OpenAI Responses compaction input contains a non-object item

Error message

OpenAI Responses compaction input contains a non-object item

What it means

When preparing OpenAI Responses input for compaction, every input item must be a plain object (Record<string, unknown>); the code validates each item after transforming the session history and throws if any element is not an object. This catches corrupted or wrongly-shaped history before it is sent to the API.

Source

Thrown at packages/agent/src/compaction/compaction.ts:1459

function buildOpenAiResponsesCompactionInput(
	messages: Message[],
	model: Model<"openai-responses" | "azure-openai-responses" | "openai-codex-responses">,
	previousReplacementHistory: Array<Record<string, unknown>> | undefined,
): Array<Record<string, unknown>> {
	const input = buildResponsesInput({
		model,
		context: { messages },
		strictResponsesPairing: model.compat.strictResponsesPairing,
		supportsImageDetailOriginal: openAiCompatSupportsImageDetailOriginal(model),
		nativeHistory: { replay: true, filterReasoning: false },
		includeThinkingSignatures: true,
		repairOrphanOutputs: true,
	});
	const nativeInput: Array<Record<string, unknown>> = [];
	for (const item of input) {
		if (!isRecord(item)) {
			throw new Error("OpenAI Responses compaction input contains a non-object item");
		}
		nativeInput.push(item);
	}
	return stripOpenAIResponsesOutputOnlyStatusesForReplay(
		previousReplacementHistory ? [...previousReplacementHistory, ...nativeInput] : nativeInput,
	);
}

/**
 * Resolve the Responses `reasoning` param for a V2 compaction request the same
 * way a normal turn does — through {@link resolveOpenAICompatPolicy}, so it
 * honors per-model effort support, `omitReasoningEffort`, disable modes, and the
 * wire-effort mapping. Returns `undefined` for non-reasoning models or when the
 * user selected `Off` (matching the normal-turn omission, not a fabricated shape).
 */
function buildCompactionV2Reasoning(
	model: Model<"openai-responses" | "azure-openai-responses" | "openai-codex-responses">,
	thinkingLevel: ThinkingLevel | undefined,

View on GitHub (pinned to 9690622007)

Solutions

Inspect the session history/messages feeding the compaction for non-object entries (null, string, number) and fix or remove them
Validate/normalize items before compaction — ensure each message passes the provider's input transform and remains an object
If a converter or plugin mutates input, fix it to emit record-shaped items matching the OpenAI Responses input schema
Re-export or rebuild the session from its source if the stored history is corrupted

Example fix

// before
input.push("summarize this") // raw string item
// after
input.push({ role: "user", content: [{ type: "input_text", text: "summarize this" }] })

Defensive patterns

Strategy: validation

Validate before calling

const bad = input.findIndex((item) => item === null || typeof item !== "object");
if (bad !== -1) {
  throw new Error(`input[${bad}] is not an object: ${JSON.stringify(input[bad])}`);
}

Type guard

function isRecord(v: unknown): v is Record<string, unknown> {
  return typeof v === "object" && v !== null && !Array.isArray(v);
}

Try / catch

try {
  await compact(session);
} catch (err) {
  if (err instanceof Error && err.message.includes("non-object item")) {
    sanitizeHistory(session); // drop/normalize malformed entries
    return compact(session);
  }
  throw err;
}

Prevention

Normalize every history item through the provider input transform before compaction
Reject/repair string or null message items at session load time
Add a record-shape type guard at the boundary where custom plugins inject messages

When it happens

Trigger: The serialized/transformed session input array contains a non-object element — e.g. a raw string, number, null, or undefined item introduced by a custom session format, a buggy converter/filter upstream, or hand-edited/corrupted session JSONL.

Common situations: Custom message transforms or plugins pushing plain strings into the input array; older session files with a different message schema merged into a new-format session; scripts that patch session history and insert raw text items.

Understand the failure class

Background: Schema validation failed / invalid input schema: payload rejected because its shape doesn't match the expected schema — this error's family across 28 libraries.

Related errors

AI-assisted analysis of can1357/oh-my-pi@9690622007 (2026-08-31). Data as JSON: /api/errors/83b8441df792bc3b. Report an issue: GitHub.