@chanlerdev/scorel 0.0.5 → 0.0.7

package/docs/CHANGELOG.md

@@ -2,6 +2,65 @@
 
 ## Unreleased
 
+## 0.0.7 - 2026-06-24
+
+### Highlights
+
+- Introduce structured system_reminder content blocks across CLI, GUI, WebUI, and daemon, replacing ad-hoc XML strings.
+- Bundle GUI runtime dependencies for self-contained execution.
+
+### Changes
+
+- System reminders now use structured `system_reminder` content blocks with origin, visibility, and scope, enabling consistent handling across all interfaces.
+- GUI's CLI runtime is now bundled with all dependencies, eliminating the need for node_modules.
+
+### Fixes
+
+- Snip tool result no longer exposes internal span IDs or event counts to the model, keeping output concise.
+
+### Breaking Changes
+
+- Protocol version incremented from 4 to 5; session headers must now carry version 5.
+
+### Verification
+
+- All existing tests pass, with new tests covering system_reminder lowering, message-attached reminders, projector filtering, and bundled runtime integrity.
+
+- Protocol version incremented to 5 with structured `system_reminder` content blocks.
+- Snip tool results now return a concise model-visible confirmation while keeping internal span details out of provider context.
+
+## 0.0.6 - 2026-06-23
+
+### Highlights
+
+- Packed GUI now bundles its own CLI runtime, enabling fully self-contained local Host startup without Node.js or global CLI.
+- Agents can now use the `snip` tool to hide completed user turns from future context — reducing token waste and keeping conversations focused.
+
+### Changes
+
+- GUI now bundles its own CLI runtime for fully self-contained Host startup (no Node.js or global CLI required).
+- Added `snip` tool that agents can call to mark a completed user turn as hidden from future LLM context.
+- Protocol version incremented to 4 with new `context_control` event and `hide_user_turn` operation.
+- UI (GUI and WebUI) now hides model-only text blocks (like `snip` reminders) from the visible transcript.
+- Provider adapters now pass each tool's own parameter schema instead of hardcoding tool-specific parameters.
+
+### Fixes
+
+- Increased Host startup timeout from 10s to 30s to prevent timeouts in slower development environments.
+
+### Breaking Changes
+
+- Protocol version incremented from 3 to 4; requires protocol-version-aware clients.
+- Packaged GUI no longer accepts `SCOREL_CLI_ENTRYPOINT` or `SCOREL_NODE_PATH` environment variables.
+
+### Verification
+
+- Unit and release tests verify bundled CLI usage in packaged GUI, snip tool end-to-end behavior, hidden spans in context builds, protocol version bump, and UI projector rendering.
+
+### Internal
+
+- Added planned spec S0107 for system reminder unification (documentation only).
+
 ## 0.0.5 - 2026-06-19
 
 ### Highlights

package/docs/ROADMAP.md

@@ -600,6 +600,9 @@ M5 WebUI 的正式产品方向记录在 [`S0030`](spec/ship/S0030-webui-product-
 | M9.F1.24 | [`S0102`](spec/ship/S0102-device-only-config.md) | Config 彻底收敛为设备级唯一配置，移除 Project config 运行时语义 | Done |
 | M9.F1.25 | [`S0103`](spec/ship/S0103-daemon-lifecycle-and-settings-resilience.md) | Daemon 生命周期按入口区分，并修复 GUI Settings remote 切换黑屏风险 | Done |
 | M9.F1.26 | [`S0105`](spec/ship/S0105-cli-update-and-gui-release.md) | CLI 命令面统一补齐、NPM 手动/自动更新、GUI release 打包和增量更新框架 | Done |
+| M9.F1.27 | [`S0106`](spec/ship/S0106-snip-context-control.md) | `context_control` 持久事件和 `snip` tool，让 agent 隐藏已完成 user turn 的未来 LLM context 投影 | Done |
+| M9.F1.28 | [`S0107`](spec/ship/S0107-system-reminder-unification.md) | 统一 system reminder 的持久化、构造、LLM 投影和 UI visibility 语义 | Done |
+| M9.F1.29 | [`S0108`](spec/ship/S0108-gui-bundled-cli-runtime.md) | GUI release 内置同版本 CLI runtime，packaged GUI 用 bundle 内可执行文件启动本地 Host | Done |
 
 **Not in M9 Follow-up**:
 
@@ -779,6 +782,9 @@ HTTP adapter 必须映射已有 Host use cases，不复制领域逻辑。
 | [`S0103`](spec/ship/S0103-daemon-lifecycle-and-settings-resilience.md) | Daemon lifecycle and Settings resilience | Done |
 | [`S0104`](spec/ship/S0104-tool-result-artifacts.md) | Tool result artifacts for oversized Bash output | Done |
 | [`S0105`](spec/ship/S0105-cli-update-and-gui-release.md) | CLI update and GUI release | Done |
+| [`S0106`](spec/ship/S0106-snip-context-control.md) | Snip context control | Done |
+| [`S0107`](spec/ship/S0107-system-reminder-unification.md) | System reminder unification | Done |
+| [`S0108`](spec/ship/S0108-gui-bundled-cli-runtime.md) | GUI bundled CLI runtime | Done |
 
 ---
 

package/docs/SHIP.md

@@ -209,6 +209,7 @@ pnpm release patch --no-generate-notes
 - 执行 WebUI production build
 - 执行 GUI production build
 - 构建 public `scorel` npm package
+- 将同版本 public `scorel` CLI runtime vendored 进 GUI app bundle，供 packaged GUI 用 bundle 内绝对路径启动本机 Host
 - 执行 `npm pack` 安装烟雾测试
 - bump 所有 package version
 - 构建 GUI macOS dmg / zip release assets，并生成 `latest-mac.yml` 与 blockmap 增量更新 metadata

package/docs/spec/channels.md

@@ -94,7 +94,7 @@ running default  -> follow_up queue
 
 ## 6. Source Reminder
 
-每个 IM turn 会在用户消息前注入 hidden channel reminder：
+每个 IM turn 会在用户消息前注入 hidden `harness_item kind="channel_context"`。`buildContext()` 会把它转成 `system_reminder` block，provider adapter 最后 lower 成类似下面的模型输入文本：
 
 ```xml
 <system-reminder>

package/docs/spec/events.md

@@ -143,7 +143,22 @@ interface CompactEvent extends PersistentEventBase {
 
 构建 context 时：从 leaf 往 root 走，遇到 CompactEvent → 注入 summary，停止继续向上。旧事件仍在 JSONL 中（可查阅），但不进入 LLM context。
 
-### 4.5 `channel_inject` — 外部来源元数据
+### 4.5 `context_control` — 控制上下文投影
+
+```typescript
+interface ContextControlEvent extends PersistentEventBase {
+  type: "context_control";
+  operation: "hide_user_turn";
+  anchorUserEventId: EventId;
+  throughEventId: EventId;
+  actor: "agent" | "user" | "system";
+  reason?: string;
+}
+```
+
+`hide_user_turn` 隐藏一个已完成 user turn span：从 `anchorUserEventId` 开始，到同一路径下一条 `user_message` 前一条事件结束。它只影响未来 `buildContext()` 的 LLM 输入投影；原始 JSONL 事件仍保留并可用于 UI、审计和 resync。
+
+### 4.6 `channel_inject` — 外部来源元数据
 
 ```typescript
 interface ChannelInjectEvent extends PersistentEventBase {
@@ -325,7 +340,7 @@ interface EventTypeHandler<T extends PersistentEvent> {
 ```typescript
 type LlmAction =
   | { action: "include"; message: ScorelMessage }       // 正常包含为一条消息
-  | { action: "merge_prev"; content: string }           // 合入前一条消息（<system-reminder> 包裹）
+  | { action: "merge_prev"; reminder: SystemReminderContentBlock } // 合入前一条 tool_result
   | { action: "skip" }                                  // 不包含在 LLM context 中
   | { action: "barrier"; summary: ScorelMessage }       // 替换上方所有消息，注入 summary，停止遍历
 ```
@@ -335,11 +350,12 @@ type LlmAction =
 | Event 类型 | convertToLlm | convertToDisplay |
 |---|---|---|
 | `message`（user/assistant/tool_result） | `include` — 原样包含 | 正常气泡 |
-| `message`（meta.source = "steer"） | `merge_prev` — 合入前一条 tool_result 的 `<system-reminder>`；无 tool_result 则 `include` 作为独立 user msg | 内联小字提示 |
+| `harness_item`（steer / runtime guidance） | `merge_prev` — 合入前一条 tool_result 的 `system_reminder` block；无 tool_result 则作为独立 user msg | 内联小字提示 |
 | `message`（meta.source = "followUp"） | 同 steer | 内联 "追加任务" 标记 |
 | `rewind` | `skip` | "回退到此处" 标记 |
 | `branch` | `skip` | "切换分支" 标记 |
 | `compact` | `barrier` — 注入 summary，停止向上遍历 | "已压缩" 折叠块 |
+| `context_control` | `filter` — 从未来 LLM context 中排除指定 user turn span | "已 snip" 标记 |
 | `channel_inject` | `skip` | 来源 badge "from Telegram" |
 | `session_info` | `skip` | "模型切换为 X" 通知 |
 | `custom` | `skip` | Extension 自定义 |
@@ -381,23 +397,40 @@ function buildContext(tree: SessionTree, leafId: EventId): ScorelMessage[] {
 
 ---
 
-## 7. `<system-reminder>` 通用 Harness 注入机制
+## 7. `system_reminder` 通用 Harness 注入机制
 
 ### 7.1 用途
 
-`<system-reminder>` 是 Scorel harness 向 LLM 传递旁路信息的统一格式。所有非用户直接输入、但需要 LLM 看到的系统级内容都用此标签包裹。
+`system_reminder` 是 Scorel harness 向 LLM 传递旁路信息的结构化 content block。所有非用户直接输入、但需要 LLM 看到的系统级内容都先表达成结构化 block；`<system-reminder>` 只是 provider adapter 最后生成的传输 envelope。
 
 ### 7.2 使用场景
 
 | 场景 | 注入内容 | 注入位置 |
 |------|---------|---------|
 | Steer（用户中途插话） | 用户的引导文字 | merge 进前一条 tool_result |
-| Hook 上下文（UserPromptSubmit 等） | hook 产出 | user message / tool_result 末尾 |
+| User message sidecar | 当前 turn 时间、用户输入引用、`snip.userMessageId` | 同一条 user message 的 `content` |
+| Hook 上下文（UserPromptSubmit 等） | hook 产出 | user message / tool_result |
 | Memory 召回 | 记忆内容 | tool_result 末尾 |
 | 系统提醒（超时、配额等） | 通知文本 | tool_result 末尾 |
 | Channel 来源标注 | 来自哪个群/频道 | user message 内 |
 
-### 7.3 格式
+### 7.3 Canonical 格式
+
+```typescript
+interface SystemReminderContentBlock {
+  type: "system_reminder";
+  kind: SystemReminderKind;
+  origin: "system" | "user" | "tool" | "skill";
+  text: string;
+  visibility: "model" | "display" | "compact";
+  scope: "message" | "turn" | "next_model_call" | "session";
+  data?: Record<string, unknown>;
+}
+```
+
+### 7.4 Provider lowering
+
+Provider adapter 把 canonical block lower 成当前模型输入格式。默认 text envelope 是：
 
 ```xml
 <system-reminder>
@@ -405,12 +438,13 @@ function buildContext(tree: SessionTree, leafId: EventId): ScorelMessage[] {
 </system-reminder>
 ```
 
-### 7.4 注入规则
+### 7.5 注入规则
 
-- **工具循环中**：merge 进最近一条 tool_result 的 content 末尾
-- **无 tool_result 时（idle / turn 结束后）**：作为独立 user message（或附加到 user message 内）
+- **跟随 user message**：作为同一条 `user_message.content` 的 sidecar block，创建时固定，保持 prompt-cache 稳定。
+- **工具循环中**：merge 进最近一条 tool_result 的 content 末尾；provider 不支持时 fallback 为紧邻 tool result batch 的 user message。
+- **无 tool_result 时（idle / turn 结束后）**：作为独立 user message。
 
-### 7.5 LLM System Prompt 声明
+### 7.6 LLM System Prompt 声明
 
 LLM 在 system prompt 中被告知：
 

package/docs/spec/runtime.md

@@ -174,15 +174,15 @@ Steer message persist 为**独立 PersistentEvent**（role = "user"，`meta.sour
 
 | 前面有 tool_result | 行为 |
 |---|---|
-| ✅ 有 | `merge_prev` — 合入前一条 tool_result content 末尾，用 `<system-reminder>` 包裹 |
+| ✅ 有 | `merge_prev` — 合入前一条 tool_result content 末尾，内容为结构化 `system_reminder` block |
 | ❌ 没有（idle 状态） | `include` — 作为独立 user message |
 
-LLM 最终看到的（工具循环中）：
+Provider lowering 后 LLM 最终看到的（工具循环中）：
 ```
 tool_result: "文件内容...\n\n<system-reminder>\n别改了，直接跑测试\n</system-reminder>"
 ```
 
-LLM 最终看到的（idle 时）：
+Provider lowering 后 LLM 最终看到的（idle 时）：
 ```
 user: "别改了，直接跑测试"
 ```

package/docs/spec/session.md

@@ -198,8 +198,8 @@ function buildContext(tree: SessionTree, leafId: EventId): ScorelMessage[] {
         messages.unshift(result.message);
         break;
       case "merge_prev":
-        // 合入 messages 中最后一条 tool_result 的 content 末尾（<system-reminder> 包裹）
-        mergeIntoPrevToolResult(messages, result.content);
+        // 合入 messages 中最后一条 tool_result 的 content 末尾（system_reminder block）
+        mergeIntoPrevToolResult(messages, result.reminder);
         break;
       case "skip":
         break;
@@ -216,7 +216,7 @@ function buildContext(tree: SessionTree, leafId: EventId): ScorelMessage[] {
 
 各事件类型的 LlmAction：
 - `message`（普通）→ `include`
-- `message`（meta.source = "steer"/"followUp"）→ `merge_prev`（前面有 tool_result 时）或 `include`（没有时）
+- `harness_item` / runtime guidance → `merge_prev`（前面有 tool_result 时）或独立 user message，内容为结构化 `system_reminder` block
 - `compact` → `barrier`（注入 summary，停止遍历）
 - `rewind` / `branch` / `channel_inject` / `session_info` / `custom` → `skip`
 - `custom_message` → `include`
@@ -345,6 +345,17 @@ if (estimateTokens(compactCandidates) > threshold) {
 - 在 compact 点之前分叉的其他分支不受影响
 - 旧事件仍在 JSONL 中，可供历史浏览
 
+### 6.3 Snip Context Control
+
+`snip` 不删除 JSONL。它 append `context_control operation="hide_user_turn"`，让后续 `buildContext()` 在 active path 上过滤一个已完成 user turn span。
+
+安全边界：
+
+- 目标必须是当前 active path 上的 `user_message`。
+- 目标之后必须存在下一条 `user_message`，因此不能 snip 当前正在执行的 turn。
+- 隐藏范围是目标 `user_message` 到下一条 `user_message` 前一条事件，保证 tool_call / tool_result 成对隐藏，不留下孤儿工具结果。
+- 原始事件仍保留在 JSONL 中，UI 和审计可以继续查看。
+
 ---
 
 ## 7. 两层消息在本模块的落点
@@ -355,7 +366,8 @@ if (estimateTokens(compactCandidates) > threshold) {
 - buildContext 通用遍历时调用每个 event 的 handler，不 hardcode 任何类型
 - `rewind` / `branch` / `channel_inject` / `session_info` / `custom` → `skip`（不进入 LLM）
 - `compact` → `barrier`（注入 summary，停止向上）
-- `message`（meta.source = "steer"）→ `merge_prev`（合入前一条 tool_result 的 `<system-reminder>`）
+- `context_control` → `filter`（从未来 LLM context 排除指定 user turn span）
+- `harness_item` / runtime guidance → `merge_prev` 或独立 user message，内容为结构化 `system_reminder` block
 
 换言之，应用层能玩的花样很多，LLM 始终只看到 handler 声明要暴露的内容。
 

package/docs/spec/ship/S0106-snip-context-control.md

@@ -0,0 +1,113 @@
+# S0106: Snip Context Control
+
+## Goal
+
+Let long-running sessions remove a completed user turn from future model context without deleting the underlying session record.
+
+The business value is context hygiene: when an earlier user turn led the agent down a noisy or obsolete path, the user or agent can mark that whole turn as no longer relevant so future model calls stop paying attention to it. The original JSONL remains the evidence chain for UI, audit, resync, and debugging.
+
+## Scope
+
+### Stable User Turn References
+
+Every persisted `user_message` already has a durable `EventId`. S0106 treats that event id as the source of truth for snip targeting. The model-facing id is a stable short alias derived from the real `EventId`; the alias resolves back to the real event id but is not a storage key.
+
+When the Host creates a `user_message`, it appends a model-only text block to that same persisted message:
+
+```text
+<system-reminder>
+snip.userMessageId: u_...
+</system-reminder>
+```
+
+The block is persisted with the message so future `buildContext()` replays do not rewrite earlier user messages and do not invalidate prompt-cache prefixes. UI projectors hide model-only text blocks from the visible transcript.
+
+### Context Control Event
+
+Add a persistent control event:
+
+```ts
+type ContextControlEvent = {
+  type: "context_control";
+  operation: "hide_user_turn";
+  anchorUserEventId: EventId;
+  throughEventId: EventId;
+  actor: "agent" | "user" | "system";
+  reason?: string;
+};
+```
+
+The event is append-only and does not participate in the conversation tree. `buildContext()` folds these events as control state and filters the active path before producing LLM messages.
+
+`hide_user_turn` semantics:
+
+- `anchorUserEventId` must identify a `user_message` on the active session path.
+- The hidden span starts at that user event.
+- The hidden span ends at the event before the next `user_message` on the same path, or the active leaf when there is no later user message.
+- The resolved end is stored as `throughEventId`.
+- The span is hidden from future `buildContext()` calls after the control event is appended.
+- Original events stay in JSONL and continue to be visible through session replay and UI projection.
+
+### `snip` Tool
+
+Expose a lazily available `Snip` runtime tool that lets the agent request hiding a completed user turn from future context. The tool accepts:
+
+```json
+{ "userMessageId": "u_...", "reason": "optional short reason" }
+```
+
+The Host validates the request, resolves the model-visible short alias to a target span, appends a `context_control` event, and returns a brief model-visible confirmation. Internal span details such as `anchorUserEventId`, `throughEventId`, and hidden event count may remain in structured tool result details for diagnostics, but provider context must only receive the concise confirmation. The tool result is still part of the current turn; the hidden span disappears on the next context build.
+
+The tool is session-context control, not a generic coding tool. It must be registered by the Host with access to the current lane, not by `createCodingTools()`.
+
+Tool parameter schemas are owned by `AgentTool` definitions. Provider adapters must pass through the tool's declared schema instead of hard-coding behavior for specific tool names such as `snip`.
+
+## Not In Scope
+
+- Physically deleting, truncating, or rewriting historical JSONL.
+- Arbitrary single-message deletion.
+- Snipping across branches.
+- User-facing GUI controls for browsing snipped spans.
+- A CLI slash command.
+- Automatic snip heuristics.
+- Reusing `compact` for snip. Compact summarizes old context; snip excludes a specific user turn.
+
+## Acceptance Criteria
+
+- `PersistentEvent` includes `context_control`.
+- Session replay folds `hide_user_turn` events into control state.
+- `buildContext()` excludes the hidden user-turn span from future LLM context.
+- The excluded span can include assistant tool calls and tool results without leaving orphan tool results in context.
+- JSONL remains append-only; original user/assistant/tool events remain loadable after snip.
+- Repeated snip of the same target is idempotent enough to keep context stable.
+- Invalid targets fail as tool errors and do not append control events.
+- `snip` is registered by the Host and can append `context_control` from a real runtime tool call path.
+- The model can discover snippable user turn ids from its normal context projection; tests must not rely on out-of-band `send_message` response data.
+- User turn ids are persisted at user-message creation time, so replaying the same turn in later provider calls does not change that message and preserves prompt-cache stability.
+- Model-only short id blocks are hidden from WebUI and GUI transcript projection.
+- The `snip` `AgentTool` schema exposes `userMessageId` and optional `reason`, and provider adapters preserve tool-owned schemas without name-specific branches.
+- Full validation passes with `pnpm typecheck && pnpm test`.
+
+## Testing Requirements
+
+- Core session tests for context-control parsing, replay, and context filtering.
+- Protocol tests for exhaustive event handling.
+- Daemon embedded test proving `snip` appends `context_control` and the next provider call no longer receives the hidden turn.
+
+## Impacted Files
+
+- `packages/protocol/src/events.ts`
+- `packages/protocol/src/index.test.ts`
+- `packages/core/src/session/index.ts`
+- `packages/core/src/session/session.test.ts`
+- `packages/core/src/tools/index.ts`
+- `packages/daemon/src/index.ts`
+- `packages/daemon/src/embedded/embedded.test.ts`
+- `docs/ROADMAP.md`
+- `README.md`
+
+## Risks And Boundaries
+
+- Hiding the wrong turn is worse than compacting too much. Target resolution must use real `EventId`s and reject non-user targets.
+- Tool-call/tool-result pairing can be broken by arbitrary deletion. This spec only hides full user-turn spans.
+- Snip is context projection, not evidence deletion. UI and session replay must still show the original events unless a later explicit product decision adds a separate display filter.

package/docs/spec/ship/S0107-system-reminder-unification.md

@@ -0,0 +1,173 @@
+# S0107: System Reminder Unification
+
+## Goal
+
+Unify Scorel's runtime reminders as structured, model-facing context fragments.
+
+The business value is prompt and transcript hygiene on a high-frequency path. Scorel will routinely attach reminders to user messages, inject reminders while a turn is running, and route reminders through different provider message formats. This must be one stable product contract, not ad-hoc `<system-reminder>` strings scattered across daemon, session replay, tool-result merge paths, UI projectors, or provider adapters.
+
+## Scope
+
+### Structured Reminder Block
+
+Add a protocol-level content block:
+
+```typescript
+type SystemReminderKind =
+  | "attachment"
+  | "time"
+  | "message_ref"
+  | "skill_listing"
+  | "skill_delta"
+  | "memory"
+  | "channel_context"
+  | "steer"
+  | "todo_nudge"
+  | "runtime_notice"
+  | "compact_summary";
+
+type SystemReminderOrigin = "system" | "user" | "tool" | "skill";
+
+type SystemReminderScope =
+  | "message"          // travels with one persisted message whenever that message is in context
+  | "turn"             // relevant to the user turn that created it
+  | "next_model_call"  // runtime nudge, consumed by the next provider call
+  | "session";         // durable session context such as initial memory
+
+type SystemReminderVisibility = "model" | "display" | "compact";
+
+interface SystemReminderContentBlock {
+  type: "system_reminder";
+  kind: SystemReminderKind;
+  origin: SystemReminderOrigin;
+  text: string;
+  visibility: SystemReminderVisibility;
+  scope: SystemReminderScope;
+  data?: Record<string, unknown>;
+}
+```
+
+`<system-reminder>` remains the model-facing transport envelope, but it is no longer stored or hand-written by feature code. Callers create structured reminder blocks; core/provider projection renders the envelope.
+
+### Two Placement Modes
+
+System reminders can appear in two product situations:
+
+1. **Message-attached reminders**: created together with a `user_message` and persisted in that message's `content`.
+   - Examples: current time for the submitted turn, `snip.userMessageId`, references to prior user messages, channel context that explains the submitted text.
+   - These are stable sidecars. They are created when the message is persisted, so replaying the same message later does not mutate historical content or break prompt-cache prefix stability.
+
+2. **Runtime injected reminders**: appended while a turn is running or between provider calls.
+   - Examples: steer, skill delta, a nudge that the model has not used `TodoWrite`, runtime notices.
+   - These remain standalone `harness_item` events because they are independent session facts. `buildContext()` lowers them into structured reminder blocks and then places them according to provider-safe rules.
+
+### Canonical Context And Provider Lowering
+
+Scorel keeps a provider-neutral context:
+
+- `ScorelMessage.content` may contain `system_reminder` blocks.
+- UI/display projectors use block type and `visibility`; they must not parse `<system-reminder>` text.
+- Provider adapters receive canonical `ScorelMessage[]` and lower `system_reminder` blocks to the provider's legal representation.
+
+Default lowering keeps the current prompt contract:
+
+```xml
+<system-reminder>
+content
+</system-reminder>
+```
+
+Provider placement rules:
+
+- User-message sidecars are rendered inside the same user message after visible user text.
+- Runtime reminders prefer merge-after-tool-result when a valid previous tool result exists.
+- If a provider cannot legally merge after a tool result, fallback to a standalone user message immediately after the tool result batch.
+- Provider-level system/developer prompt is not used for runtime reminders.
+
+### Core Helper Surface
+
+Core owns reminder construction and rendering:
+
+- `createSystemReminderBlock(input)`
+- `renderSystemReminder(block | text)`
+- `renderSystemReminderText(text)`
+- `appendSystemReminderToToolResult(message, block)`
+- `systemReminderMessage(block, meta?)`
+
+Feature code must pass semantic fields: `kind`, `origin`, `scope`, `visibility`, `text`, optional `data`. Feature code must not write `<system-reminder>` tags.
+
+### Existing Source Migration
+
+Migrate these sources to structured reminder blocks:
+
+- `snip.userMessageId` model-only block attached to every persisted user message.
+- `harness_item` conversion for memory, channel context, skill listing, skill delta, steer, runtime notice, and future todo nudges.
+- compact summary injection.
+- GUI and WebUI transcript projection for model-only blocks.
+
+`harness_item` remains the persistent event for standalone runtime/session injections. S0107 does not need a new event type unless the implementation proves `harness_item` cannot carry the contract.
+
+## Not In Scope
+
+- Changing `snip` behavior from S0106.
+- Adding UI controls for browsing hidden reminders.
+- Backfilling or migrating old session JSONL files.
+- Renaming `<system-reminder>` in the model-facing prompt.
+- Moving runtime reminders into provider-level system/developer prompts.
+- Replacing all event-type conversion with a full handler registry.
+
+## Acceptance Criteria
+
+- Protocol supports `system_reminder` content blocks.
+- No daemon or feature code hand-writes `<system-reminder>` strings.
+- `buildContext()` uses shared core helpers for `harness_item` and compact summary conversion.
+- `snip.userMessageId` is a message-attached `system_reminder` block with stable prompt-cache behavior across later turns.
+- Provider adapters lower `system_reminder` blocks through the shared renderer, including reminders inside user messages and merged tool results.
+- WebUI and GUI hide model-only reminder blocks without parsing reminder text.
+- Existing harness visibility behavior stays intact:
+  - hidden harness items do not render as visible transcript turns;
+  - display/compact harness items can still render as lightweight transcript evidence.
+- `pnpm typecheck && pnpm test` passes.
+
+## Testing Requirements
+
+- Protocol tests for `system_reminder` content block round trip / exhaustiveness.
+- Core session tests for:
+  - message-attached reminder blocks rendering to `<system-reminder>` in provider context;
+  - `harness_item` conversion producing structured reminder blocks;
+  - merge-after-tool-result behavior preserving tool result content;
+  - compact summary using the shared reminder renderer.
+- Daemon embedded test proving snip's message-attached reminder remains stable across later turns.
+- Provider adapter test proving `system_reminder` blocks are lowered to `<system-reminder>` text.
+- WebUI and GUI projector tests proving model-only reminder blocks are hidden while display harness items remain visible.
+- Static regression check that common runtime paths no longer hand-write `<system-reminder>` literals outside tests/docs and the shared renderer.
+
+## Impacted Files
+
+- `packages/protocol/src/messages.ts`
+- `packages/protocol/src/index.test.ts`
+- `packages/core/src/session/index.ts`
+- `packages/core/src/session/session.test.ts`
+- `packages/core/src/provider/pi-ai.ts`
+- `packages/core/src/provider/pi-ai.test.ts`
+- `packages/daemon/src/index.ts`
+- `packages/daemon/src/embedded/embedded.test.ts`
+- `apps/webui/lib/events/projector.ts`
+- `apps/webui/lib/events/projector.test.ts`
+- `apps/gui/src/renderer/chatbox/projector.ts`
+- `apps/gui/src/renderer/chatbox/projector.test.ts`
+- `docs/spec/events.md`
+- `docs/spec/session.md`
+- `README.md`
+
+## Risks And Boundaries
+
+- Reminder placement affects prompt-cache behavior. Do not move message-attached reminders into dynamic `buildContext()` injection.
+- Tool-result merge behavior must preserve valid assistant tool-call / tool-result replay.
+- Runtime reminders can be frequent. The data model must keep origin, kind, scope, and visibility explicit so future skill, time, todo, IM, and provider-specific rules do not become string parsing.
+- UI must use explicit metadata, not text parsing.
+- A full event handler registry remains a later refactor; S0107 should ship the stable reminder contract first.
+
+## Status
+
+Done.