fix: AI Conversations input/output correlation (0.2.11 + 0.2.12)

.claude-plugin/marketplace.json

@@ -8,7 +8,7 @@
     {
       "name": "claude-code-ai-observability",
       "description": "Realtime Sentry traces with per-turn tokens, cost, subagent and error instrumentation, and rich auto-tagging (session / git / host). Forked from sergical/claude-code-sentry-monitor (MIT).",
-      "version": "0.2.11",
+      "version": "0.2.12",
       "author": {
         "name": "Joshkop"
       },

.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "claude-code-ai-observability",
   "description": "Comprehensive AI Agent Observability plugin for Claude Code — realtime Sentry traces with per-turn tokens, cost, subagent and error instrumentation, and rich auto-tagging (session / git / host). Forked from sergical/claude-code-sentry-monitor (MIT).",
-  "version": "0.2.11",
+  "version": "0.2.12",
   "author": {
     "name": "Joshkop"
   },

CHANGELOG.md

@@ -4,6 +4,12 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
+## [0.2.12] - 2026-05-21
+
+### Fixed
+
+- **AI Conversations: inputs and outputs now align with the right turn.** Claude Code's transcript records client-only slash commands like `/model` and `/clear` as a synthetic user-line triple (`<local-command-caveat>`, `<command-name>`, `<local-command-stdout>`), and Claude Code does not fire `UserPromptSubmit` for them. The transcript reader was counting every one of those lines as a real turn, so the collector's `turnIndex` (which only counts real `UserPromptSubmit` events) drifted off-by-N from the transcript's real-turn ordinal. Because real `UserPromptSubmit` hook payloads don't carry `prompt_id`, `selectTurn` fell back to ordinal — and the ordinal pointed at the wrong transcript turn. Visible symptom: early turns showed `gen_ai.input.messages` but the chat span had no `gen_ai.output.messages` (`claude_code.usage_extraction.status = turn_had_no_usage`), and a later turn carried an *earlier* turn's response (e.g. turn 3's chat span output was actually the assistant response to turn 0). `readTranscript` now skips `isMeta:true` lines and every user line belonging to a prompt_id that has any `<local-command-*>` content, restoring 1:1 alignment with `UserPromptSubmit`. Defense-in-depth: `UserPromptSubmit` events are now read from both `prompt_id` and `promptId`.
+
 ## [0.2.11] - 2026-05-21
 
 ### Fixed

package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-code-ai-observability",
-  "version": "0.2.11",
+  "version": "0.2.12",
   "description": "Comprehensive AI Agent Observability plugin for Claude Code — realtime Sentry traces with per-turn tokens, cost, subagent and error instrumentation, and rich auto-tagging (session / git / host).",
   "type": "module",
   "scripts": {

scripts/server.js

@@ -311,7 +311,7 @@ export function startServer(sentry, config, baseAutoTags) {
         const record = getOrCreateSession(event);
         void closeCurrentTurn(record).then(() => {
             record.turnIndex += 1;
-            record.currentPromptId = event.prompt_id ?? null;
+            record.currentPromptId = event.prompt_id ?? event.promptId ?? null;
             const prompt = event.prompt ?? event.message ?? null;
             record.currentTurnPrompt = prompt;
             record.currentTurnStart = Date.now() / 1000;

scripts/transcript-reader.js

@@ -38,6 +38,21 @@ function isToolResultUserLine(content) {
 function promptIdOf(l) {
     return l.promptId ?? l.prompt_id ?? null;
 }
+function isMetaOf(l) {
+    return l.isMeta === true || l.is_meta === true;
+}
+// Client-only slash commands like /model and /clear are surfaced in the
+// transcript as a synthetic user-line triple (<local-command-caveat>,
+// <command-name>, <local-command-stdout>) but never fire UserPromptSubmit,
+// so they must NOT count as real turns. Detect them by the unmistakable
+// <local-command-*> wrappers — these only appear on client-side commands;
+// model-bound slash commands (e.g. /superpowers:foo) have <command-name>
+// without any <local-command-*> sibling.
+function isLocalCommandText(text) {
+    if (!text)
+        return false;
+    return text.startsWith("<local-command-stdout>") || text.startsWith("<local-command-caveat>");
+}
 function collectSessionDims(l, into) {
     if (into.permissionMode === undefined) {
         into.permissionMode = l.permissionMode ?? l.permission_mode;
@@ -91,6 +106,24 @@ export function readTranscript(path) {
     if (recognizedTypeLines === 0) {
         return legacyResult(path, "no recognizable transcript line types");
     }
+    // Pre-pass: any prompt_id that has at least one <local-command-*> line is
+    // a client-only slash command (/model, /clear, …). Every user line under
+    // that prompt_id must be excluded from real-turn segmentation; otherwise
+    // the synthetic triple bumps the real-turn ordinal off-by-N from the
+    // collector's turnIndex (which only counts true UserPromptSubmit events).
+    const clientOnlyPromptIds = new Set();
+    for (const l of parsed) {
+        if (l.type !== "user")
+            continue;
+        if (l.isSidechain === true)
+            continue;
+        const pid = promptIdOf(l);
+        if (!pid)
+            continue;
+        if (isLocalCommandText(textFromContent(l.message?.content))) {
+            clientOnlyPromptIds.add(pid);
+        }
+    }
     const turns = [];
     const byPromptId = new Map();
     const session = {};
@@ -104,10 +137,21 @@ export function readTranscript(path) {
                 continue;
             if (isToolResultUserLine(l.message?.content))
                 continue;
+            // Skip caveat / meta lines — these are Claude Code annotations, not
+            // user input, and never fire UserPromptSubmit.
+            if (isMetaOf(l))
+                continue;
+            const pid = promptIdOf(l);
+            // Skip every line in a client-only slash-command group (see pre-pass).
+            if (pid && clientOnlyPromptIds.has(pid))
+                continue;
+            // Defensive: also skip a bare <local-command-*> line that somehow lacks
+            // a prompt_id (older Claude Code builds).
+            if (isLocalCommandText(textFromContent(l.message?.content)))
+                continue;
             if (current)
                 turns.push(current);
             realIndex += 1;
-            const pid = promptIdOf(l);
             current = emptyTurn(pid, realIndex);
             const t = textFromContent(l.message?.content);
             if (t)

src/server.ts

@@ -371,7 +371,7 @@ export function startServer(
     const record = getOrCreateSession(event);
     void closeCurrentTurn(record).then(() => {
       record.turnIndex += 1;
-      record.currentPromptId = event.prompt_id ?? null;
+      record.currentPromptId = event.prompt_id ?? event.promptId ?? null;
       const prompt = event.prompt ?? event.message ?? null;
       record.currentTurnPrompt = prompt;
       record.currentTurnStart = Date.now() / 1000;

src/transcript-reader.ts

@@ -29,6 +29,8 @@ export interface TranscriptReadResult {
 interface Line {
   type?: string;
   isSidechain?: boolean;
+  isMeta?: boolean;
+  is_meta?: boolean;
   promptId?: string;
   prompt_id?: string;
   permissionMode?: string;
@@ -85,6 +87,22 @@ function promptIdOf(l: Line): string | null {
   return l.promptId ?? l.prompt_id ?? null;
 }
 
+function isMetaOf(l: Line): boolean {
+  return l.isMeta === true || l.is_meta === true;
+}
+
+// Client-only slash commands like /model and /clear are surfaced in the
+// transcript as a synthetic user-line triple (<local-command-caveat>,
+// <command-name>, <local-command-stdout>) but never fire UserPromptSubmit,
+// so they must NOT count as real turns. Detect them by the unmistakable
+// <local-command-*> wrappers — these only appear on client-side commands;
+// model-bound slash commands (e.g. /superpowers:foo) have <command-name>
+// without any <local-command-*> sibling.
+function isLocalCommandText(text: string | null): boolean {
+  if (!text) return false;
+  return text.startsWith("<local-command-stdout>") || text.startsWith("<local-command-caveat>");
+}
+
 function collectSessionDims(l: Line, into: SessionDimensions): void {
   if (into.permissionMode === undefined) {
     into.permissionMode = l.permissionMode ?? l.permission_mode;
@@ -139,6 +157,22 @@ export function readTranscript(path: string): TranscriptReadResult {
     return legacyResult(path, "no recognizable transcript line types");
   }
 
+  // Pre-pass: any prompt_id that has at least one <local-command-*> line is
+  // a client-only slash command (/model, /clear, …). Every user line under
+  // that prompt_id must be excluded from real-turn segmentation; otherwise
+  // the synthetic triple bumps the real-turn ordinal off-by-N from the
+  // collector's turnIndex (which only counts true UserPromptSubmit events).
+  const clientOnlyPromptIds = new Set<string>();
+  for (const l of parsed) {
+    if (l.type !== "user") continue;
+    if (l.isSidechain === true) continue;
+    const pid = promptIdOf(l);
+    if (!pid) continue;
+    if (isLocalCommandText(textFromContent(l.message?.content))) {
+      clientOnlyPromptIds.add(pid);
+    }
+  }
+
   const turns: RealTurn[] = [];
   const byPromptId = new Map<string, RealTurn>();
   const session: SessionDimensions = {};
@@ -151,9 +185,17 @@ export function readTranscript(path: string): TranscriptReadResult {
     if (l.type === "user") {
       if (l.isSidechain === true) continue;
       if (isToolResultUserLine(l.message?.content)) continue;
+      // Skip caveat / meta lines — these are Claude Code annotations, not
+      // user input, and never fire UserPromptSubmit.
+      if (isMetaOf(l)) continue;
+      const pid = promptIdOf(l);
+      // Skip every line in a client-only slash-command group (see pre-pass).
+      if (pid && clientOnlyPromptIds.has(pid)) continue;
+      // Defensive: also skip a bare <local-command-*> line that somehow lacks
+      // a prompt_id (older Claude Code builds).
+      if (isLocalCommandText(textFromContent(l.message?.content))) continue;
       if (current) turns.push(current);
       realIndex += 1;
-      const pid = promptIdOf(l);
       current = emptyTurn(pid, realIndex);
       const t = textFromContent(l.message?.content);
       if (t) current.prompt = t;

src/types.ts

@@ -117,8 +117,10 @@ export interface UserPromptSubmitEvent extends AiobsEnvelope {
   message?: string;
   /** C1: stable id of this prompt, used to correlate the collector's open
    *  turn to the transcript's real-turn line. Optional — absent on older
-   *  Claude Code; collector falls back to ordinal among real turns. */
+   *  Claude Code; collector falls back to ordinal among real turns. Some
+   *  builds emit camelCase `promptId`, so both are accepted at read time. */
   prompt_id?: string;
+  promptId?: string;
 }
 
 export interface PreToolUseEvent extends AiobsEnvelope {

tests/transcript-reader.test.ts

@@ -56,6 +56,44 @@ describe("readTranscript — turn segmentation (C1)", () => {
   });
 });
 
+describe("readTranscript — synthetic slash-command lines do not count as real turns", () => {
+  // Claude Code emits synthetic user lines for client-only slash commands like
+  // /model and /clear — typically a triple: <local-command-caveat> (isMeta:true),
+  // <command-name>/foo, <local-command-stdout>. These never fire UserPromptSubmit,
+  // so they must be excluded from the real-turn index, or the collector's
+  // turnIndex (which counts UserPromptSubmit events) drifts off-by-N from
+  // transcript ordinals and selectTurn returns the wrong turn.
+  it("skips /model-style synthetic triples and aligns ordinals with real prompts", () => {
+    const p = make(
+      [
+        // /model client-only command: 3 synthetic user lines sharing one promptId
+        JSON.stringify({ type: "user", isMeta: true, promptId: "cmd1", message: { content: "<local-command-caveat>Caveat: ...</local-command-caveat>" } }),
+        JSON.stringify({ type: "user", promptId: "cmd1", message: { content: "<command-name>/model</command-name>\n<command-args></command-args>" } }),
+        JSON.stringify({ type: "user", promptId: "cmd1", message: { content: "<local-command-stdout>Set model to Sonnet</local-command-stdout>" } }),
+        // First real prompt — UserPromptSubmit would fire here, collector turnIndex=0
+        JSON.stringify({ type: "user", promptId: "real1", message: { content: "Hello how are you" } }),
+        JSON.stringify({ type: "assistant", message: { model: "m", usage: { input_tokens: 100, output_tokens: 10 }, content: [{ type: "text", text: "Doing well, thanks!" }] } }),
+        // Second real prompt — collector turnIndex=1
+        JSON.stringify({ type: "user", promptId: "real2", message: { content: "another reply please" } }),
+        JSON.stringify({ type: "assistant", message: { model: "m", usage: { input_tokens: 200, output_tokens: 20 }, content: [{ type: "text", text: "Here you go." }] } }),
+      ].join("\n"),
+    );
+    const r = readTranscript(p);
+    expect(r.degraded).toBe(false);
+    // Only 2 REAL turns — the 3 /model synthetic lines collapse to 0.
+    expect(r.turns).toHaveLength(2);
+    // Ordinal 0 must be "Hello", not the caveat — this is what selectTurn
+    // returns when UserPromptSubmit lacks prompt_id and falls back to ordinal.
+    expect(r.turns[0].promptId).toBe("real1");
+    expect(r.turns[0].response).toBe("Doing well, thanks!");
+    expect(r.turns[1].promptId).toBe("real2");
+    expect(r.turns[1].response).toBe("Here you go.");
+    // byPromptId must NOT contain the synthetic prompt_id at all.
+    expect(r.byPromptId.has("cmd1")).toBe(false);
+    expect(selectTurn(r, undefined, 0).turn!.response).toBe("Doing well, thanks!");
+  });
+});
+
 describe("readTranscript — missing/empty file is not parse-degraded", () => {
   it("missing file → empty turns, degraded false (not format drift)", () => {
     const r = readTranscript("/tmp/__aiobs_no_such_transcript__.jsonl");