zai: PDF attachments always fail with 400 [1210] — opencode sends file_data, Z.AI accepts PDF only via file_url/file_id #49237

Summary

Reading any PDF with the read tool permanently fails a zai session. OpenCode maps application/pdf attachments to the OpenAI-compatible file.file_data (base64 data-URI) transport, which is the one PDF transport Z.AI's /paas/v4/chat/completions does not accept. The request is rejected with HTTP 400 code 1210 and the turn dies.

glm-5.3-flash reads PDFs fine — via file_url or file_id. Only file_data is rejected. So this is a transport-mapping defect, not a provider limitation.

Same class as #44486 (OpenAI Responses, input_file.file_data in tool-result replay), which was fixed.

Environment

• opencode version: 1.18.31 (Homebrew, /usr/local/Cellar/opencode/1.18.31)
• OS: macOS 15.7.9, x86_64, Darwin 24.6.0
• Terminal: WarpTerminal v0.2026.09.09.08.26.stable_02, shell /bin/zsh
• Provider / model: zai / glm-5.3-flash (variant high)
• Active plugins: none (~/.config/opencode/plugin does not exist; config is {"$schema": ...} only)
• Provider mapping (models.json): npm: @ai-sdk/openai-compatible, api: https://api.z.ai/api/paas/v4

Reproduction

1. opencode with zai/glm-5.3-flash.
2. Ask it to read any local PDF, so the read tool attaches it.
3. The next assistant turn fails:

level=ERROR message="stream error" providerID=zai modelID=glm-5.3-flash
  error.error="AI_APICallError: Failed to parse the file. Please check its accessibility and format."

The tool part itself succeeds (output: "PDF read successfully", attachments[0].mime = application/pdf). The failure happens when that attachment is hoisted into the synthetic user message and converted for the provider.

Root cause

MessageV2.toModelMessages hoists tool-result media into a synthetic user message. The H() allow-predicate only returns true for @ai-sdk/anthropic, @ai-sdk/openai, @ai-sdk/amazon-bedrock*, @ai-sdk/google-vertex/anthropic, @ai-sdk/xai (images) and @ai-sdk/google (gemini-3). For @ai-sdk/openai-compatible it returns false, so the PDF becomes a file part on a user message.

@ai-sdk/openai-compatible then hardcodes:

if (part.mediaType === "application/pdf") {
  return { type: "file", file: {
    filename: part.filename ?? "document.pdf",
    file_data: `data:application/pdf;base64,${convert(part.data)}`
  }}
}

Z.AI rejects exactly that shape for PDFs.

Evidence — transport matrix

Direct calls to https://api.z.ai/api/paas/v4/chat/completions, model: glm-5.3-flash, no opencode in the circuit:

Controls that rule out alternative explanations:

• Not the file. A 615-byte PDF I generated locally fails identically; the same bytes as image_url return 200.
• Not the account/key/network. Success and failure alternate on the same key, endpoint and TLS path within the same minute.
• Not thinking. Adding thinking:{type:"enabled"} with reasoning_effort low/high/max does not change the outcome. (Note: thinking:disabled does return 1210, with a different message — "This model always engages in thinking and cannot be disabled" — see [Bug] OpenCode Go intermittently returns 400 [1210] "thinking cannot be disabled" and MissingSessionID when routing omen-alpha requests #47872. 1210 is a generic "invalid parameter" bucket; the discriminator is the message field.)
• Not max_tokens/temperature. image_url + max_tokens:512 + temperature:0.6 returns 200.

What the spec says

Z.AI's OpenAPI for POST /paas/v4/chat/completions (VisionMultimodalContentItem → File) documents three members of file, and attaches the supported-format list only to file_url:

• file_url — "Only GLM-5.3-Flash, GLM-4.6V, GLM-4.5V supported. Supports formats such as pdf, txt, word, jsonl, xlsx, pptx"
• file_id — "The ID returned by the File Upload API, only GLM-5.3-Flash supported"
• file_data — "Base64 file content in the format data:<MIME>;base64,<BASE64_DATA>" — no format list

The docs' own "File Visual Example" uses file_url with a .pdf on glm-5.3-flash. So PDF-over-file_data was never promised; opencode's JSON is schema-conformant but picks the wrong transport.

Upload endpoint: POST /paas/v4/files, multipart, purpose enum user_data | agent. Only user_data produces a file_id the chat endpoint accepts for PDF.

Minimal repro

export ZAI_API_KEY=...   # https://z.ai/manage-apikey/apikey-list
python3 - <<'EOF'
import base64, json, os, urllib.request
KEY = os.environ["ZAI_API_KEY"]
# any valid PDF
pdf = base64.b64encode(open("small.pdf","rb").read()).decode()
body = {"model":"glm-5.3-flash","stream":False,"messages":[{"role":"user","content":[
    {"type":"file","file":{"filename":"document.pdf",
                           "file_data":"data:application/pdf;base64,"+pdf}},
    {"type":"text","text":"What is this?"}]}]}
req = urllib.request.Request("https://api.z.ai/api/paas/v4/chat/completions",
    data=json.dumps(body).encode(),
    headers={"Authorization":"Bearer "+KEY,"Content-Type":"application/json"})
try:
    print(urllib.request.urlopen(req, timeout=120).status)
except urllib.error.HTTPError as e:
    print(e.code, e.read().decode())
# -> 400 {"error":{"code":"1210","message":"Failed to parse the file. ..."}}
EOF

Swap the content part for {"type":"file","file":{"file_url":"https://cdn.bigmodel.cn/static/demo/demo1.pdf"}} to get a 200 against the same key and model.

Suggested fix

For zai (and any provider whose API is api.z.ai/api/paas/v4), don't emit file_data for application/pdf. Either:

1. Upload the attachment to POST /paas/v4/files with purpose=user_data and send {"type":"file","file":{"file_id": "<id>"}} — works for local files, which is the read-tool case; or
2. Send file_url when the attachment already has a reachable URL.

Option 1 needs a small cache keyed by content hash so a PDF is not re-uploaded on every step of the agent loop.

Worth noting for whoever picks this up: the file/image_url mutual exclusion in the same message — "Not support passing both the file and image_url or video_url parameters at the same time" — means a turn mixing a PDF and a screenshot will fail regardless of transport.

I'm happy to send a PR if the approach sounds right.