fix: forward image and PDF attachments for DMR models

[Sayt-0]

Summary

DMR-hosted models are absent from the models.dev catalog, so modelinfo.LoadCaps
always missed for dmr/<model> IDs and image/PDF attachments were silently
dropped to a text-only strategy. This adds a per-model capability opt-in so DMR
vision/PDF models can forward attachments, and forwards PDFs as a native file
content part on the Chat Completions path.

Fixes #2739.

Issue expectations mapped to the implementation

Issue element	Implementation	Status
Vision/PDF attachments always degrade to text-only	Declared capabilities restore forwarding	done
Root cause: models.dev has no dmr provider, so the lookup always misses	DMR path no longer depends on the lookup; capabilities are injected explicitly	done
Possible fix: local allowlist of known DMR models	Not chosen (brittle, needs ongoing upkeep)	rejected
Possible fix: provider-level opt-in (capability hint in model config)	Chosen: provider_opts.supports_images / supports_pdf	done
Possible fix: query a DMR capability API	Not chosen (no stable capability API exposed; adds latency and a CLI dependency)	rejected

How it works

models:
  qwen-vl:
    provider: dmr
    model: ai/qwen2.5-vl
    provider_opts:
      supports_images: true
      supports_pdf: false

provider_opts is parsed and validated at client creation, stored as
Client.attachmentCaps, and injected via oaistream.ConvertMessagesWithCaps.
The models.dev lookup is skipped on the DMR path because it can never resolve a
dmr/<model> ID.

Changes

Area	Change
pkg/model/provider/dmr/configure.go	parseBoolOpt, plus parse and validate supports_images / supports_pdf
pkg/model/provider/dmr/client.go	Client.attachmentCaps set in NewClient, injected in convertMessages
pkg/model/provider/oaistream/messages.go	ConvertMessagesWithCaps; internals thread modelinfo.ModelCapabilities, resolving caps once per batch
pkg/model/provider/oaistream/attachments.go	Removed the unused store-variant convertDocument; PDFs forwarded as a file content part instead of dropped
agent-schema.json, examples/dmr.yaml	Documented the keys; added a qwen_vision example
tests	DMR parse/conversion/end-to-end and oaistream caps/PDF tests

Behavior

Case	Before	After
DMR vision model, image attachment	dropped (text-only)	forwarded when supports_images: true
DMR model, PDF attachment	dropped	forwarded as file part when supports_pdf: true
DMR model, no declared capability	dropped	dropped (unchanged, conservative default)
Catalogued providers (openai, anthropic, ...)	models.dev lookup	unchanged

Testing

• pkg/model/provider/dmr, pkg/model/provider/oaistream, pkg/modelinfo, pkg/attachment, pkg/model/provider/openai: pass.
• pkg/config example parsing (TestParseExamples) including the new dmr.yaml entry: pass.
• golangci-lint and the repository linter: clean.

Note for reviewers

The PDF forwarding change also affects the OpenAI legacy Chat Completions path:
a model that models.dev reports as PDF-capable now receives PDFs as a file
content part rather than having them dropped (newer OpenAI models already handle
PDFs via the Responses API). The behavior is gated by resolved capabilities, so
nothing changes for models without PDF support. If restricting this strictly to
DMR seems more appropriate, it can be changed.

[docker-agent]

❌ PR Review Failed — The review agent encountered an error and could not complete the review. View logs.

[docker-agent]

Assessment: 🟡 NEEDS ATTENTION

One MEDIUM severity finding in the new PDF attachment forwarding path. The fix for DMR attachment capabilities is well-structured overall — capabilities are parsed cleanly, threaded correctly through Client.attachmentCaps, and ConvertMessagesWithCaps resolves caps once per batch as intended. The image forwarding path is correct. The concern is in how PDF content is encoded for the Chat Completions file content part.

[docker-agent]

[MEDIUM] PDF file_data passes a data URI; OpenAI Chat Completions expects plain base64

The new PDF branch builds the file_data value with dataURI(...), producing a data:application/pdf;base64,<encoded> string:

openai.FileContentPart(openai.ChatCompletionContentPartFileFileParam{
    FileData: openai.String(dataURI(doc.MimeType, doc.Source.InlineData)),
    Filename: openai.String(doc.Name),
})

The OpenAI Chat Completions file content part spec expects file_data to be a plain base64-encoded string (no data: URI prefix), identical to how the Responses API input_file.file_data is specified. The data:mimetype;base64,... format is only valid for the url field of image_url content parts (used correctly in the image branch above).

Sending a data URI prefix in file_data will likely result in a 400/422 from any endpoint that follows the OpenAI Chat Completions spec strictly — including DMR backends that proxy to an OpenAI-compatible runtime. The code comment claims "Chat Completions accepts PDFs as a file content part carrying a base64 data URI", but this is not supported by the OpenAI API reference.

Suggested fix: strip the data URI prefix and pass only the raw base64 payload:

FileData: openai.String(base64.StdEncoding.EncodeToString(doc.Source.InlineData)),