data-talk

Use when the user asks for a chart, KPI, trend, report, or multi-widget dashboard. Triggers on phrases like 图表/趋势图/可视化/报表/仪表盘/大屏/多图组合 and chart/visualization/dashboard/KPI/multi-widget/ECharts/grid layout. Covers the inline `chart` fenced code block (default single-chart path), `chart:<artifactId>` for SQL-derived charts, and dashboard schema v1 plus incremental `ui_patch` edits for persistent multi-widget layouts.

Use when an LLM agent edits workbench tabs (query_editor, artifact_preview, er_inspector, er_designer, future report_designer) and encounters version conflicts or must guard against concurrent writes. Triggers on 版本冲突 / 409 / version_conflict / anchor_not_found / anchor_ambiguous / baseVersion / expectedVersion / optimistic locking / multi-edit batch / concurrent edit / 乐观锁 / 并发编辑 / 重试. Defines the optimistic-locking contract for workspace-wide, persisted tabs that any session (including parallel agents) may have mutated since your last read, plus mandatory recovery steps on conflict.

Use when invoking the four UI tools — `datatalk_ui_find` / `datatalk_ui_read` / `datatalk_ui_patch` / `datatalk_ui_exec` — to discover, inspect, edit, or execute actions on stage tabs (workspace / query_editor / er_inspector / er_designer). Triggers on `ui_find` / `ui_read` / `ui_patch` / `ui_exec` / `apply_text_edits` / `set_context` / `操作 UI` / `编辑 tab` / `查找 tab` / `读取 tab 状态` / `修改 SQL 编辑器`. Defines parameter shapes, required fields per action, top-level `object`/`target`/`mode`/`action`/`params` semantics, and the post-edit re-read verification rule.

Premium industrial dashboards from v3 JSON. Use for dashboards, KPI boards, monitoring screens, ops cockpits, big-screen displays, or any multi-metric data visualization. The backend Bezel compiler turns JSON into HTML deterministically — no AI-generated HTML.

企业级汇报文档生成（业务月报、问题复盘、季度总结、事件postmortem）。 与 bezel skill 的边界：bezel 做实时大屏 dashboard（深色科技风、单源 widget、轮询刷新）； ledger 做冻结快照报告（白底衬线印刷品风、跨源叙事、PDF 邮件转发与打印归档）。 二者**不可互相替代**：用户说"做大屏"用 bezel，用户说"做汇报/周报/月报/复盘"用 ledger。

Use when the user asks to query, mutate, or change schema against a SQL connection. SELECT/INSERT/UPDATE/CREATE/ALTER...ADD run via `datatalk_execute_sql`; DELETE requires `confirmationId` confirm; destructive DDL (DROP/TRUNCATE/ALTER...DROP/GRANT/REVOKE) returns `redirect_to_editor`. Triggers on 查询/统计/分析/读表/取数/插入/更新/删除/建表/改表/count/select/insert/update/delete/create/alter/drop/aggregate/group by/trend/top N/explore schema/describe table. Covers execute_sql contract, DELETE confirm, DDL redirect, read_schema discovery/describe, and table-not-found probe.

Use when the user mentions a specific database product, asks about supported data source types, picks a connection kind, or asks why a feature (ER / diagnostics / mutation) is unavailable on a given dialect. Triggers on MySQL / PostgreSQL / Oracle / SQL Server / SQLite / DuckDB / ClickHouse / TiDB / MariaDB / OceanBase / StarRocks / Apache Doris / Trino / Presto / Dameng / Apache Hive / GaussDB / 方言 / dialect / 数据库类型 / 数据源类型 / 连接类型. Covers per-dialect connection kind, default port, JDBC driver, schema visibility, SQL splitter family, risk-tier highlights, and ER / diagnostics support status.

Routes uploaded files to appropriate actions based on pre-analysis summary and user intent. Activated when user message contains a file_upload part.

Collect external data via Python/Node.js scripts and write into SQL connections. 数据采集脚本、爬虫脚本、API 数据拉取、抓取网页数据、script runner、Python script、Node.js script、data scraping、web scraping、data fetching.

Use when the user manages saved data source connections or switches the session data context (创建连接 / 测试连接 / 编辑连接 / 选择连接 / 切换数据库 / 切换 schema / use xxx / 数据上下文 / connection / data context / connectionId / confirmable mutation / two-phase confirm / terminate session / optimize table). Owns the 6 session-data-context tools, the 3 connection-lifecycle tools, the confirmable mutation two-phase protocol (`confirm=true` + `confirmationToken`), plus `datatalk_terminate_session` and `datatalk_optimize_table`.

Use when the user asks about ER diagrams, schema visualisation, or schema design / fork / apply. Triggers on ER 图 / ER 设计器 / 实体关系 / 看表关系 / 设计表结构 / fork to designer / generate DDL / ER inspector / ER designer / draw ER / annotate relation. Covers the two ER tab types — er_inspector (read-only ER Diagram Viewer with annotation overlay) and er_designer (independent schema draft producing DDL) — plus dialect support (MySQL/PostgreSQL/H2/MariaDB live; SQLite CREATE-only; DuckDB/ClickHouse/Doris/StarRocks/Oracle/SQL Server unsupported) and the rule that designer DDL is never auto-applied.

Use when a SQL run fails or a performance question arrives. Triggers on SQL 报错 / 表不存在 / 列不存在 / 多候选 / 性能分析 / EXPLAIN / 索引建议 / 锁等待 / 慢查询 / 连接池 / 表空间 / SQL execution failed / table not found / Unknown column / ambiguous candidates / Select a database first / slow query / index hint / lock blocking / pool full / disk usage. Owns the three SQL-failure diagnostic classes (syntax, object-not-found incl. cross-DB probe, ambiguous candidates) and the five diagnostic tools `datatalk_explain_query` / `datatalk_index_hints` / `datatalk_lock_info` / `datatalk_pool_status` / `datatalk_table_space`.

Use when the AI is about to write SQL touching tables it has not yet read — multi-table JOIN / cross-table aggregation / first-time operation on an unfamiliar table — OR when the user's keyword does not yet map to a known table name. Triggers on 分析销售/统计/探索/不知道表名/where is the table for X/which table stores/汇总数据/give me trend of. Defines the Pre-Action Exploration Protocol: ordering get_data_context → schema_search → read_schema → execute_sql, budget 3 misses → question tool.

Use when the user wants to browse table rows, inspect sample data, run a simple table preview or row count, write SQL, open a SQL editor, or execute SQL in the editor. Triggers on SQL 编辑器/查询编辑器/打开 SQL/写 SQL/跑 SQL/查表/看表数据/open query editor/open SQL/write SQL/run query/edit SQL/show rows from. Covers the full query_editor lifecycle (open → set_context → patch /content or apply_text_edits → run_sql → focus), the boundSessionId / source / useSessionContext context model, and the 5 Query Editor Rules.

Use when producing files, deliverables, or large tool outputs in a DataTalk session — covers artifact lifecycle (temporary vs archive candidate), `datatalk_archive_artifact` / `datatalk_supersede_artifact` / `datatalk_pin_artifact` tools, and the `saved file path` continuation protocol for oversized responses. Triggers on 工件 / 产物 / 归档 / 落盘 / saved file path / supersede / pin artifact / large output / archive artifact / 链接旧产物 / 替换工件.

Use when the user asks for a business metric, uses business terms ("销售额" / "GMV" / "客单价" / "复购率" / 自然语言度量名), asks to define/lookup semantic model entities/dimensions/measures/metrics, or asks to record a verified query. Covers 6 Actions (semantic_lookup / verified_query_find / verified_query_record / semantic_propose_change / literal_mapping_add / skill_create), L0-L3 Verified Query routing, when to trigger propose_change vs record verified query. 当用户询问业务指标、使用业务术语、或需要查询/定义语义模型时使用。

Create a new business domain Semantic Model skill (YAML) for DataTalk. Triggered when the user asks to define business metrics, create a semantic model for a new domain, or build a skill for a specific business area. 当用户要求为业务域创建语义模型、定义业务指标、或创建新技能时使用。forked_from: anthropics/skills@<upstream-sha-pending>

Tab 生命周期、复用、持久化与搜索策略。触发词：复用 tab / 新开 tab / 当前编辑器 / 上一条 SQL / 继续编辑 / library vs workset / tab reuse / continuation / search tabs / locate text / inWorkset / archive tab / focus tab。指导何时复用 query_editor、何时新开 tab、library 与 workset 的区别、跨重启持久化、以及 ui_find 三动词（list/search/read）组合策略。

A skill for unit testing the SkillResourceSyncer