feat: refresh course through workflow and goal loops

web/src/data/annotations/s10.json

@@ -3,16 +3,16 @@
   "decisions": [
     {
       "id": "prompt-from-state",
-      "title": "The System Prompt Is Built from Runtime State",
-      "description": "Prompt sections are selected from real state: workspace, available tools, memory, and skills. The prompt becomes a product of the runtime rather than a single hardcoded string.",
+      "title": "Model Input Is Built from Runtime State",
+      "description": "Stable prompt sections and dynamic state are assembled at the model boundary: workspace, available tools, memory, and skills. Model input becomes a product of the runtime rather than a single hardcoded string.",
       "alternatives": "A static prompt is easier to inspect, but it goes stale as capabilities change.",
       "zh": {
-        "title": "系统提示词由运行时状态构建",
-        "description": "Prompt section 来自真实状态：workspace、可用工具、memory 和 skills。提示词是运行时的产物，而不是单个硬编码字符串。"
+        "title": "模型输入由运行时状态构建",
+        "description": "稳定 prompt section 与动态状态在模型边界组装：workspace、可用工具、memory 和 skills。模型输入是运行时的产物，而不是单个硬编码字符串。"
       },
       "ja": {
-        "title": "システムプロンプトはランタイム状態から作る",
-        "description": "workspace、利用可能ツール、memory、skills などの実状態から prompt section を選びます。プロンプトは単一の固定文字列ではなくランタイムの産物です。"
+        "title": "モデル入力はランタイム状態から作る",
+        "description": "安定した prompt section と workspace、利用可能ツール、memory、skills などの動的状態をモデル境界で組み立てます。モデル入力は単一の固定文字列ではなくランタイムの産物です。"
       }
     },
     {

web/src/data/annotations/s17.json

@@ -31,16 +31,16 @@
     },
     {
       "id": "identity-reinjection",
-      "title": "Identity Is Re-injected on Each Autonomous Turn",
-      "description": "Autonomous agents need a stable sense of who they are and what they are allowed to do. Re-injecting identity keeps a teammate from drifting into the lead's responsibilities.",
-      "alternatives": "A one-time identity prompt is shorter, but long-running loops are prone to context drift.",
+      "title": "Advanced Teaching Workaround: Identity Re-injection",
+      "description": "The simplified lesson re-injects identity after its compaction heuristic. This is optional teaching scaffolding; production compaction should preserve stable system instructions instead.",
+      "alternatives": "Keep identity in stable system instructions and preserve that boundary through compaction.",
       "zh": {
-        "title": "每个自治回合都重新注入身份",
-        "description": "自治 agent 需要稳定知道自己是谁、允许做什么。重新注入身份可以防止队友逐渐漂移到 lead 的职责上。"
+        "title": "进阶教学补丁：身份重注入",
+        "description": "简化课程在压缩启发式触发后重新注入身份。这是选学脚手架；生产实现应在压缩时保留稳定的 system 指令。"
       },
       "ja": {
-        "title": "各自律ターンでアイデンティティを再注入する",
-        "description": "自律エージェントには、自分が誰で何を許可されているかという安定した認識が必要です。アイデンティティを再注入することで、チームメイトがリードの責務へ漂うことを防ぎます。"
+        "title": "発展用の教育補助：アイデンティティ再注入",
+        "description": "簡略化した教材は圧縮ヒューリスティック後にアイデンティティを再注入します。これは任意の足場であり、本番実装では圧縮を越えて安定した system 指示を保持すべきです。"
       }
     }
   ]

web/src/data/annotations/s21.json

@@ -0,0 +1,47 @@
+{
+  "version": "s21",
+  "decisions": [
+    {
+      "id": "script-owns-fixed-orchestration",
+      "title": "Code Owns Fixed Orchestration",
+      "description": "When the stages and aggregation rules are known in advance, a workflow script makes the process parallel, reproducible, and inspectable without changing the main agent loop.",
+      "alternatives": "Letting the model choose every next step is more flexible, but slower and harder to resume for a fixed procedure.",
+      "zh": {
+        "title": "固定编排由代码负责",
+        "description": "当阶段与汇总规则事先确定时，workflow 脚本能让流程并行、可复现、可检查，同时不修改主 Agent 循环。"
+      },
+      "ja": {
+        "title": "固定された編成はコードが担う",
+        "description": "段階と集約ルールが事前に決まっているなら、workflow script は main Agent loop を変えずに処理を並列化し、再現可能で検査可能にする。"
+      }
+    },
+    {
+      "id": "semantic-journal-keys",
+      "title": "Semantic Keys Make Resume Independent of Completion Order",
+      "description": "Journal entries use stable call content rather than a shared completion counter. Concurrent calls can finish in any order and still map to the correct cached result.",
+      "alternatives": "Indexing by completion order is simpler, but replays the wrong result as soon as concurrent timing changes.",
+      "zh": {
+        "title": "语义键让恢复不依赖完成顺序",
+        "description": "Journal 用稳定的调用内容作为 key，而不是共享完成计数器。并发调用无论以什么顺序结束，都能命中正确缓存。"
+      },
+      "ja": {
+        "title": "意味キーで完了順序に依存せず再開する",
+        "description": "Journal は共有完了カウンタではなく安定した call 内容を key にする。並行 call の終了順が変わっても正しい cache result に対応できる。"
+      }
+    },
+    {
+      "id": "fail-the-workflow",
+      "title": "Orchestration Failures Propagate",
+      "description": "A failed stage, invalid structured result, corrupt journal, or exceeded run-wide limit fails the workflow instead of silently dropping an item and reporting success.",
+      "alternatives": "Best-effort collection can be useful for optional work, but it must be explicit rather than the default.",
+      "zh": {
+        "title": "编排故障必须向上传播",
+        "description": "阶段失败、结构化结果不合法、journal 损坏或超过全局限制时，workflow 直接失败，而不是静默丢项后仍报告成功。"
+      },
+      "ja": {
+        "title": "編成の失敗は上位へ伝播させる",
+        "description": "stage failure、無効な structured result、破損 journal、run-wide limit 超過は workflow を失敗させ、項目を黙って落として成功扱いしない。"
+      }
+    }
+  ]
+}

web/src/data/annotations/s22.json

@@ -0,0 +1,47 @@
+{
+  "version": "s22",
+  "decisions": [
+    {
+      "id": "host-owns-completion-gate",
+      "title": "The Host Owns the Completion Gate",
+      "description": "The working model may request to stop, but the harness evaluates the active goal before returning. Completion is a program decision at the turn boundary.",
+      "alternatives": "Asking the working model whether it is finished is simpler, but lets the same actor make and verify its own claim.",
+      "zh": {
+        "title": "完成闸门由宿主持有",
+        "description": "工作模型可以请求停止，但 harness 会在 return 前评估 active goal。是否完成是轮次边界上的程序决策。"
+      },
+      "ja": {
+        "title": "完了ゲートはホストが所有する",
+        "description": "作業モデルは停止を要求できるが、harness は return 前に active goal を評価する。完了は turn 境界でのプログラム判断である。"
+      }
+    },
+    {
+      "id": "host-assigned-evidence-origins",
+      "title": "Evidence Trust Comes from the Ingress Path",
+      "description": "Ordinary submit calls cannot attach trusted labels. Only an allowlisted host-event channel can deliver task or monitor evidence, so user and model prose cannot certify itself.",
+      "alternatives": "Trusting text content or caller-supplied labels makes the evidence boundary forgeable.",
+      "zh": {
+        "title": "证据信任来自入口路径",
+        "description": "普通 submit 不能附加可信标签；只有白名单宿主事件通道能送入 task 或 monitor 证据，因此用户与模型文本不能自证完成。"
+      },
+      "ja": {
+        "title": "証拠の信頼は入力経路から得る",
+        "description": "通常の submit は trusted label を付けられず、allowlist 済み host event channel だけが task や monitor evidence を届ける。ユーザーやモデルの文章は自己証明できない。"
+      }
+    },
+    {
+      "id": "bounded-continuation",
+      "title": "Every Automatic Continuation Needs a Budget",
+      "description": "An unmet goal queues another turn only while budget remains. Exhaustion marks the goal blocked and releases the gate instead of creating an infinite loop.",
+      "alternatives": "An unbounded goal is persistent, but an impossible condition can consume resources forever.",
+      "zh": {
+        "title": "每次自动续轮都必须有预算",
+        "description": "目标未满足时只在预算剩余时继续；耗尽后将目标标记为 blocked 并释放闸门，避免无限循环。"
+      },
+      "ja": {
+        "title": "自動継続には必ず予算を置く",
+        "description": "goal 未達時は予算が残る間だけ次の turn を追加する。使い切れば blocked にして gate を解放し、無限 loop を防ぐ。"
+      }
+    }
+  ]
+}

web/src/data/execution-flows.ts

@@ -541,6 +541,56 @@ export const EXECUTION_FLOWS: Record<string, FlowDefinition> = {
       { from: "recover", to: "context" },
     ],
   },
+  s21: {
+    nodes: [
+      { id: "start", label: "Workflow Tool Call", type: "start", x: COL_CENTER, y: 30 },
+      { id: "validate", label: "Validate Meta +\nPermission", type: "process", x: COL_CENTER, y: 120 },
+      { id: "launch", label: "async_launched", type: "process", x: COL_CENTER, y: 210 },
+      { id: "phase", label: "Select Phase", type: "process", x: COL_CENTER, y: 300 },
+      { id: "pipeline", label: "Pipeline /\nParallel Agents", type: "subprocess", x: COL_LEFT, y: 410 },
+      { id: "journal", label: "Append Journal", type: "process", x: COL_LEFT, y: 520 },
+      { id: "more", label: "More Stages?", type: "decision", x: COL_CENTER, y: 620 },
+      { id: "output", label: "Write Output", type: "process", x: COL_RIGHT, y: 520 },
+      { id: "notify", label: "task_notification", type: "end", x: COL_RIGHT, y: 620 },
+    ],
+    edges: [
+      { from: "start", to: "validate" },
+      { from: "validate", to: "launch" },
+      { from: "launch", to: "phase" },
+      { from: "phase", to: "pipeline" },
+      { from: "pipeline", to: "journal" },
+      { from: "journal", to: "more" },
+      { from: "more", to: "phase", label: "yes" },
+      { from: "more", to: "output", label: "no" },
+      { from: "output", to: "notify" },
+    ],
+  },
+  s22: {
+    nodes: [
+      { id: "start", label: "Model Wants to Stop", type: "start", x: COL_CENTER, y: 30 },
+      { id: "active", label: "Active Goal?", type: "decision", x: COL_CENTER, y: 120 },
+      { id: "evidence", label: "Read Trusted\nEvidence Window", type: "process", x: COL_CENTER, y: 220 },
+      { id: "evaluate", label: "Goal Satisfied?", type: "decision", x: COL_CENTER, y: 320 },
+      { id: "complete", label: "goal_completed", type: "end", x: COL_RIGHT, y: 430 },
+      { id: "budget", label: "Budget Left?", type: "decision", x: COL_LEFT, y: 430 },
+      { id: "continue", label: "Queue Goal\nContinuation", type: "process", x: COL_LEFT, y: 540 },
+      { id: "loop", label: "Next Agent Turn", type: "subprocess", x: COL_LEFT, y: 650 },
+      { id: "blocked", label: "goal_blocked", type: "end", x: COL_CENTER, y: 540 },
+      { id: "return", label: "Return", type: "end", x: COL_RIGHT, y: 120 },
+    ],
+    edges: [
+      { from: "start", to: "active" },
+      { from: "active", to: "return", label: "no" },
+      { from: "active", to: "evidence", label: "yes" },
+      { from: "evidence", to: "evaluate" },
+      { from: "evaluate", to: "complete", label: "yes" },
+      { from: "evaluate", to: "budget", label: "no" },
+      { from: "budget", to: "continue", label: "yes" },
+      { from: "budget", to: "blocked", label: "no" },
+      { from: "continue", to: "loop" },
+      { from: "loop", to: "start" },
+    ],
+  },
 };
 
 const CURRENT_FLOW_OVERRIDES: Record<string, FlowDefinition> = {

web/src/data/scenarios/s10.json

@@ -1,7 +1,7 @@
 {
   "version": "s10",
-  "title": "System Prompt",
-  "description": "The system prompt is assembled from runtime state and cached by a deterministic context key.",
+  "title": "Context Assembly",
+  "description": "Stable instructions and dynamic runtime state are assembled at the model boundary and cached by a deterministic context key.",
   "steps": [
     {
       "type": "system_event",

web/src/data/scenarios/s15.json

@@ -1,7 +1,7 @@
 {
   "version": "s15",
-  "title": "Agent Teams",
-  "description": "A lead agent can spawn teammate loops and coordinate through file-backed mailboxes.",
+  "title": "Agent Teams: Runtime Lab",
+  "description": "The first Agent Teams lab builds persistent teammate loops and file-backed mailboxes.",
   "steps": [
     {
       "type": "user_message",

web/src/data/scenarios/s16.json

@@ -1,7 +1,7 @@
 {
   "version": "s16",
-  "title": "Team Protocols",
-  "description": "Typed requests, request ids, and response matching turn team chat into a protocol.",
+  "title": "Agent Teams: Protocol Lab",
+  "description": "The second Agent Teams lab adds typed requests, request ids, and response matching to the s15 runtime.",
   "steps": [
     {
       "type": "user_message",

web/src/data/scenarios/s21.json

@@ -0,0 +1,43 @@
+{
+  "version": "s21",
+  "title": "Workflow Runtime",
+  "description": "A fixed review pipeline runs in parallel, records a journal, and resumes completed agent calls from cache.",
+  "steps": [
+    {
+      "type": "user_message",
+      "content": "Review this change across correctness, security, performance, and style.",
+      "annotation": "The goal is open-ended, but the review procedure has a fixed shape."
+    },
+    {
+      "type": "tool_call",
+      "toolName": "Workflow",
+      "content": "{\"name\":\"review-changes\",\"phases\":[\"Review\",\"Verify\"]}",
+      "annotation": "One tool call hands deterministic orchestration to the workflow runtime."
+    },
+    {
+      "type": "tool_result",
+      "content": "{\"status\":\"async_launched\",\"runId\":\"wf_review-changes_6779\"}",
+      "annotation": "The product runtime returns a claim ticket while work continues in the background."
+    },
+    {
+      "type": "system_event",
+      "content": "phase=Review -> pipeline([correctness, security, performance, style])",
+      "annotation": "Each item advances independently through the scripted stages."
+    },
+    {
+      "type": "system_event",
+      "content": "phase=Verify -> adversarial checks run in parallel",
+      "annotation": "Structured results cross a validation boundary before aggregation."
+    },
+    {
+      "type": "system_event",
+      "content": "journal append -> task_notification(status=completed)",
+      "annotation": "Every completed agent call is checkpointed before the final notification."
+    },
+    {
+      "type": "system_event",
+      "content": "resumeFromRunId -> unchanged calls status=cached",
+      "annotation": "Semantic keys reuse completed work without depending on concurrency order."
+    }
+  ]
+}

web/src/data/scenarios/s22.json

@@ -0,0 +1,37 @@
+{
+  "version": "s22",
+  "title": "Goal Loop",
+  "description": "A host-owned completion gate keeps the turn alive until trusted evidence satisfies the active goal.",
+  "steps": [
+    {
+      "type": "user_message",
+      "content": "/goal until tests passed and deploy green",
+      "annotation": "The command defines the condition but sits outside the evidence window."
+    },
+    {
+      "type": "system_event",
+      "content": "goal_started -> evidence window reset",
+      "annotation": "The harness, not the working model, owns the completion gate."
+    },
+    {
+      "type": "assistant_text",
+      "content": "Tests passed and deployment is green.",
+      "annotation": "Assistant prose is not trusted completion evidence."
+    },
+    {
+      "type": "system_event",
+      "content": "goal_evaluated satisfied=false -> continuation queued",
+      "annotation": "An unmet condition pushes the loop into another bounded turn."
+    },
+    {
+      "type": "tool_result",
+      "content": "<task_notification>tests passed; deploy green</task_notification>",
+      "annotation": "Only an allowlisted host event can attach a trusted evidence origin."
+    },
+    {
+      "type": "system_event",
+      "content": "goal_evaluated satisfied=true -> goal_completed",
+      "annotation": "Trusted evidence closes the goal and releases the stop gate."
+    }
+  ]
+}