Attached AVIF images are silently dropped from model context in V2 #41722

Activity

1. github-actions

   assigned

   jlongster

   on Aug 11, 2026
2. github-actions commented on Aug 11, 2026

   

   github-actionsBot

   on Aug 11, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of or closely related to existing issues. Please check:

   • Support AVIF images or auto-convert to supported format before sending to API #17772: Same AVIF image support request — OpenCode should auto-convert or skip AVIF gracefully (closed, but the underlying problem is still present in V2 per this report)
   • tui: attached image path is omitted from model context #41454: Related V2 bug where attached image metadata/path is omitted from model context
3. feiiiiii5 commented on Oct 1, 2026

   

   feiiiiii5

   on Oct 1, 2026

   More actions

   Adding scope to this report rather than opening a second one, since it is the same attachmentContent fall-through. Two things beyond the image formats:

   1. It is not only images. attachmentContent (packages/core/src/session/runner/to-llm-message.ts:75-83) branches on text/plain, application/x-directory, image/{png,jpeg,gif,webp} and application/pdf, then falls through to return [] for everything else. Mime.detect (packages/core/src/mime.ts:16) returns application/octet-stream for any non-text bytes, and session/prompt.ts:113 uses it when the attachment carries no mime — so an ordinary .docx or .zip attached from disk lands on exactly this path. That is the default, not an edge case.

   2. The message can disappear entirely. The caller is to-llm-message.ts:266-272:

   const content = [
     ...(message.skills ?? []).flatMap((skill) => (skill.text === undefined ? [] : [Message.text(skill.text)])),
     ...(message.text === "" ? [] : [Message.text(message.text)]),
     ...userAttachmentContent(message.files ?? []),
   ]
   if (content.length === 0) return []

   With an unsupported attachment and empty text, the whole user message is dropped and the step is sent with no new input.

   There is already a convention for the fix: SessionModelRequest.unsupportedParts turns input it cannot represent into a visible part — ERROR: Cannot read "logo.png" (this model does not support image input). Inform the user. (model-request.ts:129-136, test "replaces unsupported user media with a visible error"). It only inspects parts that exist, so it never sees an attachment that attachmentContent already returned [] for. The v1 migration renders the same idea as [Attachment unavailable after migration: … (application/octet-stream)] (test/v1-migration.test.ts:198).

   A visible fallback in attachmentContent would make this consistent with both. I can send a PR with that if useful — is that the shape you want, or would you prefer it surface as an explicit error?
4. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   More actions

   Thanks for the detailed report. I can confirm this on v2.2.0, and it still happens on the latest dev (b7992598).

   Steps:

   1. With a model that supports image input, paste the path of a small .avif file into the TUI prompt. The prompt shows [Image 1] and a file red.avif chip.
   2. Submit. The request sent to the model has only the text Describe this image: [Image 1] : no image part, no Attached file text, no warning.
   3. Pasting a .png the same way does send an image_url (data:image/png;base64,…) part.
   4. Even with a text-only model, the PNG gets an explicit "this model does not support image input" note, but the AVIF still disappears with nothing in its place.

   The screenshot and recording below show the AVIF turn, where the model received text only, next to the PNG turn, which included the image. The script below reproduces it with a simulated model.

   This matches what you described: to-llm-message.ts only forwards png/jpeg/gif/webp, and attachmentContent returns nothing for other image types, even though the TUI accepts them. Related: #17772 (an earlier AVIF request, where the provider rejected the image instead of OpenCode dropping it).

   recording.mp4

   

   Reproduction script (drive.ts)

   // Reproduces anomalyco/opencode#41722: a pasted AVIF image gets an [Image N]
   // label in the TUI, but the request sent to the model contains no image part.
   // A PNG is pasted the same way as a control.
   import { Effect, Stream } from "effect"
   import { Llm, OpenCodeDriver } from "opencode-drive"
   import { execFileSync } from "node:child_process"
   import { mkdirSync } from "node:fs"
   
   const dir = "/tmp/opencode/repro-41722"
   mkdirSync(dir, { recursive: true })
   const avif = `${dir}/red.avif`
   const png = `${dir}/blue.png`
   // tiny real images (requires ffmpeg with libaom)
   execFileSync("ffmpeg", ["-y", "-loglevel", "error", "-f", "lavfi", "-i", "color=c=red:s=64x64", "-frames:v", "1", "-c:v", "libaom-av1", "-still-picture", "1", avif])
   execFileSync("ffmpeg", ["-y", "-loglevel", "error", "-f", "lavfi", "-i", "color=c=blue:s=64x64", "-frames:v", "1", png])
   
   const paste = (text: string) => `\x1b[200~${text}\x1b[201~`
   
   const summarize = (body: unknown) => {
     const json = JSON.stringify(body)
     const last = (body as any)?.messages?.at(-1)
     const content = Array.isArray(last?.content) ? last.content : [last?.content]
     return {
       parts: content.map((p: any) =>
         typeof p === "string" ? `text:${JSON.stringify(p)}` : p?.type === "text" ? `text:${JSON.stringify(p.text)}` : `${p?.type}${p?.image_url ? ":" + String(p.image_url.url).slice(0, 30) : ""}`,
       ),
       hasAvif: json.includes("image/avif"),
       hasPng: json.includes("image/png"),
     }
   }
   
   export default OpenCodeDriver.use(
     {
       ...(process.env.OC_DEV ? { opencode: { dev: process.env.OC_DEV } } : {}),
       tui: { recording: true },
       // make the simulated model accept image input, like any vision model
       setup: ({ config }) =>
         Effect.sync(() => {
           const model = (config as any).providers.simulation.models["gpt-sim-model"]
           model.capabilities.input = ["text", "image"]
         }),
     },
     ({ ui, llm }) =>
       Effect.gen(function* () {
         yield* llm.serve((request, index) => {
           const s = summarize(request.body)
           console.log(`REQUEST ${index}: ${JSON.stringify(s)}`)
           return Stream.make(Llm.text(`Request ${index + 1} parts: ${s.parts.join(" | ")}`))
         })
         yield* ui.type("Describe this image: ")
         yield* ui.type(paste(avif))
         yield* ui.waitFor("[Image 1]")
         yield* ui.screenshot("avif-attached")
         yield* ui.enter()
         yield* ui.waitFor("parts:", { timeout: 20000 })
         yield* Effect.sleep("1 second")
         yield* ui.screenshot("avif-response")
   
         yield* ui.type("And this one: ")
         yield* ui.type(paste(png))
         yield* ui.waitFor("[Image")
         yield* ui.enter()
         yield* ui.waitFor("Request 2", { timeout: 20000 })
         yield* Effect.sleep("1 second")
         yield* ui.screenshot("png-response")
       }),
   )

   Run it with opencode-drive: opencode-drive run ./drive.ts. It drives opencode from your PATH; set OC_DEV to the path of an OpenCode checkout to run that from source instead.
5. opencodejr

   added

   reproduced

   on Oct 28, 2026