tool/read: full base64 data URI persisted inside tool-result message record (context bloat) #49320

Summary

The v2 read tool persists the entire base64 data URI of an image inside the tool-result record that gets stored in the session's message table — both as the type: "file" content part's uri field and again in the tool's output. Because message records are serialized back into the conversation (and clients render file parts as [file: data:image/jpeg;base64,...]), a single screenshot can inject ~100KB+ of base64 text into context — roughly the image size again — so image-heavy sessions blow past the context limit at ~2x the cost. v1 kept the binary in a separate attachments side-channel and only put a tiny "Image read successfully" string in the output.

Environment

• opencode version: v2.0.3 (verified against v2.0.4 source at v2.0.4 tag)
• OS: Darwin 25.6.0 (arm64)
• Terminal: Apple_Terminal, xterm-256color, truecolor
• Shell: /bin/zsh
• Install/channel: latest
• Active plugins: opencode-dynamic-context-pruning (local plugin), @dietrichgebert/ponytail, opencode-models-discovery@latest

Reproduction

1. In a v2 session, run the read tool on an image file (e.g., a ~300KB JPEG screenshot).
2. Inspect the persisted session message for the tool result: SQLite table session_message (column data, look at content[].state.content).
3. Observe { "type": "file", "uri": "data:image/jpeg;base64,<~400k chars>", "mime": ..., "name": ... } — the complete data URI is stored verbatim, and the tool's output field wraps the same full FileContent again.

Expected Behavior

Binary payloads should not be persisted inside the tool-result text record. The model should still receive the image as real media (image_url parts), but message records should keep a small reference — as v1 did with attachments — so persisted history stays small and clients can render a preview/thumbnail instead of a giant data URI.

Actual Behavior

The full base64 data URI is embedded in the stored message record and rendered verbatim back into agent context. Context/usage accounting then counts it again (model-request.ts boundImages() adds Buffer.byteLength(item.uri) for tool-result file image parts). Net effect: each image effectively costs ~2x in context/usage than the pixels sent to the model.

Additional Context

• DB evidence from a real session: one assistant message was 127KB; content[1].state.content[1].uri alone was 126,067 chars — the entire data:image/jpeg;base64,... string stored verbatim.
• v1 (v1.4.14, packages/opencode/src/tool/read.ts) returns { output: "Image read successfully", attachments: [{ type: "file", mime, url }] } — binary in a side-channel, never in the output text.
• v2 (packages/core/src/tool/plugin/read.ts) builds toModelContent with a file content part whose uri is the full data URI, and also returns that whole FileContent as the tool output.
• Provider adapters (packages/ai/src/protocols/openai-chat.ts lowerToolMessages/lowerMedia, also openai-responses/anthropic) do correctly lower file parts to image_url media parts — so the model request is fine; this is storage + transcript-rendering bloat, not a provider issue.
• Suggested fix areas: (1) keep binary payloads out of persisted tool-result records (reuse the v1 attachments pattern / store a file reference); (2) clients should render file parts without inlining data URIs into the transcript text; (3) revisit counting raw URI bytes in boundImages().
• Frequency: every image read; consistently reproducible.