os-dpt 0.0.1 → 0.2.0

package/server/dist/index.mjs

@@ -1531,17 +1531,20 @@ async function streamAssistantMessage(p) {
       }
     },
     async (span) => {
-      const stream = client.messages.stream({
-        model,
-        max_tokens: maxTokens,
-        system: cachedSystem(p.system),
-        messages: withConversationCacheBreakpoint(p.messages),
-        tools: p.tools,
-        // One tool_use per assistant turn so the loop can cleanly handle the
-        // ask_user_question pause without leaving sibling tool_use blocks
-        // without matching tool_results.
-        tool_choice: { type: "auto", disable_parallel_tool_use: true }
-      });
+      const stream = client.messages.stream(
+        {
+          model,
+          max_tokens: maxTokens,
+          system: cachedSystem(p.system),
+          messages: withConversationCacheBreakpoint(p.messages),
+          tools: p.tools,
+          // One tool_use per assistant turn so the loop can cleanly handle the
+          // ask_user_question pause without leaving sibling tool_use blocks
+          // without matching tool_results.
+          tool_choice: { type: "auto", disable_parallel_tool_use: true }
+        },
+        { signal: p.signal }
+      );
       if (p.onTextDelta) {
         stream.on("text", (delta) => p.onTextDelta?.(delta));
       }
@@ -2018,6 +2021,7 @@ function costFromUsage(model, usage) {
 
 // server/agent/prompt.ts
 function buildSystemPrompt(session) {
+  if (session.meta.mode === "quick-edit") return buildQuickEditPrompt(session);
   const { worksheetSlug, connectionId } = session.meta;
   const bound = !!worksheetSlug;
   const lines = [
@@ -2059,7 +2063,7 @@ function buildSystemPrompt(session) {
     "- run_sql results are capped to 50 rows by default; add LIMIT or aggregation if you need a broader view.",
     "- run_sql executes whatever SQL you pass, including DDL and DML, against the role the user connected with. Treat exploration as read-only \u2014 use SELECT. Before any INSERT/UPDATE/DELETE/TRUNCATE/CREATE/DROP/ALTER, call ask_user_question to confirm.",
     "- ask_user_question pauses the loop entirely \u2014 only one question per call, and use it sparingly.",
-    "- render_chart draws a chart inline in the chat. After run_sql, when a picture beats a table (trends \u2192 line/area, comparisons \u2192 bar, proportions \u2192 pie), call it with pre-aggregated rows. Pass the data inline, shaped to small row objects (e.g. {month, revenue}); don't dump raw wide rows."
+    "- render_chart draws a chart inline in the chat. After run_sql, when a picture beats a table (trends \u2192 line/area, comparisons \u2192 bar, part-to-whole \u2192 stacked-bar, proportions \u2192 pie, stage drop-off \u2192 funnel), call it with pre-aggregated rows. Pass the data inline, shaped to small row objects (e.g. {month, revenue}); don't dump raw wide rows."
   );
   const ctx = [];
   if (worksheetSlug) ctx.push(`- Active worksheet: ${worksheetSlug}`);
@@ -2074,6 +2078,33 @@ function buildSystemPrompt(session) {
   }
   return lines.join("\n");
 }
+function buildQuickEditPrompt(session) {
+  const { worksheetSlug, connectionId } = session.meta;
+  const lines = [
+    "You are os-dpt's inline SQL-editing agent. The user prompts you from a small floating box inside their SQL editor. Your ONLY deliverable is SQL staged into their worksheet via write_sql \u2014 the user never reads chat output in this mode, they watch the editor update.",
+    "",
+    "Output rules (hard requirements):",
+    "- Emit NO prose. No preamble, no explanation, no closing summary \u2014 no text blocks at all.",
+    "- Anything worth flagging (an assumption you made, a caveat, a data-quality quirk) goes in a SQL comment inside the query itself.",
+    "- Always finish by calling write_sql with the COMPLETE worksheet contents (drafts are overwrites, not patches), then end the turn.",
+    "",
+    "Method:",
+    "- Each user message ends with the worksheet's current contents. Treat the request as an edit to that SQL unless it clearly asks for something new; preserve parts of the worksheet the request doesn't touch.",
+    "- Check get_context first when the request depends on schema, business terms, or conventions you have not verified this session; call get_schema when tables or columns are uncertain. Do not guess names.",
+    "- Verify your SQL with run_sql before staging it. If it errors, fix and re-run until it works. Treat verification as read-only \u2014 SELECT only; never run DDL or DML in this mode.",
+    "- If no connection is bound or verification is impossible, stage your best SQL with a leading '-- not verified' comment instead of stopping.",
+    "- You cannot ask the user questions in this mode. Make the most reasonable assumption and record it as a SQL comment.",
+    "- Persist durable knowledge via update_context the moment you learn it: table/column meanings \u2192 schemas.md; corrections or errors that revealed a fact \u2192 feedback.md; standing rules and business terms \u2192 conventions.md.",
+    "- run_sql results are capped to 50 rows by default; add LIMIT or aggregation if you need a broader view."
+  ];
+  const ctx = [];
+  if (worksheetSlug) ctx.push(`- Active worksheet: ${worksheetSlug}`);
+  if (connectionId) ctx.push(`- Active connection: ${connectionId}`);
+  if (ctx.length > 0) {
+    lines.push("", "Current bindings:", ...ctx);
+  }
+  return lines.join("\n");
+}
 
 // server/agent/session.ts
 import { promises as fs14 } from "node:fs";
@@ -2085,6 +2116,7 @@ function freshMeta(input) {
   const now = nowIso();
   return {
     id: crypto4.randomUUID(),
+    mode: input.mode ?? "chat",
     createdAt: now,
     updatedAt: now,
     title: input.title ?? null,
@@ -2102,6 +2134,9 @@ function ensureUsageFields(session) {
   if (!Array.isArray(session.meta.usage)) session.meta.usage = [];
   if (!session.meta.totals) session.meta.totals = emptyTotals();
   if (typeof session.meta.standalone !== "boolean") session.meta.standalone = false;
+  if (session.meta.mode !== "chat" && session.meta.mode !== "quick-edit") {
+    session.meta.mode = "chat";
+  }
   if (typeof session.meta.titleGenerated !== "boolean") {
     session.meta.titleGenerated = session.meta.title != null;
   }
@@ -2141,6 +2176,19 @@ async function deleteChat(id) {
   assertSafeChatId(id);
   await fs14.rm(paths.chat(id), { force: true });
 }
+var QUICK_EDIT_MAX_TURNS = 8;
+var QUICK_EDIT_KEEP_TURNS = 4;
+function trimQuickEditHistory(session) {
+  if (session.meta.mode !== "quick-edit") return;
+  const turnStarts = [];
+  session.messages.forEach((m, i) => {
+    if (m.role === "user" && typeof m.content === "string") turnStarts.push(i);
+  });
+  if (turnStarts.length <= QUICK_EDIT_MAX_TURNS) return;
+  session.messages = session.messages.slice(
+    turnStarts[turnStarts.length - QUICK_EDIT_KEEP_TURNS]
+  );
+}
 async function appendMessage(session, message) {
   session.messages.push(message);
   session.meta.updatedAt = nowIso();
@@ -2242,6 +2290,32 @@ import { promises as fs16 } from "node:fs";
 function isContextFile(v) {
   return CONTEXT_FILES.includes(v);
 }
+var DETAIL_CONTEXT_LINES = 8;
+function toLines(text) {
+  return text === "" ? [] : text.replace(/\n$/, "").split("\n");
+}
+function detailSnapshots(before, after) {
+  const a = toLines(before);
+  const b = toLines(after);
+  let prefix = 0;
+  while (prefix < a.length && prefix < b.length && a[prefix] === b[prefix]) prefix++;
+  let suffix = 0;
+  while (suffix < a.length - prefix && suffix < b.length - prefix && a[a.length - 1 - suffix] === b[b.length - 1 - suffix]) {
+    suffix++;
+  }
+  const leading = Math.max(0, prefix - DETAIL_CONTEXT_LINES);
+  const trailing = Math.max(0, suffix - DETAIL_CONTEXT_LINES);
+  if (leading === 0 && trailing === 0) return { before, after };
+  let fences = 0;
+  for (let i = 0; i < leading; i++) {
+    if (/^\s{0,3}(```|~~~)/.test(a[i])) fences++;
+  }
+  return {
+    before: a.slice(leading, a.length - trailing).join("\n"),
+    after: b.slice(leading, b.length - trailing).join("\n"),
+    trimmed: { leading, trailing, ...fences % 2 === 1 ? { inFence: true } : {} }
+  };
+}
 var updateContextTool = {
   name: "update_context",
   description: "Write to one of the agent's context files (schemas.md, conventions.md, feedback.md). Call this the moment you learn something durable \u2014 without waiting to be asked: a clarified schema fact or value set, a project convention, a user correction, a data-quality quirk, or an error pattern from running SQL. Saving is cheap and git-tracked; prefer over-saving to losing a finding. Prefer 'append' for new findings; use 'replace' only when restructuring an entire file. Writes are scoped to the data source bound to this chat; with no connection bound they go to the workspace-level (unassigned) docs.",
@@ -2290,16 +2364,16 @@ var updateContextTool = {
       };
     }
     const target = paths.contextFile(input.file, connectionId);
+    let current = "";
+    try {
+      current = await fs16.readFile(target, "utf8");
+    } catch (err) {
+      if (err.code !== "ENOENT") throw err;
+    }
     let next;
     if (input.mode === "replace") {
       next = input.content.endsWith("\n") ? input.content : input.content + "\n";
     } else {
-      let current = "";
-      try {
-        current = await fs16.readFile(target, "utf8");
-      } catch (err) {
-        if (err.code !== "ENOENT") throw err;
-      }
       const stamp = (/* @__PURE__ */ new Date()).toISOString().slice(0, 10);
       const block = `## ${stamp}
 
@@ -2311,7 +2385,8 @@ ${input.content.trim()}
     return {
       toolResult: `Wrote ${next.length} bytes to context/${input.file}.md (${input.mode})`,
       isError: false,
-      uiSummary: `Updated context/${input.file}.md (${input.mode})`
+      uiSummary: `Updated context/${input.file}.md (${input.mode})`,
+      detail: { file: `${input.file}.md`, mode: input.mode, ...detailSnapshots(current, next) }
     };
   }
 };
@@ -2550,9 +2625,13 @@ var runSqlTool = {
   description: "Execute SQL against the active database connection and return the result. Use this to validate queries you've written and to inspect data. If the query errors, capture what you learned by calling update_context against feedback.md so future runs avoid the same mistake. Row results are capped (default 50, max 500) \u2014 add LIMIT or aggregation if you need a wider view. IMPORTANT: this tool executes the SQL verbatim against whatever role the connection uses. Connections default to read-only, which rejects anything but a single read statement (SELECT/WITH/SHOW/EXPLAIN/TABLE/VALUES). On a read-write connection it will also run DDL (CREATE/DROP/ALTER) and DML (INSERT/UPDATE/DELETE/TRUNCATE), so before running anything that mutates data or schema, call ask_user_question to confirm.",
   input_schema: {
     type: "object",
-    required: ["sql"],
+    required: ["sql", "name"],
     properties: {
       sql: { type: "string", description: "SQL to execute." },
+      name: {
+        type: "string",
+        description: 'A short human-readable name for this query (3\u20136 words) describing what it does, e.g. "Monthly revenue by region". Shown in the UI next to the query.'
+      },
       connection_id: {
         type: "string",
         description: "Connection UUID. Defaults to the chat's bound connection."
@@ -2623,7 +2702,7 @@ var runSqlTool = {
 };
 
 // server/agent/tools/render_chart.ts
-var CHART_TYPES = ["bar", "line", "area", "pie"];
+var CHART_TYPES = ["bar", "stacked-bar", "line", "area", "pie", "funnel"];
 var MAX_ROWS2 = 500;
 var MAX_SERIES = 8;
 function parseSeries(raw) {
@@ -2664,7 +2743,7 @@ function parseData(raw) {
 }
 var renderChartTool = {
   name: "render_chart",
-  description: `Render a chart inline in the chat to visualize query results. Supply the data rows directly (shaped from your run_sql results) along with the chart spec \u2014 the chart is self-contained. Use this after run_sql when a visualization communicates the answer better than a table: trends over time (line/area), category comparisons (bar), or proportions (pie). Keep data small and pre-aggregated (cap ${MAX_ROWS2} rows). Each data row is an object keyed by column name, e.g. \`[{ "month": "Jan", "revenue": 120 }, \u2026]\`. \`x\` names the category column; \`series\` names the numeric column(s) to plot. For pie charts only the first series is used.`,
+  description: `Render a chart inline in the chat to visualize query results. Supply the data rows directly (shaped from your run_sql results) along with the chart spec \u2014 the chart is self-contained. Use this after run_sql when a visualization communicates the answer better than a table: trends over time (line/area), category comparisons (bar), part-to-whole across categories (stacked-bar), proportions (pie), or stage-by-stage drop-off like a conversion pipeline (funnel). Keep data small and pre-aggregated (cap ${MAX_ROWS2} rows). Each data row is an object keyed by column name, e.g. \`[{ "month": "Jan", "revenue": 120 }, \u2026]\`. \`x\` names the category column; \`series\` names the numeric column(s) to plot. For pie and funnel charts only the first series is used; for funnel, order rows from the widest stage to the narrowest. Pass \`sourceQuery\` \u2014 the \`name\` you gave the run_sql call the data came from \u2014 so the chart links back to its query.`,
   input_schema: {
     type: "object",
     required: ["type", "x", "series", "data"],
@@ -2672,7 +2751,7 @@ var renderChartTool = {
       type: {
         type: "string",
         enum: CHART_TYPES,
-        description: "Chart kind: bar, line, area, or pie."
+        description: "Chart kind: bar, stacked-bar, line, area, pie, or funnel."
       },
       title: { type: "string", description: "Optional title shown above the chart." },
       x: {
@@ -2695,6 +2774,10 @@ var renderChartTool = {
         type: "array",
         description: "Row objects to chart, keyed by column name.",
         items: { type: "object" }
+      },
+      sourceQuery: {
+        type: "string",
+        description: "The `name` of the run_sql call whose results this chart plots. Lets the UI link the chart back to its source query."
       }
     }
   },
@@ -2727,7 +2810,8 @@ var renderChartTool = {
       x: input.x,
       series,
       data,
-      ...typeof input.title === "string" && input.title.trim() !== "" ? { title: input.title } : {}
+      ...typeof input.title === "string" && input.title.trim() !== "" ? { title: input.title } : {},
+      ...typeof input.sourceQuery === "string" && input.sourceQuery.trim() !== "" ? { sourceQuery: input.sourceQuery.trim() } : {}
     };
     const seriesLabel = series.map((s) => s.label ?? s.key).join(", ");
     return {
@@ -2756,8 +2840,14 @@ function findTool(name) {
   return TOOL_BY_NAME[name];
 }
 function anthropicToolDefs(opts = {}) {
-  const { worksheetBound = true } = opts;
-  return TOOLS.filter((t) => worksheetBound || t.name !== "write_sql").map((t) => ({
+  const { worksheetBound = true, mode = "chat" } = opts;
+  const withheld = /* @__PURE__ */ new Set();
+  if (!worksheetBound) withheld.add("write_sql");
+  if (mode === "quick-edit") {
+    withheld.add("render_chart");
+    withheld.add("ask_user_question");
+  }
+  return TOOLS.filter((t) => !withheld.has(t.name)).map((t) => ({
     name: t.name,
     description: t.description,
     input_schema: t.input_schema
@@ -2820,7 +2910,7 @@ async function runAgentTurn(opts) {
   );
 }
 async function runTurn(opts) {
-  const { session, emit } = opts;
+  const { session, emit, signal } = opts;
   let apiKey;
   try {
     apiKey = await getAnthropicKey();
@@ -2829,8 +2919,12 @@ async function runTurn(opts) {
     return;
   }
   const system = buildSystemPrompt(session);
-  const tools = anthropicToolDefs({ worksheetBound: !!session.meta.worksheetSlug });
+  const tools = anthropicToolDefs({
+    worksheetBound: !!session.meta.worksheetSlug,
+    mode: session.meta.mode
+  });
   for (let step = 0; step < MAX_STEPS; step += 1) {
+    if (signal?.aborted) return;
     let final;
     let model;
     let usage;
@@ -2841,6 +2935,7 @@ async function runTurn(opts) {
         system,
         tools,
         messages: session.messages,
+        signal,
         onTextDelta: (delta) => {
           textEmitChain = textEmitChain.then(
             () => Promise.resolve(emit({ type: "text_delta", text: delta }))
@@ -2852,6 +2947,7 @@ async function runTurn(opts) {
       model = result.model;
       usage = result.usage;
     } catch (err) {
+      if (signal?.aborted) return;
       await emit({ type: "error", message: err.message });
       return;
     }
@@ -2892,6 +2988,15 @@ async function runTurn(opts) {
         });
         continue;
       }
+      if (signal?.aborted) {
+        resultBlocks.push({
+          type: "tool_result",
+          tool_use_id: block.id,
+          content: "Canceled: the user aborted this run before the tool executed.",
+          is_error: true
+        });
+        continue;
+      }
       await emit({
         type: "tool_start",
         toolUseId: block.id,
@@ -2940,7 +3045,8 @@ async function runTurn(opts) {
         toolUseId: block.id,
         name: tool.name,
         ok: !execution.isError,
-        summary: execution.uiSummary
+        summary: execution.uiSummary,
+        ...execution.detail ? { detail: execution.detail } : {}
       });
       resultBlocks.push({
         type: "tool_result",
@@ -2976,7 +3082,7 @@ async function runTurn(opts) {
   });
 }
 async function resumeWithAnswer(opts) {
-  const { session, userAnswer, emit } = opts;
+  const { session, userAnswer, emit, signal } = opts;
   const pending = session.meta.pending;
   if (!pending) {
     await emit({ type: "error", message: "No pending question to resume." });
@@ -2993,7 +3099,7 @@ async function resumeWithAnswer(opts) {
   }
   await persistSession(session);
   await clearPending(session);
-  await runAgentTurn({ session, emit });
+  await runAgentTurn({ session, emit, signal });
 }
 
 // server/api/agent.ts
@@ -3055,7 +3161,8 @@ app6.post("/sessions", async (c) => {
     worksheetSlug: body.worksheetSlug ?? null,
     connectionId: body.connectionId ?? null,
     title: body.title ?? null,
-    standalone: body.standalone ?? false
+    standalone: body.standalone ?? false,
+    mode: body.mode === "quick-edit" ? "quick-edit" : "chat"
   });
   return c.json(session, 201);
 });
@@ -3149,7 +3256,9 @@ app6.post("/sessions/:id/messages", async (c) => {
     );
   }
   return streamSSE(c, async (stream) => {
+    const signal = c.req.raw.signal;
     const emit = async (event) => {
+      if (signal.aborted) return;
       await stream.writeSSE({ data: JSON.stringify(event) });
     };
     try {
@@ -3166,11 +3275,12 @@ app6.post("/sessions/:id/messages", async (c) => {
           });
           return;
         }
+        trimQuickEditHistory(session);
         await appendMessage(session, { role: "user", content: message });
         if (!session.meta.title) {
           await setTitle(session, message.slice(0, 60));
         }
-        await runAgentTurn({ session, emit });
+        await runAgentTurn({ session, emit, signal });
       });
     } catch (err) {
       await emit({ type: "error", message: err.message });
@@ -3190,7 +3300,9 @@ app6.post("/sessions/:id/respond", async (c) => {
     return c.json({ error: "No pending question to answer." }, 409);
   }
   return streamSSE(c, async (stream) => {
+    const signal = c.req.raw.signal;
     const emit = async (event) => {
+      if (signal.aborted) return;
       await stream.writeSSE({ data: JSON.stringify(event) });
     };
     try {
@@ -3207,7 +3319,7 @@ app6.post("/sessions/:id/respond", async (c) => {
           });
           return;
         }
-        await resumeWithAnswer({ session, userAnswer: answer, emit });
+        await resumeWithAnswer({ session, userAnswer: answer, emit, signal });
       });
     } catch (err) {
       await emit({ type: "error", message: err.message });