Sycun · Oct 12, 2026 · Oct 16, 2026
diff --git a/config.example.yaml b/config.example.yaml
@@ -579,6 +579,15 @@ agents_dir: agents
 # 系统会从该目录加载所有 .yaml 格式的角色配置文件
 # 每个角色应创建独立的配置文件，例如：roles/CTF.yaml, roles/默认.yaml 等
 roles_dir: roles # 角色配置文件目录（相对于配置文件所在目录）
+# ============================================
+# 任务经验草稿（实验性，仅草稿）
+# ============================================
+learning:
+  enabled: false # 实验性：任务结束后后台提炼经验草稿，不自动启用 Skill；修改后重启
+  max_jobs_per_day: 20 # UTC 自然日作业上限（重试最多 3 次），后台单并发
+  timeout_seconds: 60 # 单次模型提炼超时，最大 120 秒
+  max_input_runes: 12000 # 证据正文总字符预算，最大 32000；每条最多 2000 字符
+
 # ============================================
 # 项目管理与事实黑板
 # ============================================

diff --git a/docs/zh-CN/configuration.md b/docs/zh-CN/configuration.md
@@ -147,6 +147,22 @@ multi_agent:
 
 `agents_dir` 指向 Markdown 子代理目录。单个代理可在 front matter 中设置 `tools`、`bind_role`、`max_iterations`。
 
+## 任务经验草稿
+
+```yaml
+learning:
+  enabled: false
+  max_jobs_per_day: 20
+  timeout_seconds: 60
+  max_input_runes: 12000
+```
+
+- `enabled`：实验性功能，默认关闭。开启后任务结束时在后台提炼经验草稿，只落草稿，不自动启用任何 Skill；修改后需重启。
+- `max_jobs_per_day`：UTC 自然日内的草稿作业上限，默认 20，上限 200。
+- `timeout_seconds`：单次模型提炼超时，默认 60 秒，上限 120 秒。
+- `max_input_runes`：证据正文总字符预算，默认 12000，上限 32000，单条最多 2000 字符。
+- 三项数值省略或填 0 时取默认值，超过上限按上限截断。后台单并发执行，同一作业失败最多重试 3 次后放弃。
+
 ## 工具与 MCP
 
 ```yaml

diff --git a/internal/app/app.go b/internal/app/app.go
@@ -1027,6 +1027,7 @@ func setupRoutes(
 		// Eino ADK 单代理（ChatModelAgent + Runner；不依赖 multi_agent.enabled）
 		protected.POST("/eino-agent", agentHandler.EinoSingleAgentLoop)
 		protected.POST("/eino-agent/stream", agentHandler.EinoSingleAgentLoopStream)
+		protected.GET("/conversations/:id/learning", agentHandler.ListConversationLearning)
 		protected.GET("/hitl/pending", agentHandler.ListHITLPending)
 		protected.GET("/hitl/logs", agentHandler.ListHITLLogs)
 		protected.DELETE("/hitl/logs", agentHandler.DeleteHITLLogs)

diff --git a/internal/config/config.go b/internal/config/config.go
@@ -52,6 +52,7 @@ type Config struct {
 	Project     ProjectConfig         `yaml:"project,omitempty" json:"project,omitempty"`
 	Vision      VisionConfig          `yaml:"vision,omitempty" json:"vision,omitempty"`
 	Update      UpdateConfig          `yaml:"update,omitempty" json:"update,omitempty"`
+	Learning    LearningConfig        `yaml:"learning,omitempty" json:"learning,omitempty"`
 }
 
 // UpdateConfig 一键更新的更新源（可选）：一个仓库地址，更新跟随该仓库的默认分支。

diff --git a/internal/config/learning.go b/internal/config/learning.go
@@ -0,0 +1,23 @@
+package config
+
+// LearningConfig enables the experimental, draft-only post-task learner.
+// Settings are captured at startup; changing them requires a restart.
+type LearningConfig struct {
+	Enabled        bool `yaml:"enabled" json:"enabled"`
+	MaxJobsPerDay  int  `yaml:"max_jobs_per_day,omitempty" json:"max_jobs_per_day,omitempty"`
+	TimeoutSeconds int  `yaml:"timeout_seconds,omitempty" json:"timeout_seconds,omitempty"`
+	MaxInputRunes  int  `yaml:"max_input_runes,omitempty" json:"max_input_runes,omitempty"`
+}
+
+func (c LearningConfig) DailyLimit() int { return boundedLearningValue(c.MaxJobsPerDay, 20, 200) }
+func (c LearningConfig) Timeout() int    { return boundedLearningValue(c.TimeoutSeconds, 60, 120) }
+func (c LearningConfig) InputLimit() int { return boundedLearningValue(c.MaxInputRunes, 12000, 32000) }
+func boundedLearningValue(n, fallback, maximum int) int {
+	if n <= 0 {
+		return fallback
+	}
+	if n > maximum {
+		return maximum
+	}
+	return n
+}
diff --git a/internal/config/learning_test.go b/internal/config/learning_test.go
@@ -0,0 +1,23 @@
+package config
+
+import (
+	"gopkg.in/yaml.v3"
+	"testing"
+)
+
+func TestLearningOptInAndHardLimits(t *testing.T) {
+	var cfg Config
+	if err := yaml.Unmarshal([]byte("learning:\n  enabled: true\n  max_jobs_per_day: 999\n  timeout_seconds: 999\n  max_input_runes: 999999\n"), &cfg); err != nil {
+		t.Fatal(err)
+	}
+	if !cfg.Learning.Enabled || cfg.Learning.DailyLimit() != 200 || cfg.Learning.Timeout() != 120 || cfg.Learning.InputLimit() != 32000 {
+		t.Fatalf("invalid bounds: %+v", cfg.Learning)
+	}
+	var defaultCfg Config
+	if err := yaml.Unmarshal([]byte("version: test\n"), &defaultCfg); err != nil {
+		t.Fatal(err)
+	}
+	if defaultCfg.Learning.Enabled || defaultCfg.Learning.DailyLimit() != 20 || defaultCfg.Learning.Timeout() != 60 || defaultCfg.Learning.InputLimit() != 12000 {
+		t.Fatal("learning not opt-in or defaults changed")
+	}
+}
diff --git a/internal/database/learning.go b/internal/database/learning.go
@@ -0,0 +1,279 @@
+package database
+
+import (
+	"database/sql"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/google/uuid"
+)
+
+const LearningExtractorVersion = "draft-v1"
+
+type LearningJob struct {
+	ID               string `json:"id"`
+	ConversationID   string `json:"conversationId"`
+	ProjectID        string `json:"projectId,omitempty"`
+	OwnerUserID      string `json:"-"`
+	AnchorMessageID  string `json:"anchorMessageId"`
+	RunStatus        string `json:"runStatus"`
+	Outcome          string `json:"outcome"`
+	Status           string `json:"status"`
+	Attempts         int    `json:"attempts"`
+	CandidateJSON    string `json:"candidateJson,omitempty"`
+	Reason           string `json:"reason,omitempty"`
+	PromptTokens     int    `json:"promptTokens"`
+	CompletionTokens int    `json:"completionTokens"`
+	CreatedAt        int64  `json:"createdAt"`
+	UpdatedAt        int64  `json:"updatedAt"`
+	LeaseToken       string `json:"-"`
+	StartRowID       int64  `json:"-"`
+	EndRowID         int64  `json:"-"`
+}
+
+type LearningEvidence struct {
+	ID        string `json:"id"`
+	Kind      string `json:"kind"`
+	Text      string `json:"text"`
+	Truncated bool   `json:"truncated,omitempty"`
+}
+
+// BeginLearningRun captures identity and the user-message anchor before the run
+// starts. No transcript copies or files are created.
+func (db *DB) BeginLearningRun(runID, conversationID string) error {
+	now := time.Now().Unix()
+	_, err := db.Exec(`INSERT OR IGNORE INTO learning_jobs
+    (id,extractor_version,conversation_id,anchor_message_id,project_id,owner_user_id,start_rowid,created_at,updated_at)
+    SELECT ?,?,c.id,m.id,c.project_id,c.owner_user_id,m.rowid,?,?
+    FROM conversations c JOIN messages m ON m.conversation_id=c.id
+    WHERE c.id=? AND m.role='user' AND COALESCE(c.owner_user_id,'')<>''
+    AND (SELECT COUNT(*) FROM learning_jobs WHERE status IN ('waiting','queued','processing'))<200
+    AND m.rowid=(SELECT MAX(rowid) FROM messages WHERE conversation_id=c.id AND role='user')`,
+		runID, LearningExtractorVersion, now, now, conversationID)
+	return err
+}
+
+// FinishLearningRun is idempotent. Unknown business outcomes never become
+// successful just because the execution completed.
+func (db *DB) FinishLearningRun(runID, status string) error {
+	state, reason := "skipped", "运行未正常收尾，不提炼技能"
+	if status == "completed" || status == "failed" {
+		state, reason = "queued", ""
+	}
+	_, err := db.Exec(`UPDATE learning_jobs SET run_status=?, status=?, reason=?,
+    end_rowid=COALESCE((SELECT MIN(rowid)-1 FROM messages WHERE conversation_id=learning_jobs.conversation_id AND role='user' AND rowid>start_rowid),
+      (SELECT MAX(rowid) FROM messages WHERE conversation_id=learning_jobs.conversation_id),start_rowid),updated_at=?
+    WHERE id=? AND status='waiting'`, status, state, reason, time.Now().Unix(), runID)
+	return err
+}
+
+// RecoverLearningRuns is called before accepting requests on process startup.
+// A crash cannot be interpreted as a completed task. Expired worker leases are
+// recovered by ClaimLearningJob without inventing a task result.
+func (db *DB) RecoverLearningRuns() error {
+	_, err := db.Exec(`UPDATE learning_jobs SET status='skipped',run_status='interrupted',reason='服务中断，缺少可靠运行终态',updated_at=? WHERE status='waiting'`, time.Now().Unix())
+	return err
+}
+
+const learningJobColumns = `id,conversation_id,COALESCE(project_id,''),owner_user_id,anchor_message_id,run_status,outcome,status,attempts,candidate_json,reason,prompt_tokens,completion_tokens,created_at,updated_at,lease_token,start_rowid,end_rowid`
+
+type learningScanner interface{ Scan(...any) error }
+
+func scanLearningJob(row learningScanner) (*LearningJob, error) {
+	j := new(LearningJob)
+	err := row.Scan(&j.ID, &j.ConversationID, &j.ProjectID, &j.OwnerUserID, &j.AnchorMessageID, &j.RunStatus, &j.Outcome, &j.Status, &j.Attempts, &j.CandidateJSON, &j.Reason, &j.PromptTokens, &j.CompletionTokens, &j.CreatedAt, &j.UpdatedAt, &j.LeaseToken, &j.StartRowID, &j.EndRowID)
+	return j, err
+}
+
+// ClaimLearningJob uses one atomic UPDATE, so competing workers cannot share a
+// lease. Daily limits count jobs, including retries of already-started jobs.
+func (db *DB) ClaimLearningJob(now time.Time, dailyLimit int, lease time.Duration) (*LearningJob, error) {
+	if dailyLimit < 1 || lease <= 0 {
+		return nil, fmt.Errorf("invalid learning queue limits")
+	}
+	if _, err := db.Exec(`UPDATE learning_jobs SET status='failed',lease_token='',lease_until=0,reason='提炼重试已耗尽',updated_at=? WHERE status='processing' AND lease_until<=? AND attempts>=3`, now.Unix(), now.Unix()); err != nil {
+		return nil, err
+	}
+	day := now.UTC().Format("2006-01-02")
+	j, err := scanLearningJob(db.QueryRow(`UPDATE learning_jobs SET status='processing',attempts=attempts+1,
+    lease_token=?,lease_until=?,started_day=?,updated_at=?
+    WHERE id=(SELECT id FROM learning_jobs WHERE extractor_version=? AND attempts<3
+      AND ((status='queued' AND next_attempt_at<=?) OR (status='processing' AND lease_until<=?))
+      AND (started_day=? OR (SELECT COUNT(*) FROM learning_jobs WHERE started_day=?)<?)
+      ORDER BY created_at,id LIMIT 1) RETURNING `+learningJobColumns,
+		uuid.NewString(), now.Add(lease).Unix(), day, now.Unix(), LearningExtractorVersion, now.Unix(), now.Unix(), day, day, dailyLimit))
+	if errors.Is(err, sql.ErrNoRows) {
+		return nil, nil
+	}
+	return j, err
+}
+
+// LearningSourceValid rechecks both the original scope and current permissions.
+func (db *DB) LearningSourceValid(j *LearningJob) bool {
+	var n int
+	if err := db.QueryRow(`SELECT COUNT(*) FROM learning_jobs j JOIN conversations c ON c.id=j.conversation_id
+    JOIN messages m ON m.id=j.anchor_message_id WHERE j.id=? AND j.lease_token=? AND j.status='processing'
+    AND c.owner_user_id=j.owner_user_id AND COALESCE(c.project_id,'')=COALESCE(j.project_id,'')`, j.ID, j.LeaseToken).Scan(&n); err != nil || n != 1 {
+		return false
+	}
+	a, err := db.ResolveRBACAccess(j.OwnerUserID)
+	if err != nil || !a.User.Enabled || !a.Permissions["chat:read"] || !db.UserCanAccessResource(j.OwnerUserID, a.PermissionScopes["chat:read"], "conversation", j.ConversationID) {
+		return false
+	}
+	return j.ProjectID == "" || (a.Permissions["project:read"] && db.UserCanAccessResource(j.OwnerUserID, a.PermissionScopes["project:read"], "project", j.ProjectID))
+}
+
+// LearningRunEvidence reads a bounded slice of source messages/tool results.
+// The message rowid boundary prevents a later run leaking into a delayed job.
+func (db *DB) LearningRunEvidence(j *LearningJob, maxRunes int) ([]LearningEvidence, error) {
+	if maxRunes < 1 || maxRunes > 32000 {
+		return nil, fmt.Errorf("invalid evidence budget")
+	}
+	perItem := maxRunes / 4
+	if perItem > 2000 {
+		perItem = 2000
+	}
+	if perItem < 1 {
+		perItem = 1
+	}
+	rows, err := db.Query(`SELECT id,kind,text FROM (
+    SELECT m.id,m.role AS kind,substr(m.content,1,?) AS text,m.rowid AS seq,0 AS sub
+      FROM messages m WHERE m.conversation_id=? AND m.rowid BETWEEN ? AND ?
+    UNION ALL
+    SELECT p.id,p.event_type,substr(COALESCE(p.data,p.message,''),1,?),m.rowid,p.rowid
+      FROM process_details p JOIN messages m ON m.id=p.message_id
+      WHERE m.conversation_id=? AND m.rowid BETWEEN ? AND ? AND p.event_type IN ('tool_result','error')
+    ) ORDER BY seq,sub LIMIT 100`, perItem+1, j.ConversationID, j.StartRowID, j.EndRowID, perItem+1, j.ConversationID, j.StartRowID, j.EndRowID)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	out := []LearningEvidence{}
+	for rows.Next() {
+		var e LearningEvidence
+		if err := rows.Scan(&e.ID, &e.Kind, &e.Text); err != nil {
+			return nil, err
+		}
+		r := []rune(e.Text)
+		if len(r) > perItem {
+			r = r[:perItem]
+			e.Truncated = true
+		}
+		if len(r) > maxRunes {
+			r = r[:maxRunes]
+			e.Truncated = true
+		}
+		e.Text = string(r)
+		maxRunes -= len(r)
+		out = append(out, e)
+		if maxRunes <= 0 {
+			break
+		}
+	}
+	return out, rows.Err()
+}
+
+func (db *DB) CompleteLearningJob(j *LearningJob, status, candidate, hash, reason string, promptTokens, completionTokens int) error {
+	if status != "drafted" && status != "skipped" {
+		return fmt.Errorf("invalid learning result")
+	}
+	if !db.LearningSourceValid(j) {
+		status, candidate, hash, reason = "skipped", "", "", "来源已删除、范围变更或权限失效"
+	}
+	if status == "drafted" && !db.learningCandidateEvidenceValid(j, candidate) {
+		status, candidate, hash, reason = "skipped", "", "", "来源证据已删除或无效"
+	}
+	// Exact duplicate suppression is deliberately conversation-scoped; no
+	// implicit cross-project sharing or deletion dependencies are introduced.
+	if hash != "" {
+		if err := db.InvalidateLearningDrafts(j.ConversationID); err != nil {
+			return err
+		}
+		var count int
+		if err := db.QueryRow(`SELECT COUNT(*) FROM learning_jobs WHERE conversation_id=? AND id<>? AND content_hash=? AND status='drafted'`, j.ConversationID, j.ID, hash).Scan(&count); err != nil {
+			return err
+		}
+		if count > 0 {
+			status, candidate, hash, reason = "skipped", "", "", "相同会话已有相同技能草稿"
+		}
+	}
+	_, err := db.Exec(`UPDATE learning_jobs SET status=?,candidate_json=?,content_hash=?,reason=?,prompt_tokens=prompt_tokens+?,completion_tokens=completion_tokens+?,lease_token='',lease_until=0,updated_at=?
+    WHERE id=? AND lease_token=? AND status='processing'
+    AND (?='skipped' OR EXISTS(SELECT 1 FROM conversations c WHERE c.id=learning_jobs.conversation_id AND c.owner_user_id=learning_jobs.owner_user_id AND COALESCE(c.project_id,'')=COALESCE(learning_jobs.project_id,'')))`,
+		status, candidate, hash, reason, promptTokens, completionTokens, time.Now().Unix(), j.ID, j.LeaseToken, status)
+	return err
+}
+
+func (db *DB) RetryLearningJob(j *LearningJob, reason string, now time.Time) error {
+	status := "queued"
+	if j.Attempts >= 3 {
+		status = "failed"
+	}
+	_, err := db.Exec(`UPDATE learning_jobs SET status=?,reason=?,next_attempt_at=?,lease_token='',lease_until=0,updated_at=? WHERE id=? AND lease_token=? AND status='processing'`,
+		status, reason, now.Add(time.Minute).Unix(), now.Unix(), j.ID, j.LeaseToken)
+	return err
+}
+
+func (db *DB) ListLearningJobs(conversationID string) ([]LearningJob, error) {
+	if err := db.InvalidateLearningDrafts(conversationID); err != nil {
+		return nil, err
+	}
+	rows, err := db.Query(`SELECT `+learningJobColumns+` FROM learning_jobs j WHERE conversation_id=?
+    AND EXISTS(SELECT 1 FROM conversations c WHERE c.id=j.conversation_id AND c.owner_user_id=j.owner_user_id AND COALESCE(c.project_id,'')=COALESCE(j.project_id,''))
+    ORDER BY created_at DESC,id LIMIT 50`, strings.TrimSpace(conversationID))
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	jobs := []LearningJob{}
+	for rows.Next() {
+		j, err := scanLearningJob(rows)
+		if err != nil {
+			return nil, err
+		}
+		jobs = append(jobs, *j)
+	}
+	return jobs, rows.Err()
+}
+
+func (db *DB) learningCandidateEvidenceValid(j *LearningJob, candidate string) bool {
+	var refs struct {
+		IDs []string `json:"evidenceIds"`
+	}
+	if json.Unmarshal([]byte(candidate), &refs) != nil || len(refs.IDs) == 0 {
+		return false
+	}
+	var n int
+	err := db.QueryRow(`SELECT COUNT(*) FROM json_each(?,'$.evidenceIds') e WHERE
+    NOT EXISTS(SELECT 1 FROM messages m WHERE m.id=e.value AND m.conversation_id=? AND m.rowid BETWEEN ? AND ?)
+    AND NOT EXISTS(SELECT 1 FROM process_details p JOIN messages m ON m.id=p.message_id WHERE p.id=e.value AND m.conversation_id=? AND m.rowid BETWEEN ? AND ?)`,
+		candidate, j.ConversationID, j.StartRowID, j.EndRowID, j.ConversationID, j.StartRowID, j.EndRowID).Scan(&n)
+	return err == nil && n == 0
+}
+
+// InvalidateLearningDrafts also handles removal of individual evidence records.
+func (db *DB) InvalidateLearningDrafts(conversationID string) error {
+	_, err := db.Exec(`UPDATE learning_jobs SET status='skipped',candidate_json='',content_hash='',reason='来源证据已删除或范围变更',updated_at=?
+    WHERE conversation_id=? AND status='drafted' AND (
+      NOT EXISTS(SELECT 1 FROM conversations c WHERE c.id=learning_jobs.conversation_id AND c.owner_user_id=learning_jobs.owner_user_id AND COALESCE(c.project_id,'')=COALESCE(learning_jobs.project_id,''))
+      OR EXISTS(SELECT 1 FROM json_each(learning_jobs.candidate_json,'$.evidenceIds') e WHERE
+        NOT EXISTS(SELECT 1 FROM messages m WHERE m.id=e.value AND m.conversation_id=learning_jobs.conversation_id AND m.rowid BETWEEN start_rowid AND end_rowid)
+        AND NOT EXISTS(SELECT 1 FROM process_details p JOIN messages m ON m.id=p.message_id WHERE p.id=e.value AND m.conversation_id=learning_jobs.conversation_id AND m.rowid BETWEEN start_rowid AND end_rowid)))`, time.Now().Unix(), conversationID)
+	return err
+}
+
+func (db *DB) RecordLearningUsage(j *LearningJob, prompt, completion int) error {
+	if prompt < 0 || completion < 0 {
+		return fmt.Errorf("invalid usage")
+	}
+	_, err := db.Exec(`UPDATE learning_jobs SET prompt_tokens=prompt_tokens+?,completion_tokens=completion_tokens+? WHERE id=? AND lease_token=? AND status='processing'`, prompt, completion, j.ID, j.LeaseToken)
+	return err
+}
+
+func (db *DB) PruneLearningJobs(now time.Time) error {
+	_, err := db.Exec(`DELETE FROM learning_jobs WHERE created_at<? AND (status<>'processing' OR lease_until<=?)`, now.Add(-30*24*time.Hour).Unix(), now.Unix())
+	return err
+}