Attached images degrade to a bare text placeholder when the model lacks vision — agent has no way to read them #42758

Activity

1. github-actions

   assigned

   neriousy

   on Aug 15, 2026
2. github-actions commented on Aug 15, 2026

   

   github-actionsBot

   on Aug 15, 2026 – with GitHub Actions

   Contributor

   More actions

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

   • [FEATURE]: Allow plugins to customize behavior when a model doesn't support a media modality (image/audio/video) #36006: Same root problem — when a model doesn't support a media modality, the agent has no actionable path. That issue also proposes plugin hooks and richer replacement text (including file paths) so agents can recover. The proposals in both issues overlap substantially (exposing retrievable references / file paths, configurable replacement text, and hooks after the strip step).
   • [FEATURE]: MCP tool calls should carry current session context (session_id) #33035: Partially related — proposes exposing session_id to MCP tools/agents (one of the workarounds your issue describes needing to find images via the serve API).

   If your issue adds meaningful new detail or a distinct angle not covered by those, feel free to clarify the difference so maintainers can decide whether to merge them.
3. sumwcc commented on Aug 16, 2026

   

   sumwcc

   on Aug 16, 2026

   Author

   More actions

   Thanks for the duplicate check — #36006 is indeed the same root problem and its proposals are excellent. I want to clarify one gap my report adds that #36006 does not cover: images pasted from the clipboard have no local file path at all.

   The paste scenario

   When a user pastes an image (rather than @-referencing a file), the image never touches disk:

   • It is stored as a data:image/...;base64,... URL inside the part table of opencode.db (verified: {type:"file", mime:"image/png", filename:"image-2.png", url:"data:..."} with session_id/message_id).
   • filename is a synthetic name (image-2.png), not a path.

   So for pasted images, none of #36006's options help directly:

   • Option C (put the file path in the replacement text) — there is no path to put there.
   • Options A/B (plugin hook / custom template pointing at a path) — same problem.

   The only way to recover a pasted image is via the session/message/part identity (querying the local DB or the serve API GET /session/:id/message). The agent currently cannot know even the current session_id (no env var, no context injection).

   Concrete suggestion to cover both scenarios

   If #36006's Option A hook were extended to include the original part's identity (or its data: URL) rather than only a file path, both scenarios would be covered:

   input: {
     sessionID: string
     messageID?: string
     partID?: string
     dataURL?: string       // for pasted media that never touched disk
     filename?: string
     mime: string
     ...
   }

   That way a plugin can guide the LLM to a tool that reads pasted images from the DB/serve API (which is exactly the workaround my team built: a skill that pulls the latest image part via the serve API and forwards it to a separate vision model).

   Also worth noting: exposing the current session_id to agents (e.g. an OPENCODE_SESSION_ID env var for tool calls) would fix the related issue #33035 and remove the need to guess the session at all.

   Happy to consolidate this detail into #36006 if maintainers prefer a single issue.
4. thdxr commented on Oct 11, 2026

   

   thdxr

   on Oct 11, 2026

   Member

   More actions

   Thanks for the report! OpenCode 2.0 is now out and development is focused on V2, so we're closing issues filed against OpenCode 1. Please upgrade (opencode upgrade --major, details: https://anoma.ly/notes/opencode-2-0), and if this is still relevant on 2.0, reopen this issue or open a new one with your 2.0 version.

   — from 𝕺𝖕𝖊𝖓𝕮𝖔𝖉𝖊
5. thdxr

   closed this as not plannedon Oct 11, 2026