fix(tracing): trace waterfall + summary prompt + chat tab all lose data after each agent turn

packages/opencode/src/altimate/observability/tracing.ts

@@ -42,7 +42,7 @@ export interface TraceSpan {
   spanId: string
   parentSpanId: string | null
   name: string
-  kind: "session" | "generation" | "tool" | "text" | "span"
+  kind: "session" | "generation" | "tool" | "text" | "span" | "user-message"
   startTime: number
   endTime?: number
   status: "ok" | "error"
@@ -345,6 +345,12 @@ function formatDurationShort(ms: number): string {
 }
 // altimate_change end
 
+// altimate_change start — shared truncation cap for `logUserMessage` span input.
+// Exported so the viewer's chat-tab dedupe can compare against the same boundary
+// (otherwise it'd silently drift if either side changes the magic number).
+export const USER_MESSAGE_INPUT_MAX_CHARS = 4000
+// altimate_change end
+
 export class Trace {
   // Global active trace — set when a session starts, cleared on end.
   private static _active: Trace | null = null
@@ -490,6 +496,109 @@ export class Trace {
     this.snapshot()
   }
 
+  // altimate_change start — rehydrate a Trace from an existing on-disk file.
+  // Used by `getOrCreateTrace` on cache miss for a session whose trace file
+  // already exists (worker restart, MAX_TRACES eviction). Without this, the
+  // fresh Trace.create() + startTrace() path would push a single root span
+  // into an empty `this.spans` and the immediate snapshot would clobber the
+  // rich on-disk trace. Returns true if a usable trace was loaded; false
+  // otherwise (the caller should fall back to startTrace).
+  //
+  // Does NOT call snapshot — the file is already correct on disk; an
+  // unnecessary write here would compete with concurrent flushSync paths.
+  //
+  // Async: read the file off the event-loop hot path so a worker-restart
+  // cache miss with a large existing trace doesn't head-of-line-block all
+  // session events. The actual hot path was bounded (cache-miss only) but
+  // the sync `readFileSync` could still pause the worker noticeably for a
+  // multi-MB trace.
+  async rehydrateFromFile(sessionId: string): Promise<boolean> {
+    if (!this.snapshotDir) return false
+    const safeId = sessionId.replace(/[/\\.:]/g, "_") || "unknown"
+    const filePath = path.join(this.snapshotDir, `${safeId}.json`)
+    let raw: string
+    try {
+      raw = await fs.readFile(filePath, "utf-8")
+    } catch {
+      return false
+    }
+    let trace: TraceFile
+    try {
+      trace = JSON.parse(raw) as TraceFile
+    } catch {
+      return false
+    }
+    // `buildTraceFile` writes the sanitized form of `sessionId` (see ~line 808
+    // `.replace(/[/\\.:]/g, "_")`), so compare the same way — otherwise valid
+    // trace files with `/`, `\`, `.`, `:` in the session id would be rejected
+    // and the caller would fall back to `startTrace`, clobbering them.
+    const normalizedSessionId = sessionId.replace(/[/\\.:]/g, "_") || "unknown"
+    if (
+      !trace ||
+      trace.sessionId !== normalizedSessionId ||
+      !Array.isArray(trace.spans) ||
+      trace.spans.length === 0
+    ) {
+      return false
+    }
+    const root = trace.spans.find((s) => s.parentSpanId === null && s.kind === "session")
+    if (!root) return false
+
+    this.sessionId = sessionId
+    // Restore the original traceId so post-rehydrate snapshots/exports keep
+    // the same trace identity (downstream OTLP/Jaeger-style consumers and the
+    // trace viewer URL both depend on it being stable across rehydration).
+    if (typeof trace.traceId === "string" && trace.traceId.length > 0) {
+      this.traceId = trace.traceId
+    }
+    this.spans = trace.spans.map((s) => ({ ...s }))
+    this.rootSpanId = root.spanId
+    this.metadata = { ...(trace.metadata ?? {}) }
+    this.startTime = root.startTime
+    if (trace.summary) {
+      this.totalTokens = trace.summary.totalTokens ?? 0
+      this.totalCost = trace.summary.totalCost ?? 0
+      this.toolCallCount = trace.summary.totalToolCalls ?? 0
+      this.generationCount = trace.summary.totalGenerations ?? 0
+      if (trace.summary.tokens) {
+        this.tokensBreakdown = {
+          input: trace.summary.tokens.input ?? 0,
+          output: trace.summary.tokens.output ?? 0,
+          reasoning: trace.summary.tokens.reasoning ?? 0,
+          cacheRead: trace.summary.tokens.cacheRead ?? 0,
+          cacheWrite: trace.summary.tokens.cacheWrite ?? 0,
+        }
+      }
+    }
+    // Mid-session rehydration: the root span's endTime (if any) was set by a
+    // prior endTrace; clear it so the trace doesn't render as "completed."
+    const r = this.spans.find((s) => s.spanId === this.rootSpanId)
+    if (r) {
+      delete (r as { endTime?: number }).endTime
+      r.status = "ok"
+    }
+    // Close any in-flight generation spans as interrupted. The transient state
+    // these spans depended on (`this.currentGenerationSpanId`, `generationText`,
+    // `generationToolCalls`, `pendingToolResults`) lives only in memory and
+    // can't be reconstructed from the on-disk file. If we leave open spans
+    // open, the next `step-finish` event for that turn drops at the
+    // `if (!this.currentGenerationSpanId) return` guard in `logStepFinish`
+    // and any later `logToolCall` falls back to attaching tool spans to
+    // `rootSpanId`, silently degrading the trace shape. Marking them
+    // interrupted preserves the partial data and makes the boundary visible.
+    const now = Date.now()
+    for (const s of this.spans) {
+      if (s.kind === "generation" && s.endTime === undefined) {
+        s.endTime = now
+        s.status = "error"
+        s.statusMessage = "interrupted (worker restart / cache eviction before step-finish)"
+      }
+    }
+    this.endTraceStarted = false
+    return true
+  }
+  // altimate_change end
+
   /**
    * Enrich the trace with model/provider info from the first assistant message.
    * Called when the message.updated event fires with assistant role.
@@ -515,6 +624,46 @@ export class Trace {
     if (prompt) this.metadata.prompt = prompt
   }
 
+  // altimate_change start — set only the user prompt without mutating title.
+  // The bus emits an auto-generated session title (Path C, `session.updated`)
+  // separately from the user's actual prompt text (Path B, `message.part.updated`).
+  // Capturing the prompt via `setTitle(text, text)` would race the title-agent:
+  // if the user text part arrived after `session.updated`, the nice generated
+  // title ("Greeting") would regress to the raw user input ("hi"). Use this
+  // method for prompt-only capture.
+  setPrompt(prompt: string) {
+    if (prompt) this.metadata.prompt = prompt
+  }
+  // altimate_change end
+
+  // altimate_change start — record an individual user message as a span so the
+  // chat tab can render multi-turn conversations. Without this, the viewer's
+  // chat tab can only display `metadata.prompt` at the top — a single string —
+  // and every later user message is silently dropped from the conversation
+  // rendering. Tracked as `kind: "user-message"` so the viewer can interleave
+  // these with `kind: "generation"` spans by startTime.
+  logUserMessage(text: string) {
+    if (!this.rootSpanId) return
+    if (!text) return
+    try {
+      const now = Date.now()
+      this.spans.push({
+        spanId: randomUUIDv7(),
+        parentSpanId: this.rootSpanId,
+        name: "user-message",
+        kind: "user-message",
+        startTime: now,
+        endTime: now,
+        status: "ok",
+        input: text.length > USER_MESSAGE_INPUT_MAX_CHARS ? text.slice(0, USER_MESSAGE_INPUT_MAX_CHARS) : text,
+      })
+      this.snapshot()
+    } catch {
+      // best-effort
+    }
+  }
+  // altimate_change end
+
   /**
    * Open a generation span from a step-start event.
    */

packages/opencode/src/altimate/observability/viewer.ts

@@ -12,7 +12,7 @@
  * Branded with Altimate Trace colors. Includes share/export features for virality.
  */
 
-import type { TraceFile } from "./tracing"
+import { USER_MESSAGE_INPUT_MAX_CHARS, type TraceFile } from "./tracing"
 
 export function renderTraceViewer(trace: TraceFile, options?: { live?: boolean; apiPath?: string }): string {
   const traceJSON = JSON.stringify(trace).replace(/<\//g, "<\\/")
@@ -1310,13 +1310,48 @@ function showDetail(span) {
 (function() {
   var el = document.getElementById('v-chat');
   var html = '';
+  // Build a chronologically sorted list of conversation turns by interleaving
+  // user-message spans with generation spans. We render \`metadata.prompt\` at
+  // the top as a fallback whenever it carries a value not already represented
+  // by a user-message span — that covers three cases:
+  //   1. Pre-fix traces with no user-message spans (just the legacy prompt).
+  //   2. Mixed traces: an older session rehydrated with only metadata.prompt,
+  //      then continued under the new code which added user-message spans for
+  //      later turns. Without this, the legacy first turn would be dropped.
+  //   3. Brand-new traces where the first user-message span input equals
+  //      metadata.prompt — we skip the duplicate to avoid double-rendering.
+  var userMsgs = spans.filter(function(s){return s.kind==='user-message';});
+  var gens = spans.filter(function(s){return s.kind==='generation';});
   if (t.metadata.prompt) {
-    html += '<div class="chat-msg user"><div class="chat-role">\\u25B6 You</div>';
-    html += '<div class="chat-bubble">' + e(t.metadata.prompt) + '</div></div>';
+    // \`logUserMessage\` truncates span input to USER_MESSAGE_INPUT_MAX_CHARS
+    // (see tracing.ts); \`metadata.prompt\` stores the full string. For prompts
+    // longer than the truncation length, strict equality would miss the dedupe
+    // and the same text would render twice. Match against the truncated form as
+    // well. The constant is interpolated at render time so the two sides can't
+    // drift.
+    var USER_MSG_TRUNCATE = ${USER_MESSAGE_INPUT_MAX_CHARS};
+    var promptStr = String(t.metadata.prompt);
+    var promptTruncated = promptStr.slice(0, USER_MSG_TRUNCATE);
+    var promptAlreadyInSpan = userMsgs.some(function(u){
+      return typeof u.input === 'string' && (u.input === promptStr || u.input === promptTruncated);
+    });
+    if (!promptAlreadyInSpan) {
+      html += '<div class="chat-msg user"><div class="chat-role">\\u25B6 You</div>';
+      html += '<div class="chat-bubble">' + e(promptStr) + '</div></div>';
+    }
   }
-  var gens = spans.filter(function(s){return s.kind==='generation';});
-  gens.forEach(function(gen) {
-    var tools = spans.filter(function(s){return s.parentSpanId===gen.spanId && s.kind==='tool';});
+  var turns = userMsgs.concat(gens).sort(function(a, b) { return (a.startTime||0) - (b.startTime||0); });
+  turns.forEach(function(s) {
+    if (s.kind === 'user-message') {
+      var utxt = typeof s.input === 'string' ? s.input : (s.input != null ? JSON.stringify(s.input) : '');
+      if (!utxt) return;
+      html += '<div class="chat-msg user"><div class="chat-role">\\u25B6 You</div>';
+      html += '<div class="chat-bubble">' + e(utxt) + '</div></div>';
+      return;
+    }
+    // Generation: render its tool children first, then the agent response.
+    var gen = s;
+    var tools = spans.filter(function(c){return c.parentSpanId===gen.spanId && c.kind==='tool';});
     if (tools.length) {
       tools.forEach(function(tool) {
         html += '<div class="chat-tool' + (tool.status === 'error' ? ' err' : '') + '">';

packages/opencode/src/cli/cmd/tui/routes/session/index.tsx

@@ -183,11 +183,24 @@ export function Session() {
     return new CustomSpeedScroll(3)
   })
 
-  createEffect(() => {
-    if (session()?.workspaceID) {
-      sdk.setWorkspace(session()?.workspaceID)
-    }
-  })
+  // altimate_change start — gate setWorkspace on actual workspaceID change.
+  // A plain inline `createEffect` callback that reads the session signal would
+  // re-fire whenever ANY field on the signal changes (message count, status,
+  // parts) — including the cascade of updates at agent-finish. Every spurious
+  // fire propagates into `worker.setWorkspace` → `startEventStream` →
+  // `sessionTraces.clear()` → next snapshot overwrites the rich on-disk trace
+  // with a near-empty one. The `on()` projector below restricts SolidJS dirty-
+  // tracking to the workspaceID value alone, so the effect only fires when
+  // that field actually changes.
+  createEffect(
+    on(
+      () => session()?.workspaceID,
+      (workspaceID) => {
+        if (workspaceID) sdk.setWorkspace(workspaceID)
+      },
+    ),
+  )
+  // altimate_change end
 
   createEffect(async () => {
     await sync.session