PDF parser FaaS call failed on large context, no single PDF at fault

Description

Description:
When the session context grows, the provider fails with:
"Upstream request failed: [invalid_request_error] failed to parse PDF at input[XXX].output[1]: PDF parser FaaS call failed"
The input index varies between runs (observed 197, 201, 205, 214).

Reproduction:

1. Tested 3 PDF files individually with Muse Spark 1.3 via OpenCode Zen: all 3 load fine alone (5.5 MB / 57 pages v1.7, 4.0 MB / 8 pages v1.4, 2.9 MB / 8 pages v1.4).
2. As the session context grows past ~200 inputs, the same error appears even without loading a new PDF (also seen while fetching a web page).
3. So no single PDF is at fault. The FaaS ingestion stage breaks on large contexts.

Attachments:
No file attached on purpose. The slides are copyrighted training material (reproduction prohibited), and since every file loads fine on its own, an attachment would not help reproduce. I can share the largest file (5.5 MB, 57 pages) privately on request.

Environment: Windows, OpenCode Desktop, model opencode/muse-spark-1.3-contributor-free.

Expected: files that load fine individually keep working in a larger session, or a clear error naming the offending input.

Actual: generic FaaS crash with a shifting input index.

Workaround used: convert PDFs locally with pdftotext and load only the .md text.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

No repository file or test is named. Start by reproducing a session with more than 200 inputs, compare PDF loading alone with loading in the growing context, and inspect the provider PDF-ingestion request path. Done means large-context PDFs continue working or the error clearly identifies the offending input.