fix(core): recover sessions after HTTP 413 payload rejections #53107

Activity

1. github-actions

   assigned

   kitlangton

   on Oct 4, 2026
2. github-actions commented on Oct 4, 2026

   

   github-actionsBot

   on Oct 4, 2026 – with GitHub Actions

   Contributor

   More actions

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

   • #36552: V2 inline-image payloads cause HTTP 413 failures and prevent session recovery.
   • #14562: Earlier report of HTTP 413 errors with images leaving sessions unrecoverable.
3. rezenderaul

   added a commit that references this issue on Oct 4, 2026

   test(session): lock 413 payload rejection as context overflow

   ...

   e692c7e
4. olobarius commented on Oct 18, 2026

   

   olobarius

   on Oct 18, 2026

   More actions

   Additional data point: the oversized-body rejection does not always come from the provider. With the third-party Anthropic auth plugin @ex-machina/opencode-anthropic-auth (<= 2.0.0-next.5), the plugin's own pre-send check throws Anthropic request body exceeds 10485760 byte limit before any HTTP request is made. OpenCode records it as type: "unknown" without a status, so neither the payload-too-large path (#53108) nor the unclassified-4xx path (#56767) applies, and the session dead-ends the same way.

   Session: 20 screenshots ≈ 13.7 MB since the last compaction, far below the token limit, so token-based auto-compaction never triggered. A manual compaction recovered it (the compaction request was ~175 KB because media are not sent). The plugin has since raised its limit to 32 MiB (ex-machina-co/opencode-anthropic-auth#282), but errors thrown from request hooks for oversized bodies (or any failure before a response on a request above a few MiB) could also be treated as payload rejections. OpenCode 2.0.32, Windows.
5. opencodejr commented on Oct 27, 2026

   

   opencodejrBot

   on Oct 27, 2026

   Contributor

   More actions

   Thanks for the detailed report. I've reproduced this on 2.0.20, and it still happens on the latest dev (c250c272).

   What I did: I pointed a mock OpenAI-compatible gateway at OpenCode. It answers any request body over 200 KB with HTTP 413 and {"error":{"code":"proxy_error","message":"Request body exceeds the configured limit."}}. The model has a 1M-token context and compaction.auto is on. I sent one text turn, then a turn with an image of about 480 KB, then a plain --continue.

   What happened: the image turn and the next turn both fail with Error: Request body exceeds the configured limit. The gateway only ever receives requests that still contain the image, and no compaction request is sent.

   Control: if only the error code changes to request_too_large, OpenCode does send a compaction request without the image. That matches your analysis: the runner in packages/core/src/session/runner/step.ts only starts recovery for context-overflow, not payload-too-large. In that control run, the request rebuilt after compaction still included the image and was rejected again, so that is worth checking in a fix as well.

   The script below runs the reproduction. Set ERR_CODE=request_too_large for the control. I couldn't search existing issues, so I haven't checked how #36552 relates.

   Reproduction script (run.sh)

   #!/usr/bin/env bash
   # Reproduces #53107: HTTP 413 "payload-too-large" never triggers recovery compaction.
   # Usage: ERR_CODE=proxy_error|request_too_large ./run.sh   (OC="bun ... packages/cli/src/index.ts" to run from source)
   set -u
   cd "$(dirname "$0")"
   OC=${OC:-opencode}
   PORT=${PORT:-4599}
   ERR_CODE=${ERR_CODE:-proxy_error} PORT=$PORT bun mock.ts > mock.log 2>&1 &
   MOCK=$!; sleep 1; pkill -f "[s]erve --service"; sleep 1
   PROJ=$(mktemp -d /tmp/opencode/p413.XXXX); cp big.png "$PROJ/"; mkdir -p "$PROJ"/{home,data,cfg,state,cache}; cd "$PROJ"; git init -q
   # From source: the CLI respawns itself as `bun src/index.ts serve --service`, which needs the opentui preload.
   [ "$OC" != opencode ] && echo 'preload = ["/workspace/repo/packages/tui/node_modules/@opentui/solid/scripts/preload.js"]' > bunfig.toml
   export HOME=$PROJ/home XDG_DATA_HOME=$PROJ/data XDG_CONFIG_HOME=$PROJ/cfg XDG_STATE_HOME=$PROJ/state XDG_CACHE_HOME=$PROJ/cache
   export OPENCODE_CONFIG_CONTENT=$(cat <<JSON
   {"update":"disable","model":"gw/big","compaction":{"auto":true},
    "providers":{"gw":{"name":"Gateway","package":"aisdk:@ai-sdk/openai-compatible",
     "settings":{"apiKey":"k","baseURL":"http://127.0.0.1:$PORT/v1"},
     "models":{"big":{"name":"Big","capabilities":{"tools":true,"input":["text","image"],"output":["text"]},
      "cost":{"input":0,"output":0},"limit":{"context":1000000,"output":10000}}}}}}
   JSON
   )
   echo "== turn 1 (text)"; timeout 120 $OC run "hello" 2>&1 | tail -5
   echo "== turn 2 (with image, body > 200KB gateway limit)"; timeout 120 $OC run --continue -f big.png "what is in this image" 2>&1 | tail -8
   echo "== turn 3 (continue)"; timeout 120 $OC run --continue "continue" 2>&1 | tail -8
   pkill -f "[s]erve --service"; kill $MOCK; echo "== mock log"; cat "$OLDPWD/mock.log" 2>/dev/null || cat /workspace/repro/mock.log

   Run it: bash run.sh
6. opencodejr

   added

   reproduced

   on Oct 27, 2026