tracer-sh 0.2.5 → 0.2.7

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tracer-sh",
-  "version": "0.2.5",
+  "version": "0.2.7",
   "type": "module",
   "description": "Local-first debugging & analysis platform",
   "license": "SEE LICENSE IN LICENSE",

package/packages/server/dist/index.js

@@ -16548,6 +16548,49 @@ var TOOL_NAMES = {
 var CLIENT_TOOL_NAMES = Object.fromEntries(
   Object.entries(TOOL_NAMES).map(([k, v]) => [k, `tool-${v}`])
 );
+var ANALYSIS_MARKER = "<analysis>";
+function isAnalysisMessage(msg) {
+  if (msg.role !== "assistant" || !msg.parts) return false;
+  return msg.parts.some((p) => {
+    if (p.type === CLIENT_TOOL_NAMES.BEGIN_ANALYSIS) return true;
+    if (p.type !== "text") return false;
+    const text3 = p.text;
+    return typeof text3 === "string" && text3.includes(ANALYSIS_MARKER);
+  });
+}
+function compactionUpTo(messages, boundaryIdx) {
+  const boundary = messages[boundaryIdx];
+  if (!boundary || boundary.role !== "assistant") return null;
+  if (!isAnalysisMessage(boundary)) return boundaryIdx + 1;
+  return boundaryIdx >= 1 ? boundaryIdx : null;
+}
+function findAnalysisMarker(parts) {
+  const toolIdx = parts.findIndex((p) => p.type === CLIENT_TOOL_NAMES.BEGIN_ANALYSIS);
+  if (toolIdx !== -1) return { kind: "tool", partIdx: toolIdx };
+  for (let i = 0; i < parts.length; i++) {
+    const p = parts[i];
+    if (p.type !== "text") continue;
+    const text3 = p.text;
+    if (typeof text3 !== "string") continue;
+    const idx = text3.indexOf(ANALYSIS_MARKER);
+    if (idx !== -1) return { kind: "text", partIdx: i, charIdx: idx };
+  }
+  return null;
+}
+function splitAtAnalysis(parts) {
+  const marker26 = findAnalysisMarker(parts);
+  if (!marker26) return null;
+  if (marker26.kind === "tool") {
+    return { before: parts.slice(0, marker26.partIdx), analysis: parts.slice(marker26.partIdx + 1) };
+  }
+  const p = parts[marker26.partIdx];
+  const beforeText = p.text.slice(0, marker26.charIdx);
+  const afterText = p.text.slice(marker26.charIdx + ANALYSIS_MARKER.length);
+  return {
+    before: [...parts.slice(0, marker26.partIdx), ...beforeText.trim() ? [{ ...p, text: beforeText }] : []],
+    analysis: [...afterText.trim() ? [{ ...p, text: afterText }] : [], ...parts.slice(marker26.partIdx + 1)]
+  };
+}
 var ImportedAnalysisSchema = external_exports.object({
   v: external_exports.literal(1),
   kind: external_exports.literal("analysis"),
@@ -20371,6 +20414,13 @@ var chatSessions = sqliteTable("chat_sessions", {
   status: text("status").notNull().default("idle"),
   /** Null for normal sessions. "imported" for sessions re-hydrated from a dropped analysis PNG. */
   kind: text("kind"),
+  /** Compaction: LLM-generated (possibly user-edited) summary of the first summaryUpTo messages. */
+  summary: text("summary"),
+  /** Count of leading messages covered by the summary. Index-based because
+   *  assistant messages can carry empty ids; prefixes are stable here (messages
+   *  are only appended or suffix-truncated, never reordered). */
+  summaryUpTo: integer2("summary_up_to"),
+  summaryCreatedAt: integer2("summary_created_at"),
   createdAt: integer2("created_at").notNull().$defaultFn(() => unixNow()),
   updatedAt: integer2("updated_at").notNull().$defaultFn(() => unixNow())
 }, (t2) => [
@@ -20514,6 +20564,9 @@ function runSetup() {
       messages TEXT NOT NULL,
       status TEXT NOT NULL DEFAULT 'idle',
       kind TEXT,
+      summary TEXT,
+      summary_up_to INTEGER,
+      summary_created_at INTEGER,
       created_at INTEGER NOT NULL DEFAULT (unixepoch()),
       updated_at INTEGER NOT NULL DEFAULT (unixepoch())
     );
@@ -20617,7 +20670,12 @@ function runSetup() {
   for (const ddl of [
     `ALTER TABLE sub_agent_runs ADD COLUMN session_id TEXT`,
     `ALTER TABLE tool_memories ADD COLUMN review_note TEXT`,
-    `ALTER TABLE chat_sessions ADD COLUMN kind TEXT`
+    `ALTER TABLE chat_sessions ADD COLUMN kind TEXT`,
+    `ALTER TABLE chat_sessions ADD COLUMN summary TEXT`,
+    `ALTER TABLE chat_sessions ADD COLUMN summary_up_to INTEGER`,
+    `ALTER TABLE chat_sessions ADD COLUMN summary_created_at INTEGER`,
+    // Drops the short-lived id-based boundary column (never shipped in a release).
+    `ALTER TABLE chat_sessions DROP COLUMN summary_up_to_id`
   ]) {
     try {
       sqlite.exec(ddl);
@@ -37940,7 +37998,7 @@ function injectMemories(prompt, memoryContext) {
   const memoryBlock = `
 
 ## ${MEMORY_SECTION_NAME}
-These OVERRIDE any conflicting instructions above \u2014 they are verified fixes from past errors:
+These OVERRIDE any conflicting instructions above \u2014 they are lessons recorded from real failures in past sessions:
 ${lines.join("\n")}
 `;
   const firstBreak = prompt.indexOf("\n\n");
@@ -46233,6 +46291,8 @@ You MUST save a memory for every query failure:
 - Wrong field names \u2192 "Don't use X, use Y for [purpose]"
 - Wrong event types \u2192 "Don't query X for [goal], use Y instead"
 
+**Corrections must be demonstrated, not guessed.** Only write "use Y instead" if Y was actually run in this session and succeeded. If no working alternative was demonstrated, save only the mistake ("Don't use X in [context]") \u2014 a wrong correction is worse than none, because memories override future instructions.
+
 ### From struggle patterns (evaluated)
 Review the full session timeline for patterns where the agent struggled \u2014 multiple attempts with variations before finding the correct approach. Look for:
 - Repeated EMPTY results with name/field/value variations followed by eventual success
@@ -46591,7 +46651,7 @@ function sanitizeNrqlRows(rows) {
 }
 
 // src/lib/shared-prompts.ts
-var UNIFIED_ROLE_INTRO = `You are Tracer, an observability expert in a direct conversation with a developer. You have DIRECT access to the query tools of multiple providers at once \u2014 each provider's syntax, fields, and debugging guidance are documented below. Pick the right provider(s) for each question; when a question spans providers, query them and correlate across the results in one investigation. You have full conversation history and can reference previous messages. You run as an AUTONOMOUS MULTI-STEP AGENT \u2014 after each tool call you automatically receive results and CAN (and often SHOULD) make additional tool calls before finishing.`;
+var UNIFIED_ROLE_INTRO = `You are Tracer, an observability expert in a direct conversation with a developer. You have DIRECT access to the query tools of multiple providers at once \u2014 each provider's syntax, common fields, and debugging guidance are documented below. Pick the right provider(s) for each question; when a question spans providers, query them and correlate across the results in one investigation. You have full conversation history and can reference previous messages. You run as an AUTONOMOUS MULTI-STEP AGENT \u2014 after each tool call you automatically receive results and CAN (and often SHOULD) make additional tool calls before finishing.`;
 function buildUnifiedModePrompt(providerFragments, maxSteps) {
   return `${UNIFIED_ROLE_INTRO}
 
@@ -46600,6 +46660,8 @@ ${buildRules({ investigation: true })}
 
 ${DETECTIVE_MINDSET}
 
+${EVIDENCE_GROUNDING}
+
 ${EXECUTION_DISCIPLINE}
 
 ${providerFragments.join("\n\n---\n\n")}
@@ -46641,13 +46703,20 @@ You have limited steps. Every query must earn its place. Your goal is the **fast
 3. **"Is there a single query that could answer multiple questions at once?"** \u2014 Combine work. Pack information density per query.
 
 **"Good enough" beats "complete."** The user can always ask follow-up questions. Don't anticipate them \u2014 answer what was asked.`;
+var EVIDENCE_GROUNDING = `## Grounded in Evidence
+
+Your only sources of truth are the literal text of tool results from this session, what the user has stated in the conversation, and what this prompt documents. Anything else is unknown \u2014 including the meaning of the data you retrieve.
+
+1. **Field names and values are opaque labels.** Never translate or assign meaning to a field name, enum value, code, or flag beyond its literal text \u2014 systems attach internal meanings you cannot know. Report the raw value; if its meaning matters and is undocumented, say so.
+2. **Absence requires an empty probe.** Only claim something is missing, absent, or "not on file" if a query that would have returned it came back empty. Not having looked is not evidence of absence.
+3. **Separate facts, deductions, and gaps.** Facts restate query results. Deductions must follow from stated facts alone \u2014 present them as deductions and name the supporting results; correlation across results is not causation. Gaps are reported as "the data does not show X" \u2014 never filled with a plausible story.`;
 var NO_FIXES_RULE = `**NEVER suggest fixes, remediation, next steps, or actions.** Forbidden phrasings include: "consider," "you should," "try," "might want to," "recommend," "could help," "suggests [action]," "would resolve," "to fix this." Any sentence about what to DO about the problem is forbidden, regardless of phrasing. Your job ends at "here is what happened and the evidence." The developer decides what to do.`;
 var EXECUTION_DISCIPLINE = `## Execution Discipline
 
 For multi-step investigations:
 1. **Step N: [Goal]** \u2014 state what gap this fills
 2. **Tool call** \u2192 ONE query
-3. **\u2192 Found:** [data] **\u2192 So what:** [inference]
+3. **\u2192 Found:** [data] **\u2192 So what:** [only what this data supports \u2014 if it needs an assumption, it's a gap, not a finding]
 4. **\u2192 Can I answer now?** \u2014 If YES: respond. If NO: state what's missing.
 
 For simple questions (counts, lookups), skip this \u2014 just answer directly.`;
@@ -46661,6 +46730,7 @@ function analysisBlock() {
 
 1. **Think first** \u2014 before writing anything, plan the evidence chain in your head:
    - Known facts from query results, inferences that follow from them, and remaining gaps.
+   - Self-audit each claim against the tool results already in this session: if no specific result backs it, drop it or present it explicitly as unverified. This is an in-head check \u2014 never run extra investigation queries for it.
    - Which queries best VISUALIZE each finding \u2014 these become the tool calls you will run in this section.
    - Do not start writing until you have a clear chain and a concrete list of visuals to run.
 2. ${markerStep}
@@ -46686,7 +46756,7 @@ You have a maximum of ${maxSteps} steps. Most investigations should finish in 3-
 
 ## Final Reminders
 - **Tool calls are the evidence.** Every substantive claim in your response needs a visual \u2014 even if the same query already ran during investigation, re-run it here. The analysis section must be self-contained.
-- **Follow the Detective mindset:** correlation \u2260 causation, no gap-filling, no fixes. Every claim traces to a specific query result. Say "insufficient data" when data is missing.`;
+- **Stay Grounded in Evidence:** every claim maps to a specific tool result; values mean only what their literal text says; absence claims need an empty probe; gaps are stated as "the data does not show". No fixes.`;
 }
 
 // src/lib/prompt-builder.ts
@@ -46715,6 +46785,8 @@ ${buildRules({ investigation: true, extraRules: config2.extraRules })}
 
 ${DETECTIVE_MINDSET}
 
+${EVIDENCE_GROUNDING}
+
 ${config2.insideOutDebugging}
 
 ${EXECUTION_DISCIPLINE}
@@ -56769,6 +56841,137 @@ var memoryRouter = router({
   })
 });
 
+// src/agents/utility/summary.ts
+var SUMMARY_SYSTEM_PROMPT = `You are compacting an AI debugging-assistant conversation into a detailed summary. Your summary will permanently REPLACE the original messages as the assistant's only memory of them, so anything you omit is lost forever. The assistant must be able to continue the investigation from your summary alone without redoing any completed work.
+
+Write the summary as markdown with exactly these sections:
+
+## Original intent
+What the user set out to do, in their own framing. Include later refinements or pivots of the goal.
+
+## User requests & decisions
+Chronological list of every instruction, question, constraint, correction, and approval or rejection the user gave, and whether each was fulfilled. The assistant must be able to tell from this section alone what the user has and has not asked for.
+
+## Investigation log
+Numbered, chronological. Each entry pairs an action with its result:
+1. One line on what was done and why, then the exact tool call, query, or API call (with its parameters and the time range it covered) in a fenced code block.
+
+   \u2192 Result: what it returned, with the key values verbatim. Render a result with more than one row (facets, top-N lists, table rows) as a markdown table; copy up to roughly 20 rows verbatim and note what was cut.
+
+Never record an action without its result \u2014 an unpaired action forces the assistant to re-run it.
+
+## Key findings & data (verbatim)
+The distilled facts the investigation established, with exact values copied verbatim \u2014 never paraphrase these:
+- IDs of any kind (trace IDs, entity GUIDs, account/project IDs, session IDs)
+- Exact error messages and stack-trace lines
+- File paths, service names, host names, URLs
+- Numbers: counts, rates, percentages, latencies, timestamps, time ranges
+Do not re-copy queries or result tables that already appear in the Investigation log \u2014 state the facts they established and name the log entry they came from.
+
+## What did NOT work (dead ends)
+Approaches tried and abandoned, queries that errored or returned empty, hypotheses ruled out \u2014 and WHY each failed. This prevents the assistant from repeating them. If nothing failed, write "None."
+
+## Conclusions & current state
+Each conclusion the assistant reached, stated together with the evidence supporting it, so it is never re-derived. What was communicated or delivered to the user (answers, recommendations, reports), and any artifacts produced.
+
+## Open items
+Unresolved questions, pending next steps, anything the user asked for that has not been delivered yet. If none, write "None."
+
+Rules:
+- Be detailed. Length is not a concern; losing information is. A long, precise summary is always better than a short, vague one.
+- Copy identifiers, queries, errors, and numbers character-for-character from the conversation.
+- Format for scanning: queries and commands go in fenced code blocks, multi-row results in markdown tables, and inline identifiers (service names, error classes, IDs, paths) in backticks.
+- State each piece of data in full exactly once, in the section where it belongs; later mentions reference it instead of repeating it.
+- Always pair what was run with what it returned.
+- If the conversation contains an analysis or post-mortem report (the begin_analysis tool marks where one starts), carry its content through verbatim in the relevant sections instead of re-summarizing it.
+- If an existing summary of older messages is provided, merge it with the new segment into ONE self-contained summary covering everything. Preserve all verbatim data from the existing summary unless the new segment explicitly supersedes it.
+- Do not add commentary, advice, or information that is not in the conversation.
+- Output only the summary markdown, nothing else.`;
+var TOOL_INPUT_CHAR_LIMIT = 2e3;
+var TOOL_OUTPUT_CHAR_LIMIT = 6e3;
+var GENERATION_TIMEOUT_MS = 5 * 6e4;
+function truncate(value, limit) {
+  if (value === void 0) return "(none)";
+  let text3;
+  try {
+    text3 = typeof value === "string" ? value : JSON.stringify(
+      value,
+      (_key, v) => typeof v === "string" && v.length > limit ? v.slice(0, limit) : v
+    );
+  } catch {
+    text3 = String(value);
+  }
+  return text3.length > limit ? `${text3.slice(0, limit)}\u2026 (truncated)` : text3;
+}
+function serializeMessagesForSummary(messages) {
+  const blocks = [];
+  messages.forEach((msg, i) => {
+    const lines = [`### Message ${i + 1} \u2014 ${msg.role}`];
+    for (const part of msg.parts) {
+      if (part.type === "text") {
+        lines.push(part.text);
+        continue;
+      }
+      if (part.type === "reasoning" || part.type === "step-start" || part.type.startsWith("data-") || part.type === "file") {
+        continue;
+      }
+      const p = part;
+      if (p.toolCallId) {
+        const name26 = p.type.startsWith("tool-") ? p.type.slice(5) : p.type;
+        const output = p.output !== void 0 ? p.output : p.errorText !== void 0 ? `error: ${p.errorText}` : void 0;
+        lines.push(
+          `[tool: ${name26}]`,
+          `input: ${truncate(p.input, TOOL_INPUT_CHAR_LIMIT)}`,
+          `output: ${truncate(output, TOOL_OUTPUT_CHAR_LIMIT)}`
+        );
+      }
+    }
+    blocks.push(lines.join("\n"));
+  });
+  return blocks.join("\n\n");
+}
+async function generateSessionSummary(db2, opts) {
+  const resolved = resolveModel(db2);
+  if ("error" in resolved) {
+    console.warn("[summary] Cannot generate summary:", resolved.error);
+    return { error: resolved.error, config: true };
+  }
+  const serialized = serializeMessagesForSummary(opts.messages);
+  let userContent = opts.priorSummary ? `## Existing summary of older messages (merge into your output)
+${opts.priorSummary}
+
+## New conversation segment to incorporate
+${serialized}` : `## Conversation to summarize
+${serialized}`;
+  if (opts.keptAnalysis) {
+    userContent += `
+
+Note: the assistant's final analysis of the last exchange is preserved verbatim in the conversation right after your summary \u2014 record the work and results above without inventing or restating its conclusions.`;
+  }
+  try {
+    const { text: text3, usage } = await generateText({
+      model: resolved.model,
+      temperature: 0,
+      system: SUMMARY_SYSTEM_PROMPT,
+      messages: [{ role: "user", content: userContent }],
+      providerOptions: resolved.providerOptions,
+      abortSignal: AbortSignal.timeout(GENERATION_TIMEOUT_MS)
+    });
+    const summary = text3.trim();
+    if (!summary) return { error: "Summary generation returned no content" };
+    recordAgentRun(db2, {
+      sessionId: opts.sessionId,
+      agentType: "summary",
+      model: resolved.modelId,
+      usage: extractUsage(usage, resolved.modelId)
+    });
+    return { summary };
+  } catch (err) {
+    console.warn("[summary] Failed to generate summary:", err);
+    return { error: "Summary generation failed" };
+  }
+}
+
 // src/trpc/routers/sessions.router.ts
 var AGENT_TYPE_LABELS = {
   chat: "Chat",
@@ -56776,7 +56979,8 @@ var AGENT_TYPE_LABELS = {
   gcp: "GCP sub-agent",
   posthog: "PostHog sub-agent",
   title: "Title gen",
-  memory: "Memory"
+  memory: "Memory",
+  summary: "Compaction"
 };
 var sessionsRouter = router({
   list: publicProcedure.query(({ ctx }) => {
@@ -56811,7 +57015,10 @@ var sessionsRouter = router({
       status: row.status,
       kind: row.kind,
       messages,
-      updatedAt: row.updatedAt
+      updatedAt: row.updatedAt,
+      summary: row.summary,
+      summaryUpTo: row.summaryUpTo,
+      summaryCreatedAt: row.summaryCreatedAt
     };
   }),
   getCost: publicProcedure.input(external_exports.object({ id: external_exports.string() })).query(({ ctx, input }) => {
@@ -56904,19 +57111,104 @@ var sessionsRouter = router({
     return { id };
   }),
   truncateMessages: publicProcedure.input(external_exports.object({ id: external_exports.string(), keepCount: external_exports.number().int().min(0) })).mutation(({ ctx, input }) => {
-    const row = ctx.db.select({ messages: chatSessions.messages }).from(chatSessions).where(eq(chatSessions.id, input.id)).get();
-    if (!row) return { success: false };
+    const row = ctx.db.select({ messages: chatSessions.messages, summaryUpTo: chatSessions.summaryUpTo }).from(chatSessions).where(eq(chatSessions.id, input.id)).get();
+    if (!row) return { success: false, summaryCleared: false };
     let messages = [];
     try {
       messages = JSON.parse(row.messages);
     } catch {
-      return { success: false };
+      return { success: false, summaryCleared: false };
     }
     const truncated = messages.slice(0, input.keepCount);
+    const keptAnalysisBoundary = row.summaryUpTo != null && isAnalysisMessage(messages[row.summaryUpTo] ?? { role: "" });
+    const sourceUpTo = row.summaryUpTo != null ? row.summaryUpTo + (keptAnalysisBoundary ? 1 : 0) : 0;
+    const summaryStale = row.summaryUpTo != null && input.keepCount < sourceUpTo;
     ctx.db.update(chatSessions).set({
       messages: JSON.stringify(truncated),
-      updatedAt: unixNow()
+      updatedAt: unixNow(),
+      ...summaryStale ? { summary: null, summaryUpTo: null, summaryCreatedAt: null } : {}
     }).where(eq(chatSessions.id, input.id)).run();
+    return { success: true, summaryCleared: summaryStale };
+  }),
+  compact: publicProcedure.input(external_exports.object({ id: external_exports.string(), upToIndex: external_exports.number().int().min(0) })).mutation(async ({ ctx, input }) => {
+    const row = ctx.db.select().from(chatSessions).where(eq(chatSessions.id, input.id)).get();
+    if (!row) throw new TRPCError({ code: "NOT_FOUND", message: "Session not found" });
+    if (row.status === "streaming" || ctx.activeStreams.has(input.id)) {
+      throw new TRPCError({ code: "CONFLICT", message: "Cannot compact while a response is in progress" });
+    }
+    let messages;
+    try {
+      messages = JSON.parse(row.messages);
+    } catch {
+      throw new TRPCError({ code: "INTERNAL_SERVER_ERROR", message: "Session messages are corrupted" });
+    }
+    const boundaryIdx = input.upToIndex;
+    if (boundaryIdx >= messages.length) {
+      throw new TRPCError({ code: "BAD_REQUEST", message: "Message not found in session" });
+    }
+    if (messages[boundaryIdx].role !== "assistant") {
+      throw new TRPCError({ code: "BAD_REQUEST", message: "Can only compact up to an assistant message" });
+    }
+    const summaryUpTo = compactionUpTo(messages, boundaryIdx);
+    if (summaryUpTo == null) {
+      throw new TRPCError({ code: "BAD_REQUEST", message: "There are no messages to summarize before the analysis" });
+    }
+    const keptAnalysis = isAnalysisMessage(messages[boundaryIdx]);
+    const incremental = !!row.summary && row.summaryUpTo != null && row.summaryUpTo < summaryUpTo;
+    const segment = messages.slice(incremental ? row.summaryUpTo : 0, boundaryIdx + 1);
+    if (incremental) {
+      const head = segment[0];
+      if (isAnalysisMessage(head)) {
+        const split = splitAtAnalysis(head.parts);
+        if (split) segment[0] = { ...head, parts: split.analysis };
+      }
+    }
+    if (keptAnalysis) {
+      const last = segment[segment.length - 1];
+      const split = splitAtAnalysis(last.parts);
+      if (split) segment[segment.length - 1] = { ...last, parts: split.before };
+    }
+    const result = await generateSessionSummary(ctx.db, {
+      sessionId: input.id,
+      priorSummary: incremental ? row.summary : void 0,
+      messages: segment,
+      keptAnalysis
+    });
+    if ("error" in result) {
+      throw new TRPCError({
+        code: result.config ? "PRECONDITION_FAILED" : "INTERNAL_SERVER_ERROR",
+        message: result.error
+      });
+    }
+    const fresh = ctx.db.select({ messages: chatSessions.messages, status: chatSessions.status }).from(chatSessions).where(eq(chatSessions.id, input.id)).get();
+    const sourceUpTo = boundaryIdx + 1;
+    let prefixUnchanged = false;
+    if (fresh && fresh.status !== "streaming" && !ctx.activeStreams.has(input.id)) {
+      try {
+        const freshMessages = JSON.parse(fresh.messages);
+        prefixUnchanged = JSON.stringify(freshMessages.slice(0, sourceUpTo)) === JSON.stringify(messages.slice(0, sourceUpTo));
+      } catch {
+      }
+    }
+    if (!prefixUnchanged) {
+      throw new TRPCError({ code: "CONFLICT", message: "The conversation changed while the summary was being generated \u2014 try again" });
+    }
+    const summary = result.summary;
+    const summaryCreatedAt = unixNow();
+    ctx.db.update(chatSessions).set({ summary, summaryUpTo, summaryCreatedAt }).where(eq(chatSessions.id, input.id)).run();
+    return { summary, summaryUpTo, summaryCreatedAt };
+  }),
+  updateSummary: publicProcedure.input(external_exports.object({ id: external_exports.string(), summary: external_exports.string().min(1) })).mutation(({ ctx, input }) => {
+    const row = ctx.db.select({ summaryUpTo: chatSessions.summaryUpTo }).from(chatSessions).where(eq(chatSessions.id, input.id)).get();
+    if (!row) throw new TRPCError({ code: "NOT_FOUND", message: "Session not found" });
+    if (row.summaryUpTo == null) {
+      throw new TRPCError({ code: "CONFLICT", message: "The summary no longer exists" });
+    }
+    ctx.db.update(chatSessions).set({ summary: input.summary }).where(eq(chatSessions.id, input.id)).run();
+    return { success: true };
+  }),
+  clearSummary: publicProcedure.input(external_exports.object({ id: external_exports.string() })).mutation(({ ctx, input }) => {
+    ctx.db.update(chatSessions).set({ summary: null, summaryUpTo: null, summaryCreatedAt: null }).where(eq(chatSessions.id, input.id)).run();
     return { success: true };
   })
 });
@@ -57495,7 +57787,11 @@ function loadSessionMessages(db2, sessionId, newMessage) {
       console.warn(`[chat] Corrupted session ${sessionId}, starting fresh`);
     }
   }
-  return [...sanitizeMessages(previous), newMessage];
+  return {
+    messages: [...sanitizeMessages(previous), newMessage],
+    summary: existing?.summary ?? null,
+    summaryUpTo: existing?.summaryUpTo ?? null
+  };
 }
 function finalizeSession(sessionId, context2, broadcaster) {
   if (!context2.activeStreams.has(sessionId)) return;
@@ -57503,7 +57799,7 @@ function finalizeSession(sessionId, context2, broadcaster) {
   broadcaster.finish();
   context2.activeStreams.delete(sessionId);
 }
-async function processLLMStream(sessionId, messages, context2, broadcaster, serverAbort, collectTools, sessionTitle, model, modelId, providerOptions) {
+async function processLLMStream(sessionId, messages, context2, broadcaster, serverAbort, collectTools, sessionTitle, model, modelId, providerOptions, compaction) {
   const writer = {
     write: (part) => {
       const p = part;
@@ -57514,7 +57810,29 @@ async function processLLMStream(sessionId, messages, context2, broadcaster, serv
   };
   const collected = collectTools(writer);
   const tools = collected.tools;
-  const modelMessages = await convertToModelMessages(messages, {
+  let modelInput = messages;
+  let summaryForPrompt = null;
+  if (compaction.summary && compaction.summaryUpTo && compaction.summaryUpTo < messages.length) {
+    const tail = messages.slice(compaction.summaryUpTo);
+    if (tail[0].role === "user") {
+      modelInput = tail;
+      summaryForPrompt = compaction.summary;
+    } else {
+      const split = splitAtAnalysis(tail[0].parts);
+      if (split && split.analysis.length > 0) {
+        tail[0] = { ...tail[0], parts: split.analysis };
+        modelInput = [
+          { id: "", role: "user", parts: [{ type: "text", text: "(The conversation up to this point was compacted into the summary in your instructions.)" }] },
+          ...tail
+        ];
+        summaryForPrompt = compaction.summary;
+      }
+    }
+  }
+  if (compaction.summary && !summaryForPrompt) {
+    console.warn(`[chat] Ignoring stale compaction boundary for ${sessionId} (summaryUpTo=${compaction.summaryUpTo}, messages=${messages.length})`);
+  }
+  const modelMessages = await convertToModelMessages(modelInput, {
     tools,
     convertDataPart: () => void 0
   });
@@ -57535,6 +57853,16 @@ ${fragments.join("\n\n")}` : `${basePrompt}
 No observability providers are currently configured. If the user asks about observability data, let them know they can connect providers in the Settings page.`;
   }
   systemPrompt += "\n\n" + getCurrentDateBlock(context2.db);
+  if (summaryForPrompt) {
+    systemPrompt += `
+
+## Earlier conversation summary
+The earlier part of this conversation was compacted to save context. The summary below replaces those messages and is authoritative: the work it describes is already done \u2014 do NOT redo it. Reuse its recorded results, identifiers, queries, and conclusions.
+
+<conversation_summary>
+${summaryForPrompt}
+</conversation_summary>`;
+  }
   const result = streamText({
     model,
     temperature: 0,
@@ -57648,7 +57976,7 @@ No observability providers are currently configured. If the user asks about obse
   clearTimeout(timeoutId);
   finalizeSession(sessionId, context2, broadcaster);
 }
-async function runChatAgent({ sessionId, messages, context: context2, collectTools, sessionTitle, modelOverride }) {
+async function runChatAgent({ sessionId, messages, summary, summaryUpTo, context: context2, collectTools, sessionTitle, modelOverride }) {
   const resolved = modelOverride ?? resolveModel(context2.db);
   if ("error" in resolved) return { error: resolved.error };
   const { model, modelId, providerOptions } = resolved;
@@ -57680,7 +58008,8 @@ async function runChatAgent({ sessionId, messages, context: context2, collectToo
     sessionTitle,
     model,
     modelId,
-    providerOptions
+    providerOptions,
+    { summary, summaryUpTo }
   ).catch((err) => {
     console.error(`[chat] Unhandled error in LLM processing for ${sessionId}:`, err);
     finalizeSession(sessionId, context2, broadcaster);
@@ -57804,7 +58133,7 @@ function nextYPosition(db2, dashboardId) {
 }
 function collectDashboardTools(registry2, db2, writer, dashboardId) {
   const dbId = dashboardId ?? "";
-  const { tools, promptFragments, connectedProviders } = collectBaseTools(registry2, db2, writer);
+  const { tools, promptFragments, connectedProviders } = collectBaseTools(registry2, db2, writer, "unified");
   const defaultProvider = connectedProviders[0];
   tools.create_widget = tool({
     description: "Create a new dashboard widget. The query will be validated by executing it first. The widget auto-positions below existing widgets.",
@@ -57943,7 +58272,7 @@ The global date picker already shows the active time range, so titles should des
 
 ## Scope
 You are managing widgets for the current dashboard only. The widget list above shows only this dashboard's widgets.`;
-  const systemPrompt = [basePrompt, providerContext, widgetContext, ...promptFragments].join("\n\n");
+  const systemPrompt = [basePrompt, EVIDENCE_GROUNDING, providerContext, widgetContext, ...promptFragments].join("\n\n");
   return { tools, systemPrompt };
 }
 
@@ -58250,7 +58579,7 @@ function getMonitorContext(db2) {
 ${lines.join("\n")}`;
 }
 function collectMonitorTools(registry2, db2, writer) {
-  const { tools, promptFragments, connectedProviders } = collectBaseTools(registry2, db2, writer);
+  const { tools, promptFragments, connectedProviders } = collectBaseTools(registry2, db2, writer, "unified");
   const defaultProvider = connectedProviders[0];
   tools.create_monitor = tool({
     description: "Create a new monitor that periodically checks a query and alerts when a condition is met.",
@@ -58408,7 +58737,7 @@ The condition is a JS expression evaluated against the query \`result\` array. E
 
 ## Scope
 You are managing all monitors. The monitor list above shows all existing monitors.`;
-  const systemPrompt = [basePrompt, providerContext, monitorContext, ...promptFragments].join("\n\n");
+  const systemPrompt = [basePrompt, EVIDENCE_GROUNDING, providerContext, monitorContext, ...promptFragments].join("\n\n");
   return { tools, systemPrompt };
 }
 
@@ -58448,7 +58777,7 @@ function generateSessionTitle(db2, sessionId, userMessage) {
 function registerChatRoutes(app, context2) {
   app.post("/api/chat", async (c) => {
     const { id, message, activeProvider } = await c.req.json();
-    const messages = loadSessionMessages(context2.db, id, message);
+    const { messages, summary, summaryUpTo } = loadSessionMessages(context2.db, id, message);
     if (messages.length === 1) {
       const textPart = message.parts?.find((p) => p.type === "text");
       if (textPart) {
@@ -58462,6 +58791,8 @@ function registerChatRoutes(app, context2) {
     const result = await runChatAgent({
       sessionId: id,
       messages,
+      summary,
+      summaryUpTo,
       context: context2,
       collectTools: (writer) => collectChatTools(context2.providers, context2.db, writer, scopedProvider, mode),
       sessionTitle: (updatedMessages) => {
@@ -58477,10 +58808,12 @@ function registerChatRoutes(app, context2) {
   app.post("/api/dashboard-chat", async (c) => {
     const { id, message, dashboardId } = await c.req.json();
     const sessionId = dashboardSessionId(dashboardId);
-    const messages = loadSessionMessages(context2.db, sessionId, message);
+    const { messages, summary, summaryUpTo } = loadSessionMessages(context2.db, sessionId, message);
     const result = await runChatAgent({
       sessionId,
       messages,
+      summary,
+      summaryUpTo,
       context: context2,
       collectTools: (writer) => collectDashboardTools(context2.providers, context2.db, writer, dashboardId),
       sessionTitle: () => "Dashboard Builder"
@@ -58491,10 +58824,12 @@ function registerChatRoutes(app, context2) {
   app.post("/api/monitor-chat", async (c) => {
     const { message } = await c.req.json();
     const sessionId = SESSION_PREFIX.MONITORS;
-    const messages = loadSessionMessages(context2.db, sessionId, message);
+    const { messages, summary, summaryUpTo } = loadSessionMessages(context2.db, sessionId, message);
     const result = await runChatAgent({
       sessionId,
       messages,
+      summary,
+      summaryUpTo,
       context: context2,
       collectTools: (writer) => collectMonitorTools(context2.providers, context2.db, writer),
       sessionTitle: () => "Monitor Builder"
@@ -58645,7 +58980,7 @@ function registerApiRoutes(app, context2) {
       role: "user",
       parts: [{ type: "text", text: message }]
     };
-    const messages = loadSessionMessages(context2.db, sessionId, userMessage);
+    const { messages, summary, summaryUpTo } = loadSessionMessages(context2.db, sessionId, userMessage);
     if (messages.length === 1) {
       generateSessionTitle(context2.db, sessionId, message);
     }
@@ -58658,6 +58993,8 @@ function registerApiRoutes(app, context2) {
     const result = await runChatAgent({
       sessionId,
       messages,
+      summary,
+      summaryUpTo,
       context: context2,
       collectTools: (writer) => {
         const collected = collectChatTools(context2.providers, context2.db, writer, scopedProvider, mode);

package/packages/web/dist/assets/{SearchableSelect-B006wjjM.js → SearchableSelect-BPzDtKhB.js}

@@ -1 +1 @@
-import{a as q,r as s,j as r}from"./index-B7y1naN3.js";var P=q();function $(c){const n=c?`tracer:starred:${c}`:null,[i,S]=s.useState(()=>{if(!n)return new Set;try{const a=localStorage.getItem(n);return a?new Set(JSON.parse(a)):new Set}catch{return new Set}});return[i,a=>{S(u=>{const t=new Set(u);return t.has(a)?t.delete(a):t.add(a),n&&localStorage.setItem(n,JSON.stringify([...t])),t})}]}function I({options:c,value:n,onChange:i,placeholder:S="Select...",storageKey:v,fitContent:a,disabled:u}){const[t,l]=s.useState(!1),[w,y]=s.useState(""),f=s.useRef(null),j=s.useRef(null),N=s.useRef(null),g=s.useRef(null),[x,C]=$(v),[m,L]=s.useState({top:0,left:0,minWidth:0});s.useEffect(()=>{if(!t)return;const e=d=>{f.current?.contains(d.target)||j.current?.contains(d.target)||l(!1)};return document.addEventListener("mousedown",e),()=>document.removeEventListener("mousedown",e)},[t]);const R=s.useCallback(()=>{if(!f.current)return;const e=f.current.getBoundingClientRect();L({top:e.bottom+4,left:e.left,minWidth:e.width})},[]);s.useEffect(()=>{t&&(R(),y(""),requestAnimationFrame(()=>N.current?.focus()))},[t,R]);const h=c.find(e=>e.value===n),p=s.useMemo(()=>{const e=w.toLowerCase();return[...e?c.filter(o=>o.label.toLowerCase().includes(e)||o.value.toLowerCase().includes(e)):c].sort((o,b)=>{const k=x.has(o.value)?0:1,E=x.has(b.value)?0:1;return k!==E?k-E:o.label.localeCompare(b.label)})},[c,w,x]);s.useEffect(()=>{if(t&&n&&g.current){const e=g.current.querySelector(`[data-value="${CSS.escape(n)}"]`);e&&e.scrollIntoView({block:"nearest"})}},[t,n]);const W=t?P.createPortal(r.jsxs("div",{ref:j,className:"fixed z-[100] bg-white border border-[#d4d2cd] rounded shadow-lg",style:{top:m.top,left:m.left,minWidth:m.minWidth,width:a?"max-content":m.minWidth},children:[r.jsx("div",{className:"p-1.5 border-b border-[#e8e6e1]",children:r.jsx("input",{ref:N,type:"text",value:w,onChange:e=>y(e.target.value),placeholder:"Search...",className:"w-full px-2 py-1.5 text-xs text-[#2c2c2c] font-sans bg-[#f5f4f0] border border-[#e8e6e1] rounded focus:outline-none focus:border-[#2b5ea7] placeholder:text-[#9c9890]",onKeyDown:e=>{e.key==="Escape"&&l(!1),e.key==="Enter"&&p.length>0&&(i(p[0].value),l(!1))}})}),r.jsx("div",{ref:g,className:"max-h-[280px] overflow-y-auto",children:p.length===0?r.jsx("div",{className:"px-3 py-3 text-xs text-[#9c9890] text-center",children:"No projects found"}):p.map(e=>{const d=e.value===n,o=x.has(e.value);return r.jsxs("div",{"data-value":e.value,className:`flex items-center gap-1.5 px-2 py-1.5 text-xs font-sans cursor-pointer transition-colors ${d?"bg-[#2b5ea7]/10 text-[#2b5ea7]":"text-[#2c2c2c] hover:bg-[#f5f4f0]"}`,onClick:()=>{i(e.value),l(!1)},children:[r.jsx("button",{type:"button",onClick:b=>{b.stopPropagation(),C(e.value)},className:`shrink-0 w-4 h-4 flex items-center justify-center text-[10px] transition-colors ${o?"text-[#d4a017]":"text-[#d4d2cd] hover:text-[#9c9890]"}`,title:o?"Unstar":"Star to pin to top",children:o?"★":"☆"}),r.jsx("span",{className:a?"whitespace-nowrap":"truncate flex-1",children:e.label})]},e.value)})})]}),document.body):null;return r.jsxs(r.Fragment,{children:[r.jsxs("button",{ref:f,type:"button",onClick:()=>!u&&l(!t),disabled:u,className:"w-full bg-white border border-[#d4d2cd] rounded px-3 py-2 text-xs text-[#2c2c2c] font-sans text-left flex items-center justify-between focus:outline-none focus:border-[#2b5ea7] hover:border-[#b0ada6] transition-colors disabled:opacity-50 disabled:cursor-not-allowed",children:[r.jsx("span",{className:h?"text-[#2c2c2c] truncate":"text-[#9c9890]",children:h?h.displayLabel??h.label:S}),r.jsx("span",{className:"text-[#9c9890] text-[10px] ml-2 shrink-0 transition-transform duration-200",style:{transform:t?"rotate(180deg)":"rotate(0deg)"},children:"▾"})]}),W]})}export{I as S,P as r,$ as u};
+import{a as q,r as s,j as r}from"./index-IihLiR83.js";var P=q();function $(c){const n=c?`tracer:starred:${c}`:null,[i,S]=s.useState(()=>{if(!n)return new Set;try{const a=localStorage.getItem(n);return a?new Set(JSON.parse(a)):new Set}catch{return new Set}});return[i,a=>{S(u=>{const t=new Set(u);return t.has(a)?t.delete(a):t.add(a),n&&localStorage.setItem(n,JSON.stringify([...t])),t})}]}function I({options:c,value:n,onChange:i,placeholder:S="Select...",storageKey:v,fitContent:a,disabled:u}){const[t,l]=s.useState(!1),[w,y]=s.useState(""),f=s.useRef(null),j=s.useRef(null),N=s.useRef(null),g=s.useRef(null),[x,C]=$(v),[m,L]=s.useState({top:0,left:0,minWidth:0});s.useEffect(()=>{if(!t)return;const e=d=>{f.current?.contains(d.target)||j.current?.contains(d.target)||l(!1)};return document.addEventListener("mousedown",e),()=>document.removeEventListener("mousedown",e)},[t]);const R=s.useCallback(()=>{if(!f.current)return;const e=f.current.getBoundingClientRect();L({top:e.bottom+4,left:e.left,minWidth:e.width})},[]);s.useEffect(()=>{t&&(R(),y(""),requestAnimationFrame(()=>N.current?.focus()))},[t,R]);const h=c.find(e=>e.value===n),p=s.useMemo(()=>{const e=w.toLowerCase();return[...e?c.filter(o=>o.label.toLowerCase().includes(e)||o.value.toLowerCase().includes(e)):c].sort((o,b)=>{const k=x.has(o.value)?0:1,E=x.has(b.value)?0:1;return k!==E?k-E:o.label.localeCompare(b.label)})},[c,w,x]);s.useEffect(()=>{if(t&&n&&g.current){const e=g.current.querySelector(`[data-value="${CSS.escape(n)}"]`);e&&e.scrollIntoView({block:"nearest"})}},[t,n]);const W=t?P.createPortal(r.jsxs("div",{ref:j,className:"fixed z-[100] bg-white border border-[#d4d2cd] rounded shadow-lg",style:{top:m.top,left:m.left,minWidth:m.minWidth,width:a?"max-content":m.minWidth},children:[r.jsx("div",{className:"p-1.5 border-b border-[#e8e6e1]",children:r.jsx("input",{ref:N,type:"text",value:w,onChange:e=>y(e.target.value),placeholder:"Search...",className:"w-full px-2 py-1.5 text-xs text-[#2c2c2c] font-sans bg-[#f5f4f0] border border-[#e8e6e1] rounded focus:outline-none focus:border-[#2b5ea7] placeholder:text-[#9c9890]",onKeyDown:e=>{e.key==="Escape"&&l(!1),e.key==="Enter"&&p.length>0&&(i(p[0].value),l(!1))}})}),r.jsx("div",{ref:g,className:"max-h-[280px] overflow-y-auto",children:p.length===0?r.jsx("div",{className:"px-3 py-3 text-xs text-[#9c9890] text-center",children:"No projects found"}):p.map(e=>{const d=e.value===n,o=x.has(e.value);return r.jsxs("div",{"data-value":e.value,className:`flex items-center gap-1.5 px-2 py-1.5 text-xs font-sans cursor-pointer transition-colors ${d?"bg-[#2b5ea7]/10 text-[#2b5ea7]":"text-[#2c2c2c] hover:bg-[#f5f4f0]"}`,onClick:()=>{i(e.value),l(!1)},children:[r.jsx("button",{type:"button",onClick:b=>{b.stopPropagation(),C(e.value)},className:`shrink-0 w-4 h-4 flex items-center justify-center text-[10px] transition-colors ${o?"text-[#d4a017]":"text-[#d4d2cd] hover:text-[#9c9890]"}`,title:o?"Unstar":"Star to pin to top",children:o?"★":"☆"}),r.jsx("span",{className:a?"whitespace-nowrap":"truncate flex-1",children:e.label})]},e.value)})})]}),document.body):null;return r.jsxs(r.Fragment,{children:[r.jsxs("button",{ref:f,type:"button",onClick:()=>!u&&l(!t),disabled:u,className:"w-full bg-white border border-[#d4d2cd] rounded px-3 py-2 text-xs text-[#2c2c2c] font-sans text-left flex items-center justify-between focus:outline-none focus:border-[#2b5ea7] hover:border-[#b0ada6] transition-colors disabled:opacity-50 disabled:cursor-not-allowed",children:[r.jsx("span",{className:h?"text-[#2c2c2c] truncate":"text-[#9c9890]",children:h?h.displayLabel??h.label:S}),r.jsx("span",{className:"text-[#9c9890] text-[10px] ml-2 shrink-0 transition-transform duration-200",style:{transform:t?"rotate(180deg)":"rotate(0deg)"},children:"▾"})]}),W]})}export{I as S,P as r,$ as u};