④段整套同步（六节＋ P/S/F/J ＋模板）＋ open-design 移到技能根

一、④段整套同步（用户选候选 A） - 新增 `references/stage-proto-doc/assets/产品文档模板.html`（16 KB／247 行，从备份拷、转 LF）。 - ④段 SKILL.md 换成备份 15:28 那版，并按模板口径改齐： · 「产品文档框架」从「四块＋一附录」改成**六节** —— 页面结构／功能清单／使用场景／操作流程／跨页与界面状态／待确认与偏离；与模板同构，⛔ 不从空白页重写结构。 · 编号从 R/S/F 改成 **P 页面／S 场景／F 功能／J 二级功能**，各自唯一、按展示顺序通排。 · 保留旧版有价值的要求：每项功能补齐做什么／入口在哪／输入边界／异常与恢复；重点功能三处标红；横向流程图每步写清页面／功能／操作／完成判据。 · 完成标准「七节」→「六节」，自查清单同步换掉「需求」那批条目，补模板与配图两条。 · 路径 `designs/<项目>/` → `docs/pm/<项目>/ui/`（2 处）。 · 补回本机那条「每份产出文档开头都要有一张图」的引用（指向主 skill，段内不复述）。 - 主 skill 同步：四段结构表④行、④段沿革、画图小节（④段画页面结构图＋用户路径）。二、open-design 移到技能根（用户选候选 A） - `git mv product-planning/references/stage-delivery/assets/open-design → open-design/`（4178 文件）。 - `.gitignore` 白名单加 `!/open-design/` —— 否则移出去就不再纳入版本控制。 - 引用改 11 处：路径形态 `assets/open-design/...` → `<技能根>/open-design/...`；「随段携带的自带件」改称「同根技能」；「路径怎么找」那条改成自带件（段内相对路径）与同根技能（`<技能根>/<技能名>/`）两类分开写。 - open-design 自己的 ATTRIBUTION.md 两处同步（它现在与 oil-ui-pro 同层）。 - 副作用（有益）：命名闸门读数由 2063 降到 225 —— 此前它把 open-design 里 919 个第三方 md 当成规则一起扫了；现在只扫本包 23 个 md。已确认④段的 H1 与 4a/4b/4c 子步仍在检查范围内。三、验收 - check_naming.py --ws E:/ProgramData/AIProject/contentm_agent ＝通过 225 / 失败 0 / 提示 27。 - 活文件里 `assets/open-design` 残留 0；工作区 CRLF 文件 0。
2026-10-15 21:17:28 +08:00
1 parent fa0a9026c8
commit 5afeb6b78a
@@ -0,0 +1,99 @@
+# Live Artifact Schema Reference
+
+Live artifacts are stored as daemon-owned project files under `.live-artifacts/<artifactId>/`. Agents author the source files, then register them through daemon tooling. The daemon assigns IDs, project scope, timestamps, run scope, and refresh status.
+
+## Source files
+
+| File | Owner | Purpose |
+| --- | --- | --- |
+| `artifact.json` | agent-authored input, daemon-validated | Artifact metadata, preview settings, document metadata, and source descriptors. Must not contain daemon-owned fields. |
+| `template.html` | agent-authored | `html_template_v1` template used to render the preview. |
+| `data.json` | agent-authored then refresh-runner-updated | Canonical preview data. API `document.dataJson` is only a derived cache. |
+| `provenance.json` | agent-authored then refresh-runner-updated | Source summary and generation notes. |
+| `index.html` | daemon-derived | Generated preview output. Do not treat as source of truth. |
+
+## Create/update input
+
+`artifact.json` should match `LiveArtifactCreateInput` or `LiveArtifactUpdateInput` from `packages/contracts/src/api/live-artifacts.ts`.
+
+Allowed agent-owned top-level fields:
+
+- `title`
+- `slug`
+- `sessionId`
+- `pinned`
+- `status`
+- `preview`
+- `document`
+
+Daemon-owned fields are rejected in agent input:
+
+- `id`
+- `projectId`
+- `createdAt`
+- `updatedAt`
+- `createdByRunId`
+- `schemaVersion`
+- `refreshStatus`
+- `lastRefreshedAt`
+
+## HTML document contract
+
+MVP documents use `html_template_v1`:
+
+```json
+{
+  "format": "html_template_v1",
+  "templatePath": "template.html",
+  "generatedPreviewPath": "index.html",
+  "dataPath": "data.json",
+  "dataJson": {}
+}
+```
+
+`template.html + data.json` is rendered by the daemon into `index.html` and the preview route.
+
+### Binding rules
+
+- Use escaped interpolation: `{{data.path.to.value}}`.
+- Paths must start with `data` and use dot-separated keys; numeric array indexes are allowed as path segments.
+- Supported structural directive: `data-od-repeat="item in data.items"` for one-level array repeats.
+- Nested repeats, conditionals, filters, helper functions, partials, and expression evaluation are not supported.
+- Raw HTML insertion is forbidden: no triple braces, ampersand interpolation, `data-od-html`, `data-od-raw`, or equivalent.
+- Interpolation in text and ordinary attributes is HTML-escaped by default.
+- Do not interpolate inside tag names, attribute names, comments, `<script>`, `<style>`, `<iframe srcdoc>`, event-handler attributes, or unsupported URL-bearing attributes.
+
+## Bounded JSON limits
+
+All persisted JSON values must fit the shared bounded JSON envelope:
+
+| Limit | Value |
+| --- | ---: |
+| Maximum object/array depth | 8 |
+| Maximum keys per object | 100 |
+| Maximum array length | 500 |
+| Maximum string length | 16 KiB |
+| Maximum serialized JSON size | 256 KiB |
+
+Forbidden keys anywhere in persisted JSON include `raw`, `rawResponse`, `payload`, `body`, `headers`, `cookie`, `authorization`, `token`, `secret`, `credential`, and `password`.
+
+## Minimal static artifact input
+
+```json
+{
+  "title": "Release Status",
+  "preview": { "type": "html", "entry": "index.html" },
+  "document": {
+    "format": "html_template_v1",
+    "templatePath": "template.html",
+    "generatedPreviewPath": "index.html",
+    "dataPath": "data.json",
+    "dataJson": {
+      "summary": {
+        "title": "Release Status",
+        "status": "On track"
+      }
+    }
+  }
+}
+```
@@ -0,0 +1,119 @@
+# Connector Policy Reference
+
+Live artifacts may use connector or local data, but they must persist only compact, preview-oriented data and provenance. Never persist credentials or raw provider envelopes inside live artifact files.
+
+## Connector safety model
+
+Connector tools are classified by side effect and approval requirement:
+
+- `read` + `auto`: eligible for agent preview and potential refresh.
+- `write` + `confirm`: not refreshable; requires explicit user confirmation if exposed later.
+- `destructive` + `disabled`: never refreshable.
+- `unknown` + `confirm` or `disabled`: fail closed until classified.
+
+If a tool name, scope, or description suggests write/create/update/delete/admin/send/post/manage behavior, treat it as write-capable unless the daemon catalog explicitly proves otherwise. Destructive hints must be disabled for refresh.
+
+## Execution boundaries
+
+- Use daemon wrapper commands or `/api/tools/connectors/*`; do not call provider APIs directly from the artifact workflow when a daemon connector exists.
+- Tool endpoints require the injected `OD_TOOL_TOKEN`; do not invent or pass `projectId`.
+- Agent calls and refresh-runner calls must share the same daemon connector execution service.
+- Re-check connector status, allowlists, current scopes, tool safety, and refresh eligibility at execution time.
+- For connector-backed refresh, saved `connectorId`, `accountLabel`, tool name, input shape, and approval policy must still match current connector state.
+
+## Connector listing
+
+List connectors before using connector-backed data:
+
+```bash
+"$OD_NODE_BIN" "$OD_BIN" tools connectors list --format compact
+```
+
+The compact result includes each connector's `id`, display metadata, `status`, optional `accountLabel`, and callable tool summaries with `name`, `description`, `safety`, and `inputSchema`. Use this output to select a connector and tool; do not guess tool names.
+
+Only execute tools from connectors whose status is `connected`. Local/public connectors may already be connected by the daemon; OAuth-backed connectors must be connected by the user through the UI before agent execution.
+
+If the user already named a connector or app, treat that as the intended data source. For example, “create a Notion live artifact” means: list connectors, find `notion`, and if it is `connected`, use its read-only tools instead of asking where the Notion data comes from. Ask a follow-up only when the matching connector is missing/unconnected, when several connected matches are equally plausible, or when there is no searchable topic/page/database clue in the user’s request.
+
+For Notion, prefer this selection order:
+
+1. Use `notion.notion_search` with a concise query derived from the user’s requested artifact/topic.
+2. Use `notion.notion_fetch_database` only when the user provided a database id or a prior search result identifies a specific database.
+3. If the user simply says “Notion live artifact” with no topic, ask what Notion page/database/topic to visualize or whether to search broadly.
+
+## Connector execution
+
+Create a bounded JSON object input file that matches the selected tool's `inputSchema`, then execute through the wrapper:
+
+```bash
+"$OD_NODE_BIN" "$OD_BIN" tools connectors execute --connector "$CONNECTOR_ID" --tool "$TOOL_NAME" --input input.json
+```
+
+The wrapper reads `OD_NODE_BIN`, `OD_BIN`, `OD_DAEMON_URL`, and `OD_TOOL_TOKEN`, sends the request to `/api/tools/connectors/execute`, and prints compact JSON. Successful output includes `connectorId`, optional `accountLabel`, `toolName`, `safety`, `outputSummary`, redacted `output`, and daemon metadata. On failure, fix the input/schema/connection issue and retry; do not bypass connector validation with direct provider calls.
+
+Execution is fail-closed:
+
+- connector and tool IDs must be in the daemon catalog allowlist;
+- the connector must still be connected and not disabled;
+- current runtime safety must be `read` + `auto` for agent execution;
+- input must match the current tool schema;
+- run rate limits and total call limits apply;
+- output is size-bounded and redacted before the agent receives it.
+
+Use execution output as an intermediate source only. Normalize it into `data.json` and provenance, keeping only fields the preview needs.
+
+## Read-only refresh rules
+
+Connector-backed live artifact refresh is allowed only for tools that remain read-only and refresh-eligible at refresh time. A saved refresh source must include non-sensitive connector metadata and permission state, for example:
+
+```json
+{
+  "type": "connector_tool",
+  "toolName": "github.public_repo_summary",
+  "input": { "owner": "open-design", "repo": "open-design" },
+  "connector": {
+    "connectorId": "github_public",
+    "accountLabel": "public",
+    "toolName": "github.public_repo_summary"
+  },
+  "outputMapping": {
+    "dataPaths": [{ "from": "summary", "to": "repository" }],
+    "transform": "metric_summary"
+  },
+  "refreshPermission": "manual_refresh_granted_for_read_only"
+}
+```
+
+During refresh, the daemon revalidates `connectorId`, `accountLabel`, tool name, saved input schema, and allowlist membership. If anything drifts, the refresh fails without changing the previous valid preview.
+
+Never mark write, destructive, unknown, confirmation-required, disabled, unconnected, or schema-drifted connector tools as refreshable.
+
+## Persistence rules
+
+Persist only:
+
+- compact normalized values needed by the preview in `data.json`;
+- high-level provenance in `provenance.json`;
+- connector references and refresh metadata in `sourceJson`.
+
+Never persist:
+
+- OAuth tokens, API keys, cookies, headers, authorization values, or session material;
+- raw provider HTTP bodies, envelopes, payloads, or full responses;
+- credential-like values under alternate names;
+- connector credentials under `.live-artifacts/`.
+
+Credential storage is daemon-controlled and outside project artifact directories. Artifacts may contain connector IDs and non-sensitive account labels only.
+
+## Credential handling constraints
+
+- Do not ask the user for connector secrets inside the artifact workflow.
+- Do not ask the user to re-specify a data source that is already named and connected; inspect the connector catalog first.
+- Do not write OAuth material, API keys, cookies, sessions, HTTP request metadata, or provider auth state into `artifact.json`, `data.json`, `provenance.json`, tile JSON, snapshots, refresh history, or `.live-artifacts/`.
+- Do not include secret-like values in connector tool inputs or source metadata. If a connector requires credentials, the daemon-owned connector UI/storage must handle them outside project artifacts.
+- Safe persisted connector references are limited to catalog IDs, tool names, non-sensitive account labels, selected normalized output fields, and concise provenance notes.
+- If connector output contains unredacted sensitive or envelope-like fields, stop and return a validation/safety error instead of storing it.
+
+## Output protection
+
+Connector outputs must be bounded and redacted before returning to agents or entering artifact files. Use compact summaries and selected fields. If redaction cannot prove the result is safe, fail with a validation error instead of storing it.
@@ -0,0 +1,78 @@
+# Refresh Contract Reference
+
+Refresh updates live artifact data without redesigning the presentation. The refresh runner updates `data.json`, provenance, and audit history; it does not allow arbitrary template rewrites.
+
+## Refreshable source metadata
+
+Refreshable documents use `sourceJson`:
+
+```json
+{
+  "type": "connector_tool",
+  "toolName": "list_releases",
+  "input": {},
+  "connector": {
+    "connectorId": "github",
+    "accountLabel": "example/org",
+    "toolName": "list_releases"
+  },
+  "outputMapping": {
+    "dataPaths": [{ "from": "items", "to": "releases" }],
+    "transform": "compact_table"
+  },
+  "refreshPermission": "manual_refresh_granted_for_read_only"
+}
+```
+
+Supported source types:
+
+- `local_file`
+- `daemon_tool`
+- `connector_tool`
+
+Supported output transforms:
+
+- `identity`
+- `compact_table`
+- `metric_summary`
+
+## Source execution model
+
+- `refreshPermission` is retained for backward compatibility with older artifacts, but the refresh runner does not require a separate connector approval step.
+- If a safe source descriptor exists, manual refresh executes it through daemon-owned local or connector wrappers.
+- Write, destructive, unknown, disabled, unconnected, or schema-drifted connector tools should not be authored as refresh sources.
+
+## Connector-backed refresh
+
+Connector-backed refresh sources use the same connector execution service as agent-initiated connector calls. Do not call provider APIs directly from refresh logic or from skill-authored scripts.
+
+Before creating a connector-backed refresh source:
+
+1. List connectors with `"$OD_NODE_BIN" "$OD_BIN" tools connectors list --format compact`.
+2. If the user named a connector/source and it is connected, select that connector directly instead of asking where the source is. Then select a tool whose safety is `read` + `auto` and whose catalog metadata marks it refresh-eligible.
+3. Execute once with `"$OD_NODE_BIN" "$OD_BIN" tools connectors execute --connector <id> --tool <name> --input input.json` to produce compact normalized preview data.
+4. Store only non-sensitive connector references, the bounded input object, output mapping, and compatibility `refreshPermission` in `sourceJson`.
+
+On each refresh, the daemon must re-check connector status, account label, allowlist membership, input schema, and output protection. If any check fails or output protection rejects the result, refresh fails all-or-nothing and preserves the previous valid preview.
+
+Persisted connector refresh metadata may include `connectorId`, `toolName`, non-sensitive `accountLabel`, bounded `input`, `outputMapping`, and compatibility `refreshPermission`. It must not include credentials, auth/session material, raw provider envelopes, or unbounded provider responses.
+
+## Commit behavior
+
+Refresh is all-or-nothing:
+
+1. Acquire one active refresh lock per artifact.
+2. Execute each refreshable source with timeouts and current safety checks.
+3. Build candidate `data.json`, provenance, and preview.
+4. Validate all candidates with the same schemas used for create/update.
+5. Commit only if every refreshable source succeeds.
+6. Preserve the previous valid preview if any step fails.
+
+Refresh IDs must be monotonic so stale runs cannot overwrite newer committed data.
+
+## Audit storage
+
+- Append compact records to `refreshes.jsonl`.
+- Successful refresh snapshots live under `snapshots/<refreshId>/` and may include `data.json` and provenance.
+- Failed refreshes are summarized in `refreshes.jsonl` without leaking raw provider output or credentials.
+- On daemon startup, stale running refreshes should be marked failed or timed out while preserving the last valid preview.