Bound repeated generation input in long Codex rollouts

plugins/tracing/src/trace.ts

@@ -31,14 +31,17 @@ import type {
 } from "./types.js";
 import { debugLog, toText } from "./utils.js";
 
+const MAX_REPEATED_INPUT_BYTES = 64 * 1024 * 1024;
+
 async function loadSession(file: string): Promise<RolloutLine[]> {
   const lines: RolloutLine[] = [];
   const input = createReadStream(file, { encoding: "utf-8" });
   for await (const raw of createInterface({ input, crlfDelay: Infinity })) {
     const trimmed = raw.trim();
     if (!trimmed) continue;
     try {
-      lines.push(JSON.parse(trimmed) as RolloutLine);
+      const line = JSON.parse(trimmed) as RolloutLine;
+      if (line.type !== "compacted") lines.push(line);
     } catch {
       // skip malformed lines rather than aborting the whole upload
     }
@@ -435,6 +438,7 @@ async function emitTurn(
     unannouncedSubagents?: SubagentRollout[];
     inheritableTurnIds?: ReadonlySet<string>;
     historyPrefix?: ChatMlMessage[];
+    omitRepeatedInput: boolean;
   },
 ): Promise<void> {
   const isSubagent = sessionMeta.isSubagentThread === true || ctx.parentObservation != null;
@@ -492,10 +496,18 @@ async function emitTurn(
       const generation = startObservation(
         isSubagent ? "LLM Subagent" : "LLM",
         {
-          input: attachToolDefinitions(
-            generationInput(systemMessage, historyPrefix, turn, i),
-            turn.toolDefinitions,
-          ),
+          input: ctx.omitRepeatedInput
+            ? [
+                {
+                  role: "system",
+                  content:
+                    "[Generation input omitted from trace: estimated repeated context exceeds 64 MB]",
+                },
+              ]
+            : attachToolDefinitions(
+                generationInput(systemMessage, historyPrefix, turn, i),
+                turn.toolDefinitions,
+              ),
           output: buildGenerationOutput(step),
           model: turn.model,
           ...(turn.reasoningEffort
@@ -634,10 +646,15 @@ export async function convertRollout(
     stoppedTurnId?: string;
   },
 ): Promise<string[]> {
+  const rolloutBytes = (await fs.stat(rolloutFile)).size;
   const { sessionMeta, turns } = parseSession(await loadSession(rolloutFile));
+  // ponytail: file size overestimates prompt size; count parsed message bytes if this hides prompts too often.
+  const omitRepeatedInput =
+    rolloutBytes * turns.reduce((total, turn) => total + turn.steps.length, 0) >
+    MAX_REPEATED_INPUT_BYTES;
 
   const historyPrefixes: ChatMlMessage[][] = [];
-  {
+  if (!omitRepeatedInput) {
     const seen: ChatMlMessage[] = [];
     for (const turn of turns) {
       historyPrefixes.push([...seen]);
@@ -710,6 +727,7 @@ export async function convertRollout(
         unannouncedSubagents: subagentsFor(turnIndex),
         inheritableTurnIds,
         historyPrefix: historyPrefixes[turnIndex],
+        omitRepeatedInput,
       });
     }
     return [];
@@ -747,6 +765,7 @@ export async function convertRollout(
         unannouncedSubagents: subagentsFor(turnIndex),
         inheritableTurnIds,
         historyPrefix: historyPrefixes[turnIndex],
+        omitRepeatedInput,
       });
 
     if (attached) {

plugins/tracing/test/trace.test.ts

@@ -271,6 +271,62 @@ describe("convertRollout", () => {
     fs.rmSync(dir, { recursive: true, force: true });
   });
 
+  it("bounds repeated generation input while keeping a large one-step prompt", async () => {
+    const dir = fs.mkdtempSync(path.join(os.tmpdir(), "lf-codex-long-"));
+    const day = path.join(dir, "sessions", "2026", "06", "03");
+    fs.mkdirSync(day, { recursive: true });
+    const timestamp = "2026-06-03T09:00:00Z";
+    for (const steps of [1, 80]) {
+      const file = path.join(day, `rollout-${steps}.jsonl`);
+      const lines: unknown[] = [
+        { timestamp, type: "session_meta", payload: { id: `sess-${steps}` } },
+        { timestamp, type: "event_msg", payload: { type: "task_started", turn_id: "t1" } },
+        {
+          timestamp,
+          type: "event_msg",
+          payload: { type: "user_message", message: "x".repeat(1_000_000) },
+        },
+      ];
+      for (let step = 0; step < steps; step++)
+        lines.push(
+          {
+            timestamp,
+            type: "response_item",
+            payload: {
+              type: "message",
+              role: "assistant",
+              content: [{ type: "output_text", text: `Answer ${step}` }],
+            },
+          },
+          { timestamp, type: "event_msg", payload: { type: "token_count", info: {} } },
+        );
+      lines.push({
+        timestamp,
+        type: "event_msg",
+        payload: { type: "task_complete", turn_id: "t1" },
+      });
+      fs.writeFileSync(file, lines.map((line) => JSON.stringify(line)).join("\n"));
+      await convertAndMark(file, { config: baseConfig, stoppedTurnId: "t1" });
+      const generations = exporter
+        .getFinishedSpans()
+        .filter((span) => obsType(span) === "generation");
+      expect(generations).toHaveLength(steps);
+      const input = JSON.parse(attr(generations.at(-1)!, "langfuse.observation.input"));
+      if (steps === 1) expect(input[0].content).toHaveLength(1_000_000);
+      else
+        expect(input).toEqual([
+          {
+            role: "system",
+            content:
+              "[Generation input omitted from trace: estimated repeated context exceeds 64 MB]",
+          },
+        ]);
+      expect(attr(turnRoots()[0]!, "langfuse.observation.input")).toHaveLength(1_000_000);
+      exporter.reset();
+    }
+    fs.rmSync(dir, { recursive: true, force: true });
+  });
+
   it("sends attached images as multimodal content and the loaded tools on the call", async () => {
     const URI = "data:image/png;base64,iVBORw0KGgo=";
     const dir = fs.mkdtempSync(path.join(os.tmpdir(), "lf-codex-media-"));