opencode-go: long sessions stuck at provider input limit; auto-compaction never triggers #56296

Activity

1. opencode-agent commented on Oct 17, 2026

   

   opencode-agentBot

   on Oct 17, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed report. I reproduced this on v2.0.32, and it still happens on the latest dev (512165cec).

   To test it, I pointed OpenCode at a mock OpenAI-compatible provider that advertises a 1M context and answers normal requests with HTTP 400 and the gateway's exact message, This text request is too large for the model's configured input limit.. Compaction requests get a successful reply.

   • With that message, opencode run fails with the error, and no compaction request is ever sent.
   • With prompt is too long in its place (a message OpenCode already recognises as context overflow) and everything else the same, OpenCode sends compaction requests right after the 400.

   So the recover-once compaction works, but this gateway's wording isn't on the list of context-overflow errors in packages/ai/src/provider-error.ts, so it never starts. That matches what you saw. The script below reproduces it.

   The other half of your report, the catalog advertising a 1M context while the gateway enforces about 270k, is on the opencode-go side, and a mock can't confirm it. It is tracked in #50446 (same pattern on qwen3.8-flash) and #50761 (route-dependent limits on deepseek-v4.1-flash). Until this is fixed, compacting manually, as you did, is the workaround.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Repro for #56296: a 400 "This text request is too large for the model's configured input limit."
   # is not treated as context overflow, so no compact-and-retry happens.
   # Usage: ./repro.sh            -> uses `opencode` from PATH
   #        OC_DEV=/path/to/repo ./repro.sh -> runs from source
   # MSG overrides the error message (e.g. MSG="prompt is too long" as a control).
   set -u
   DEFAULT_MSG="This text request is too large for the model's configured input limit."
   MSG="${MSG:-$DEFAULT_MSG}"
   W=$(mktemp -d); PORT=${PORT:-18765}
   cat > "$W/mock.ts" <<'TS'
   let n = 0
   const msg = process.env.MSG!
   Bun.serve({ port: Number(process.env.PORT), async fetch(req) {
     const url = new URL(req.url)
     if (url.pathname.endsWith("/models")) return Response.json({ data: [{ id: "flash", object: "model" }] })
     const body: any = await req.json().catch(() => ({}))
     n++
     const text = JSON.stringify(body.messages ?? []).toLowerCase()
     const isSummary = text.includes("into a structured summary") || text.includes("consolidated summary")
     console.error(`[mock] request #${n} messages=${body.messages?.length} summaryRequest=${isSummary}`)
     // Simulate the gateway: the full (oversized) history is rejected; a compaction/summary request succeeds.
     if (!isSummary) return Response.json({ error: { type: "unsupported_request", code: "unsupported_request", message: msg } }, { status: 400 })
     const chunk = (d: any) => `data: ${JSON.stringify({ id: "x", object: "chat.completion.chunk", created: 0, model: "flash", choices: [d] })}\n\n`
     const s = chunk({ index: 0, delta: { role: "assistant", content: "summary ok" } }) + chunk({ index: 0, delta: {}, finish_reason: "stop" }) +
       `data: ${JSON.stringify({ id: "x", object: "chat.completion.chunk", created: 0, model: "flash", choices: [], usage: { prompt_tokens: 10, completion_tokens: 2, total_tokens: 12 } })}\n\ndata: [DONE]\n\n`
     return new Response(s, { headers: { "content-type": "text/event-stream" } })
   } })
   TS
   MSG="$MSG" PORT=$PORT bun "$W/mock.ts" 2> "$W/mock.log" & MP=$!
   sleep 1
   mkdir -p "$W/proj" "$W/home"
   cat > "$W/proj/opencode.json" <<JSON
   { "\$schema": "https://opencode.ai/config.json",
     "provider": { "mock": { "npm": "@ai-sdk/openai-compatible", "name": "Mock gateway",
       "options": { "baseURL": "http://127.0.0.1:$PORT/v1", "apiKey": "x" },
       "models": { "flash": { "name": "flash", "limit": { "context": 1000000, "output": 32000 } } } } } }
   JSON
   cd "$W/proj"
   export HOME="$W/home" XDG_CONFIG_HOME="$W/home/.config" XDG_DATA_HOME="$W/home/.data" XDG_STATE_HOME="$W/home/.state" XDG_CACHE_HOME="$W/home/.cache"
   if [ -n "${OC_DEV:-}" ]; then
     export BUN_OPTIONS="--conditions=browser --preload=$OC_DEV/packages/tui/node_modules/@opentui/solid/scripts/preload.js"
     OC=(bun --conditions=browser --preload=$OC_DEV/packages/tui/node_modules/@opentui/solid/scripts/preload.js $OC_DEV/packages/cli/src/index.ts)
   else OC=(opencode); fi
   echo "version: $("${OC[@]}" --version)"
   timeout 90 "${OC[@]}" run -m mock/flash "hi" 2>&1 | tail -15
   echo "--- mock log ---"; cat "$W/mock.log"
   kill $MP; pkill -x opencode.exe 2>/dev/null; [ -n "${OC_DEV:-}" ] && pkill -f "$OC_DEV/packages/cli/src/index"; sleep 1

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

   added

   reproduced

   on Oct 17, 2026
3. z99733543-cell commented on Oct 17, 2026

   

   z99733543-cell

   on Oct 17, 2026

   Author

   More actions

   Thanks — the reproduction makes sense (the wording isn't in the context-overflow classification, so the recover-once never starts).

   One more data point from the same machine/time, suggesting the compact-and-retry would be worth extending to provider 400s of this class generally: a second failure mode on the same model wedges sessions identically — if any message/tool-result contains an image/file part (e.g. read on a .png), every subsequent request fails with:

   400 [unsupported_request] Message content supports text content only. Images, audio and files are not enabled.
   

   while the catalog advertises capabilities.input: ["text","image"] for the model. Like the size case, the session only unblocks via manual POST /api/session/{id}/compact (the summary path drops attachments).

   Also, FWIW, the size side looks route-dependent rather than a fixed ~270k cap: the same account/day shows successful requests at 272k–900k tokens on this model (max 899,929; incl. one 884,448 fully uncached), and the two wedging points measured ≈276k and ≈302k prompt tokens — consistent with the known #50761 route-dependent limits.
4. rekram1-node

   self-assigned this

   on Oct 20, 2026
5. rekram1-node

   closed this as completedin #57982on Oct 22, 2026