DELIVERABLE_TERMS = new RegExp(ACTION_VERB_PATTERN + `|\\b${ENGLISH_ACTION_VERB}\\b`, "i");
 
 function normalizeText(value: string | undefined): string {
   return String(value || "").replace(/\s+/g, " ").trim();
@@ -140,10 +143,18 @@ function cleanTitle(text: string): string {
   return (boundary > 0 ? head.slice(0, boundary) : head).join("").replace(/[，,；;：:\s]+$/u, "");
 }
 
+function stripActionPrefix(text: string): string {
+  return normalizeText(text)
+    .replace(/^(todo|fixme)\s*[:：-]\s*/i, "")
+    .replace(/^(下一步|后续|待办)\s*[:：-]?\s*/u, "")
+    .replace(/^(请|需要|必须)\s*/u, "")
+    .replace(/^(follow up|follow-up|fix)\s*[:：-]?\s*/i, "");
+}
+
 function titleFromText(text: string, fallback: string): string {
   const followUpMatch = text.match(new RegExp(`(?:下一步|后续)\\s*(?:请|需要|必须)?\\s*(${ACTION_VERB_PATTERN}[^。！？\\n]*)`, "u"))?.[1];
   const failureRepairMatch = text.match(new RegExp(`(?:验证未通过|验证失败|测试未通过|测试失败|command failed|exit code [1-9]\\d*|exited with code [1-9]\\d*)[，,。\\s]*(?:请|需要|必须)?\\s*(${ACTION_VERB_PATTERN}[^。！？\\n]*)`, "iu"))?.[1];
-  const explicit = text.match(/\b(?:TODO|FIXME)\b\s*[:：-]\s*([^。！？\n]+)/i)?.[1] ||
+  const explicit = text.match(/(?:^|[\s（(])(?:TODO|FIXME)\b\s*[:：-]\s*([^。！？\n]+)/i)?.[1] ||
     text.match(/待办\s*[:：-]\s*([^。！？\n]+)/u)?.[1] ||
     followUpMatch ||
     failureRepairMatch ||
@@ -171,10 +182,31 @@ function isStatusReport(text: string): boolean {
   return false;
 }
 
+function isProcessCheckText(text: string): boolean {
+  return PROCESS_CHECK_PHRASES.test(normalizeText(text));
+}
+
+function isRewriteableProcessCheck(text: string): boolean {
+  return PROCESS_CHECK_REWRITEABLE.test(normalizeText(text));
+}
+
+function isPureProcessCheck(text: string): boolean {
+  const normalized = normalizeText(text);
+  if (!normalized || !isProcessCheckText(normalized)) return false;
+  if (isRewriteableProcessCheck(normalized)) return false;
+  return !DURABLE_DELIVERABLE_TERMS.test(stripActionPrefix(normalized).replace(PROCESS_CHECK_PHRASES, ""));
+}
+
+function confidenceFor(reason: ActionCandidate["reason"], description: string): number {
+  if (isRewriteableProcessCheck(description)) return 0.56;
+  return reason === "follow_up" ? 0.62 : 0.72;
+}
+
 function sentenceReason(text: string): ActionCandidate["reason"] | null {
   if (isToolTrace(text) || isStatusReport(text)) return null;
+  if (isPureProcessCheck(text)) return null;
   if (/不应生成待办|不要生成待办|无需生成待办|不是待办|not an action|not a todo/i.test(text)) return null;
-  if (/\b(?:TODO|FIXME)\b\s*[:：-]\s*\S/i.test(text) || /待办\s*[:：-]\s*\S/u.test(text)) return "todo";
+  if (/(?:^|[\s（(])(?:TODO|FIXME)\b\s*[:：-]\s*\S/i.test(text) || /待办\s*[:：-]\s*\S/u.test(text)) return "todo";
   if (hasExplicitFollowUpAction(text)) return "follow_up";
   if (/\bblocked\b|未完成|被阻塞|阻塞/u.test(text)) return "blocked";
   if (hasExplicitFailureRepair(text)) return "validation_failed";
@@ -301,7 +333,7 @@ export function extractActionCandidatesFromObservations(
       source: "observation",
       sourceObservationIds: [obs.id],
       tags: tagsFor(reason),
-      confidence: reason === "follow_up" ? 0.62 : 0.72,
+      confidence: confidenceFor(reason, description),
       reason,
     });
   }
@@ -331,7 +363,7 @@ export function extractActionCandidatesFromTurns(
       source: "browser-review",
       sourceObservationIds: [],
       tags: tagsFor(reason),
-      confidence: reason === "follow_up" ? 0.62 : 0.72,
+      confidence: confidenceFor(reason, description),
       reason,
     });
   }

src/functions/todo-extract-langextract.py

@@ -21,30 +21,57 @@
 
 PROMPT = textwrap.dedent(
     """\
-    Extract actionable todos from AI agent session text.
+    Extract mature, user-facing todo cards from AI agent session text.
+    When refreshAction metadata is provided, use it as advisory context about the
+    existing card. Prefer a clearer title/description suggested by cleanup
+    metadata only when the nearby source quote supports it.
     Extract only UNRESOLVED, actionable items: explicit next actions, follow-ups,
     failed validations, blocked work, and in-progress work that still needs doing.
     DO NOT extract completed work, results, status reports, confirmations, or
     narration of what was already done (e.g. "…已通过", "…都能显示", "确认…完成").
-    Ignore background summaries, generic facts, read-only tool traces, and
-    anything without a source quote.
+    Ignore background summaries, generic facts, read-only tool traces, transient
+    agent process narration, and anything without a source quote.
     Use exact source text as extraction_text. Put fields in attributes:
     title, description, confidence, timeBucket, typeBucket, dedupeKey.
-    title must be a CRISP, SPECIFIC action summary — a concrete verb + the
-    specific object (+ the concrete target/outcome when it adds signal), so each
-    card's title alone makes clear what THIS todo is. Keep it short, but do not
-    drop the object or constraint just to hit a character target. DO NOT pad
-    with vague filler such as
+    The generated card must read like a mature todo app item, not a transcript
+    snippet. title must be a CRISP, SCANNABLE user-facing action summary — a
+    concrete verb + the semantic object (+ the concrete target/outcome when it
+    adds signal), so the title alone makes clear what THIS todo is. Preserve the
+    source language unless the source itself mixes languages. Keep it short, but
+    do not drop the user-relevant object or constraint just to hit a character
+    target.
+    Precise technical identifiers must be preserved, but they should usually go
+    in description rather than the title when they are long: branch names,
+    commit hashes, file paths, URLs, session ids, package names, and raw repos.
+    Prefer title "推送当前工作分支到远程仓库" with description
+    "分支：codex/todo-cleanup-flash-model。" over title
+    "推送 codex/todo-cleanup-flash-model 分支到远程仓库".
+    A good title may contain tightly related steps when they serve one outcome,
+    e.g. "修正目录显示文字（去掉重复编号）并更新页码缓存后重渲染".
+    description must be one concise sentence about the remaining user-relevant
+    work; do not start with "I will / 我会 / 现在 / 接下来".
+    DO NOT pad with vague filler such as
     "全面了解 / 了解现状 / 梳理现状 / 获取信息 / 进行 / 处理" — name the actual thing
     (the repo, the file, the bug, the command's purpose), not the process of
     "understanding" it. Prefer "克隆 AI-Todo 仓库" over "克隆仓库并全面了解其状况".
+    Do NOT turn agent workflow/status chores into todos unless a durable user
+    outcome remains. Examples to skip: "做最后一次状态确认", "启动后做健康检查",
+    "确认工作区干净", "服务可用", "健康检查已完成", "重启后再测一次" when it is just
+    suggested procedure or stale troubleshooting, and "我会/接下来/现在确认…".
+    Negative example: source text "我会做最后一次状态确认，确保工作区干净、当前分支和 PR 链接明确。服务可用，健康检查已完成。"
+    should produce no todo.
+    If text might be useful but is ambiguous or process-like, set confidence
+    below the direct-create threshold (0.55-0.81) so it goes to human review.
     dedupeKey must be a short STABLE slug of the core action+object (e.g.
     clone-aitodo-repo, read-project-config), the SAME for two todos that are the
     same task regardless of wording, so reworded duplicates collapse.
     Never use raw command JSON, file paths, screenshots, toolUseId/call IDs,
     shell flags, logs, or truncated trace fragments as title. If the only source
     text is a tool log, command payload, path, or JSON object, do not extract a
     todo.
+    Never extract dangling/truncated titles such as
+    "准备推送分支 codex/todo-cleanup-flash-model 到"; if context cannot repair
+    them into a complete, supported action, produce no todo.
     Never extract tool-call echo lines (starting with ⏺ or containing Bash(/Shell(),
     service-status reports (e.g. "服务可用", "Viewer:"/"Health:" URL lists), or
     git-ref fragments — these are not todos.
@@ -119,6 +146,11 @@ def main() -> int:
         print(json.dumps({"todos": []}, ensure_ascii=False))
         return 0
 
+    refresh_action = payload.get("refreshAction")
+    if isinstance(refresh_action, dict):
+        refresh_json = json.dumps(refresh_action, ensure_ascii=False, sort_keys=True)
+        text = f"[refreshAction]\n{refresh_json}\n\n{text}"
+
     examples = [
         lx.data.ExampleData(
             text="[obs:obs_1]\n后续需要修复 CI 失败，并重新跑测试。",
@@ -171,6 +203,74 @@ def main() -> int:
                 )
             ],
         ),
+        lx.data.ExampleData(
+            text="[obs:obs_4]\n下一步需要修复深色模式按钮对比度，避免主操作在暗色背景下不可读。",
+            extractions=[
+                lx.data.Extraction(
+                    extraction_class="todo",
+                    extraction_text="修复深色模式按钮对比度，避免主操作在暗色背景下不可读",
+                    attributes={
+                        "title": "修复深色模式按钮对比度",
+                        "description": "修复深色模式按钮对比度，避免主操作在暗色背景下不可读。",
+                        "confidence": 0.9,
+                        "timeBucket": "current",
+                        "typeBucket": "follow_up",
+                        "dedupeKey": "fix-dark-mode-button-contrast",
+                    },
+                )
+            ],
+        ),
+        lx.data.ExampleData(
+            text="[obs:obs_5]\n下一步需要推送 codex/todo-cleanup-flash-model 分支到远程仓库。",
+            extractions=[
+                lx.data.Extraction(
+                    extraction_class="todo",
+                    extraction_text="推送 codex/todo-cleanup-flash-model 分支到远程仓库",
+                    attributes={
+                        "title": "推送当前工作分支到远程仓库",
+                        "description": "推送当前工作分支到远程仓库。分支：codex/todo-cleanup-flash-model。",
+                        "confidence": 0.88,
+                        "timeBucket": "current",
+                        "typeBucket": "to_start",
+                        "dedupeKey": "push-current-branch",
+                    },
+                )
+            ],
+        ),
+        lx.data.ExampleData(
+            text="[obs:obs_7]\n下一步需要修正目录显示文字（去掉重复编号）并更新页码缓存后重渲染。",
+            extractions=[
+                lx.data.Extraction(
+                    extraction_class="todo",
+                    extraction_text="修正目录显示文字（去掉重复编号）并更新页码缓存后重渲染",
+                    attributes={
+                        "title": "修正目录显示文字（去掉重复编号）并更新页码缓存后重渲染",
+                        "description": "修正目录显示文字，去掉重复编号，并更新页码缓存后重渲染。",
+                        "confidence": 0.9,
+                        "timeBucket": "current",
+                        "typeBucket": "follow_up",
+                        "dedupeKey": "fix-toc-numbering-rerender",
+                    },
+                )
+            ],
+        ),
+        lx.data.ExampleData(
+            text="[obs:obs_6]\n建议处理顺序：2. 重启 Codex desktop app 后再测一次。",
+            extractions=[
+                lx.data.Extraction(
+                    extraction_class="todo",
+                    extraction_text="重启 Codex desktop app 后再测一次",
+                    attributes={
+                        "title": "验证重启后的 Codex desktop app",
+                        "description": "重启 Codex desktop app 后再验证问题是否仍存在。",
+                        "confidence": 0.65,
+                        "timeBucket": "recent",
+                        "typeBucket": "follow_up",
+                        "dedupeKey": "verify-codex-desktop-after-restart",
+                    },
+                )
+            ],
+        ),
     ]
     model_id = model_id_from_env()
     result = lx.extract(
@@ -243,7 +343,15 @@ class DummyLx:
         assert params["use_schema_constraints"] is False
         params = extract_kwargs(DummyLx, "custom/openai-compatible-model", DummyConfig)
         assert params["use_schema_constraints"] is False
-        assert "CRISP, SPECIFIC" in PROMPT
+        assert "CRISP, SCANNABLE" in PROMPT
+        assert "codex/todo-cleanup-flash-model" in PROMPT
+        assert "推送当前工作分支到远程仓库" in PROMPT
+        assert "修正目录显示文字" in PROMPT
+        assert "mature todo app item" in PROMPT
+        assert "做最后一次状态确认" in PROMPT
+        assert "Negative example" in PROMPT
+        assert "refreshAction metadata" in PROMPT
+        assert "0.55-0.81" in PROMPT
         assert "dedupeKey must be a short STABLE slug" in PROMPT
         assert "克隆 AI-Todo 仓库" in PROMPT
         print("ok")