Activity

1. opencode-agent commented on Oct 12, 2026

   

   opencode-agentBot

   on Oct 12, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed analysis! I think this is a duplicate of #52448. In that issue, an MCP tool result with a media type the route can't encode (application/vnd.apache.parquet on OpenAI Chat) is saved in the history, and every later request in the session then fails. It's the same cause you traced: unsupportedMedia only covers types that map to a modality, so other types reach the route's lowerMedia and get rejected.

   Possible duplicates and closely related issues, ordered from most to least likely:

   • mcp: binary file attachment in tool result permanently breaks the session #52448: same problem (unsupported file type from an MCP tool result permanently breaks the session), reported on the OpenAI Chat route. Open PR fix(ai): prevent unsupported tool files from breaking Chat sessions #52517 targets it, but so far it only changes the OpenAI Chat adapter, so the Anthropic Messages route from your report (text/x-patch) would still fail. Your notes on the Anthropic path and the core-level fix would be a useful addition there.
   • session: provider image rejection bricks session — generic 400 error, no recovery path #52042: provider rejects an attachment and the session can't recover (a more general version of the same "stuck session" symptom).
   • bug: tool-result media 400s whole request for models without attachment capability #41163: media in tool results makes the whole request fail for models that can't take attachments.
   • provider: PDF attachment bricks session on GLM models (upstream json_parse_error at messages[N]) #53809: a PDF attachment breaks the session on GLM models.

   The bot will close this issue as a duplicate in 1 day unless you reply explaining how your problem is different.
2. opencode-agent

   added

   pending close

   triaging

   on Oct 12, 2026
3. maggu2810 commented on Oct 13, 2026

   

   maggu2810

   on Oct 13, 2026

   Author

   More actions

   This is not about binary but a general mime type regression (I think so).
   And easily to reproduce with the Atlassian MCP.
   Reading attachements high potentially result into non working session that has to be reverted to the previous message.

   Now I got:
   Anthropic Messages does not support media type text/markdown
   because the Jira ticket has a .md file attached.
4. opencode-agent

   removed

   triaging

   pending close

   on Oct 13, 2026
5. opencode-agent commented on Oct 13, 2026

   

   opencode-agentBot

   on Oct 13, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the follow-up. You're right that this goes beyond the binary-file case in #52448, so I'm keeping this issue open.

   I reproduced it on v2.0.24, and it still happens on the latest dev (4ffede4):

   1. A minimal MCP server has one tool that returns what an attachment download returns: a text part, a text/markdown resource blob and a PNG image.
   2. The model is on the Anthropic Messages route (@opencode/ai/providers/anthropic), pointed at a local mock endpoint. The mock only makes the model call the tool through Code Mode.
   3. The tool call succeeds. The next request fails with Anthropic Messages does not support media type text/markdown and never reaches the endpoint, because it fails while OpenCode builds the request.
   4. An unrelated follow-up in the same session fails with the same error.

   The script below reproduces it end to end, with no provider credentials needed.

   The rejecting check in the Anthropic request builder hasn't changed since at least v2.0.17. So rather than a recent regression, it looks like no text/* type other than text/plain has ever been handled on this route. Core only replaces unsupported media for image, audio, video and PDF types, so text/markdown and text/x-patch reach the builder unchanged. Related: #52448 (same problem on OpenAI Chat). Its open fix, #52517, only changes the OpenAI Chat adapter and does not cover the Anthropic route.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Reproduction for anomalyco/opencode#54798.
   # An MCP tool returns a text/markdown resource blob. On the Anthropic Messages route,
   # the request that sends the tool result back fails locally, and so does every later turn.
   # Needs bun and opencode v2 on PATH (or OC="<command>"). No provider credentials:
   # a local mock Anthropic endpoint only scripts the model's tool calls.
   set -u
   OC=${OC:-opencode}
   W=$(mktemp -d /tmp/oc54798.XXXX); mkdir -p "$W/project"; cd "$W"
   
   cat > mcp-server.ts <<'TS'
   import { createInterface } from "node:readline"
   const send = (m: unknown) => process.stdout.write(JSON.stringify(m) + "\n")
   const png = "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg=="
   const md = Buffer.from("# Notes\n\nREADME.md attached to the ticket.\n").toString("base64")
   createInterface({ input: process.stdin }).on("line", (line) => {
     const msg = JSON.parse(line)
     if (msg.id === undefined) return
     if (msg.method === "initialize")
       return send({ jsonrpc: "2.0", id: msg.id, result: { protocolVersion: msg.params.protocolVersion ?? "2025-06-18", capabilities: { tools: {} }, serverInfo: { name: "attach", version: "1.0.0" } } })
     if (msg.method === "tools/list")
       return send({ jsonrpc: "2.0", id: msg.id, result: { tools: [{ name: "get_attachments", description: "Download ticket attachments", inputSchema: { type: "object", properties: {} } }] } })
     if (msg.method === "tools/call")
       return send({ jsonrpc: "2.0", id: msg.id, result: { content: [
         { type: "text", text: "Downloaded 2 attachments" },
         { type: "resource", resource: { uri: "file:///README.md", mimeType: "text/markdown", blob: md } },
         { type: "image", data: png, mimeType: "image/png" },
       ] } })
     send({ jsonrpc: "2.0", id: msg.id, result: {} })
   })
   TS
   
   cat > mock-anthropic.ts <<'TS'
   // Mock Anthropic Messages endpoint. Asks the model to run Code Mode `execute` with the MCP tool
   // (retrying while the MCP catalog is still loading), then answers with plain text.
   let n = 0
   const sse = (events: any[]) => events.map((e) => `event: ${e.type}\ndata: ${JSON.stringify(e)}\n\n`).join("")
   Bun.serve({ port: 4555, async fetch(req) {
     const body: any = await req.json().catch(() => ({}))
     n++
     const raw = JSON.stringify(body.messages ?? [])
     const results = (raw.match(/tool_result/g) ?? []).length
     const unknown = (raw.match(/Unknown tool/g) ?? []).length
     const hasExecute = (body.tools ?? []).some((t: any) => t.name === "execute")
     console.log(`[mock] request #${n} received (tool results in history: ${results})`)
     const start = { type: "message_start", message: { id: "msg_" + n, type: "message", role: "assistant", model: "claude-test", content: [], stop_reason: null, usage: { input_tokens: 1, output_tokens: 0 } } }
     const callTool = hasExecute && results === unknown && results < 8 && !raw.includes("unrelated")
     const events = callTool
       ? [start,
          { type: "content_block_start", index: 0, content_block: { type: "tool_use", id: "toolu_" + n, name: "execute", input: {} } },
          { type: "content_block_delta", index: 0, delta: { type: "input_json_delta", partial_json: JSON.stringify({ code: "return await tools.attach.get_attachments({})" }) } },
          { type: "content_block_stop", index: 0 },
          { type: "message_delta", delta: { stop_reason: "tool_use" }, usage: { output_tokens: 1 } },
          { type: "message_stop" }]
       : [start,
          { type: "content_block_start", index: 0, content_block: { type: "text", text: "" } },
          { type: "content_block_delta", index: 0, delta: { type: "text_delta", text: "MOCK-REPLY-OK" } },
          { type: "content_block_stop", index: 0 },
          { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 1 } },
          { type: "message_stop" }]
     return new Response(sse(events), { headers: { "content-type": "text/event-stream" } })
   } })
   TS
   
   cat > project/opencode.jsonc <<JSON
   {
     "model": "mockant/claude-test",
     "providers": {
       "mockant": {
         "package": "@opencode/ai/providers/anthropic",
         "settings": { "baseURL": "http://127.0.0.1:4555/v1", "apiKey": "test" },
         "models": { "claude-test": { "capabilities": { "input": ["text", "image", "pdf"] } } }
       }
     },
     "mcp": { "servers": { "attach": { "type": "local", "command": ["bun", "$W/mcp-server.ts"] } } }
   }
   JSON
   (cd project && git init -q)
   
   bun mock-anthropic.ts > mock.log 2>&1 & MOCK=$!
   trap 'kill $MOCK 2>/dev/null' EXIT
   sleep 1
   export XDG_DATA_HOME=$W/xdg/data XDG_CONFIG_HOME=$W/xdg/config XDG_STATE_HOME=$W/xdg/state XDG_CACHE_HOME=$W/xdg/cache
   cd project
   SID=ses_repro54798_$$
   echo "== $($OC --version)"
   echo "== turn 1: the model calls the MCP tool, which returns a text/markdown attachment"
   timeout 120 $OC run --standalone --auto -s $SID "Read the ticket attachments and summarize them" 2>&1 | tail -8
   echo "== turn 2: unrelated follow-up in the same session"
   timeout 90 $OC run --standalone --auto -s $SID "hello, unrelated question" 2>&1 | tail -4
   echo "== mock log (requests that failed locally never show up here)"
   cat ../mock.log

   Run it: bash repro.sh
6. opencode-agent

   added

   reproduced

   on Oct 13, 2026
7. rekram1-node

   self-assigned this

   on Oct 19, 2026
8. rekram1-node

   closed this as completedin #57640on Oct 21, 2026