can1357/oh-my-pi · error

GPT-5 Harmony leak recurred after truncate-and-resume recove

Error message

GPT-5 Harmony leak recurred after truncate-and-resume recovery (${signalListLabel(err.detection.signals)}).

What it means

When the GPT-5 model leaks Harmony-format markup into its output, the agent loop attempts a truncate-and-resume recovery (replacing the leaked assistant message and resuming). This error is thrown after that recovery has succeeded twice already within the window (harmonyTruncateResumeCount >= 2): the leak keeps recurring, so the loop escalates instead of continuing to paper over it. The message lists the detection signals that identified the leak.

Source

Thrown at packages/agent/src/agent-loop.ts:1274

						signal,
						stream,
						telemetry,
						invokeAgentSpan,
						stepCounter,
						streamFn,
						harmonyRetryAttempt,
						hostToolChoice,
						softRequirementState.forcedToolChoice,
						preparedProviderCall,
					);
					harmonyRetryAttempt = 0;
					harmonyTruncateResumeCount = 0;
				} catch (err) {
					if (!(err instanceof HarmonyLeakInterruption)) throw err;
					if (err.recovered) {
						if (harmonyTruncateResumeCount >= 2) {
							await emitHarmonyAudit(config, err, "escalated", harmonyRetryAttempt);
							throw new Error(
								`GPT-5 Harmony leak recurred after truncate-and-resume recovery (${signalListLabel(err.detection.signals)}).`,
							);
						}
						harmonyTruncateResumeCount++;
						recovered = err.recovered;
						message = recovered.message;
						await emitHarmonyAudit(config, err, "truncate_resume", harmonyRetryAttempt);
						// A recovered message completes the turn, so the abort-retry counter
						// resets like the normal success path (the truncate-resume counter
						// keeps accumulating for its cross-turn cap).
						harmonyRetryAttempt = 0;
					} else {
						if (harmonyRetryAttempt >= 2) {
							await emitHarmonyAudit(config, err, "escalated", harmonyRetryAttempt);
							throw new Error(
								`GPT-5 Harmony leak persisted after ${harmonyRetryAttempt} retries (${signalListLabel(err.detection.signals)}).`,
							);
						}

View on GitHub (pinned to 9690622007)

Solutions

1. Switch to a different model or model snapshot that does not leak Harmony tokens.
2. Review/adjust the system prompt for content that induces the model to echo its chat template.
3. If this is a provider-side regression, report it; meanwhile pin a known-good model revision via the catalog config.
4. Temporarily raise the truncate-resume cap only if continued degraded output is preferable to failing the turn.

Example fix

// before
const agent = createAgent({ model: "gpt-5-preview-0902" });
// after
const agent = createAgent({ model: "gpt-5" }); // known-good revision; or catch and fall back:
try {
  await runAgent(agent);
} catch (err) {
  if (String(err).includes("Harmony leak recurred")) await runAgent(fallbackAgent);
  else throw err;
}

function isHarmonyLeakEscalation(err: unknown): err is Error {
  return err instanceof Error && err.message.includes("Harmony leak recurred after truncate-and-resume");
}

try {
  await runTurn();
} catch (err) {
  if (isHarmonyLeakEscalation(err)) {
    await runTurn({ model: fallbackModel }); // retry with a non-leaking model
  } else {
    throw err;
  }
}

When it happens

Trigger: Streaming an assistant response from a GPT-5-class model where HarmonyLeakInterruption with recovered=true is raised on the third occurrence within the cap — i.e. the model repeatedly emits Harmony tokens (<|start|>, channel markers, etc.) even after two truncate-and-resume recoveries.

Common situations: A specific model snapshot/revision regressed into emitting its internal chat template; aggressive system prompts or few-shot content that triggers template echo; misrouted model IDs hitting a Harmony-trained model without the provider-side template stripping.

Related errors

• GPT-5 Harmony leak persisted after ${harmonyRetryAttempt} re
• inspect_image model returned no text output.

AI-assisted analysis of can1357/oh-my-pi@9690622007 (2026-08-31). Data as JSON: /api/errors/e3ea3a10a0f14e14. Report an issue: GitHub.