SSE JSON parse failure with GLM-5.1 (Z.AI) — malformed JSON in content field kills stream

[darkhipo]

Description

When using GLM-5.1 via Z.AI's OpenAI-compatible API, the SSE stream occasionally contains invalid JSON. The model hallucinates SSE-formatted text (e.g. data: {"id":"...","choices":[...]}) as its actual response content. Z.AI's server does not properly escape the quotes in the content field, producing malformed JSON that fails parsing.

Observed Error

JSON parsing failed: Text: {"id":"20260420032348d6275404213948ac","created":1776626628,"object":"chat.completion.chunk","model":"glm-5.1","choices":[{
  "index":0,"delta":{"role":"assistant","content":"data: {"id":"20260420032457d424e96cb1da4f11","created":1776626698,"object":"chat.completion.chunk",
  "model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant","content":"Two"}}]}.
  Error message: JSON Parse error: Expected '}'

Root Cause

The SSE data: line is:

data: {"id":"...","choices":[{"delta":{"content":"data: {"id":"...","choices":[{"delta":{"content":"Two"}}]}}]}

The inner " in the hallucinated "data: {"id":"..." breaks the outer JSON string boundary. Z.AI's server should be escaping these quotes (\") but isn't.

Where It Fails

1. EventSourceParserStream (eventsource-parser) correctly extracts the line after data:
2. safeParseJSON() in @ai-sdk/provider-utils/src/parse-json-event-stream.ts calls JSON.parse() on the malformed string → throws
3. The error propagates up through processor.ts and kills the stream

Suggested Fix

Skip-and-continue: When a chunk fails JSON parsing, log a warning and continue to the next chunk instead of killing the entire stream. The hallucinated SSE chunks are typically noise — the real content arrives in adjacent valid chunks. A single malformed chunk shouldn't terminate the conversation.

// In the TransformStream inside parseJsonEventStream:
async transform({ data }, controller) {
  if (data === "[DONE]") return;
  try {
    controller.enqueue(await safeParseJSON({ text: data, schema }));
  } catch (err) {
    // Provider sent malformed JSON in this chunk — skip it
    // rather than killing the entire stream.
    console.warn("SSE chunk JSON parse failed, skipping:", data.slice(0, 200));
  }
}

Environment

• opencode v1.14.x
• Provider: Z.AI (api.z.ai/api/coding/paas/v4)
• Model: glm-5.1
• Stream mode: SSE

Reproduction

Any multi-turn conversation with GLM-5.1 via opencode will eventually trigger this. The model sporadically hallucinates SSE chunks in its response text. The failure is intermittent but frequent enough to disrupt normal usage.

Workaround

We added content sanitization in our downstream provider layer that strips data: {...} artifacts from response content, but the fix belongs in the SSE parser itself so all providers benefit.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• [BUG] GLM-5 via NVIDIA NIM emits malformed MCP tool JSON (Expected '}') in OpenCode v1.2.6 #13900: GLM-5 (Z.AI/NVIDIA NIM) emitting malformed JSON in tool calls with the same Expected '}' parse error
• [bug] GLM 5 keeps screwing up the json parsing of read tool calling #13982: GLM-5 screwing up JSON parsing in tool calling (same model family, same root behavior)
• [BUG] GLM-5 model pollutes context window with broken thinking output #16903: GLM-5 (Z.AI) polluting context with broken/malformed output — different symptom but same underlying model misbehavior
• Fair amount of this: Error message: JSON Parse error: Unterminated string #20786: Recurring JSON Parse error: Unterminated string — same class of SSE/stream JSON parse failure

If your issue is specifically about the stream being killed on a parse error (rather than the malformed JSON itself), it may still warrant its own report — but please review the above to avoid duplicate effort.

[darkhipo]

Reproduction from MultiAgency body session (2026-04-30)

Running opencode 1.4.10 (homebrew, darwin-arm64) with GLM-5.1 via z.ai. During a long coding session with heavy MCP tool usage, the SSE stream dies with:

JSON parsing failed: Text: {"id":"20260430204055297b97348c8e426c","created":1777552855,"object":"chat.cdata: {"id":"2026043020420635c6abc56a2741ac",
"created":1777552926,"object":"chat.completion.chunk","model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant","content":"11"}}]}.
Error message: JSON Parse error: Expected '}'

The malformed text shows two SSE data: frames concatenated — the first object is truncated mid-field ("object":"chat.c before "data:" from the next frame appears inline). This is the GLM-5 model generating truncated JSON in streaming chunks, not a network issue.

Environment:

• opencode 1.4.10 (homebrew)
• macOS darwin-arm64 (M1 Air)
• Model: zai-coding-plan/glm-5.1
• Session had 50+ tool calls before failure

The skip-and-continue approach would solve this — a single malformed chunk should not kill the entire conversation. The real content is in adjacent valid chunks.

[darkhipo]

Update: Filed with z.ai as well

We filed a detailed root-cause analysis on the z.ai side: zai-org/GLM-5#66

Two distinct failure modes identified

1. Truncated SSE chunks — the server emits a data: frame containing incomplete JSON (cut off mid-field, e.g. "object":"chat.c before the next data: appears inline). This is a server-side framing bug.

2. Hallucinated SSE content — the model generates data: {"id":"..."} as literal text in its response, and the server doesn't escape the inner quotes. This is a model output + server escaping issue.

Both produce the same symptom: JSON Parse error: Expected '}'.

Our environment (for the record)

• opencode 1.4.10 (homebrew, darwin-arm64)
• macOS M1 Air 8GB
• Model: zai-coding-plan/glm-5.1 via api.z.ai
• 50+ tool calls before failure
• Using MultiAgency body with 60+ MCP tools

The skip-and-continue approach suggested in the issue body would be a good client-side mitigation for both modes. But the root cause fix needs to happen server-side at z.ai.

[water-in-stone]

@darkhipo There is a general method to handle this problem: we could use jsonrepair

by the way, jsonrepair is a popular tool used for fixing malformed JSON text from LLMs. Its key features include:

• Automatically completing missing quotes, brackets, and commas
• Repairing unescaped double quotes (which directly addresses the notebook issue)
• Streaming support (jsonrepair/stream)
• Zero dependencies, written in native TypeScript