config: legacy provider shape silently drops compatibility for custom provider models #55652

Activity

1. opencode-agent commented on Oct 15, 2026

   

   opencode-agentBot

   on Oct 15, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed report. I can confirm this on 2.0.29, and it still happens on the latest dev (4397ee44).

   With the acme provider in the old provider / npm / options shape and a provider-level compatibility: { reasoningField: "reasoning_content" }, GET /api/model returns acme/glm-5.2 without compatibility. The same provider in the providers / package / settings shape returns compatibility: { reasoningField: "reasoning_content" }. The script below runs both configs against opencode serve and prints the result for each.

   It looks like the config loader reads old-shape providers with the v1 provider schema. That schema has no compatibility field, so the key is dropped with no warning. For now, writing the provider in the providers shape (as you did) works around it.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Reproduces #55652: provider-level `compatibility` dropped for legacy V1 provider shape.
   # Usage: ./repro.sh [opencode command...]   (default: opencode)
   set -u
   OC=("${@:-opencode}")
   [ $# -eq 0 ] && OC=(opencode)
   run() { # $1=label $2=config json
     local dir; dir=$(mktemp -d)
     mkdir -p "$dir/home/.config/opencode" "$dir/proj"
     printf '%s' "$2" > "$dir/home/.config/opencode/opencode.json"
     echo "== $1"
     local port=$((20000 + RANDOM % 20000))
     export HOME="$dir/home" XDG_CONFIG_HOME="$dir/home/.config" XDG_DATA_HOME="$dir/home/.local/share" \
        XDG_STATE_HOME="$dir/home/.local/state" XDG_CACHE_HOME="$dir/home/.cache" OPENCODE_PASSWORD=repro-pass
     (cd "$dir/proj" && exec "${OC[@]}" serve --hostname 127.0.0.1 --port "$port" >"$dir/serve.log" 2>&1) &
     local pid=$!
     for _ in $(seq 1 120); do curl -s -o /dev/null "http://127.0.0.1:$port/" && break; sleep 1; done
     (cd "$dir/proj" && "${OC[@]}" api --server "http://127.0.0.1:$port" get /api/model 2>&1) | python3 -c '
   import json,sys
   raw=sys.stdin.read()
   try: d=json.loads(raw)
   except Exception: print(raw[:2000]); sys.exit()
   items=d if isinstance(d,list) else d.get("data",d.get("items",d))
   if isinstance(items,dict): items=list(items.values())
   for m in items:
     if isinstance(m,dict) and m.get("providerID")=="acme":
       print(json.dumps({k:m.get(k) for k in ("id","providerID","package","compatibility")}))
   '
     kill $pid 2>/dev/null; wait $pid 2>/dev/null
   }
   run "legacy V1 shape (provider/npm/options)" '{"provider":{"acme":{"npm":"@ai-sdk/openai-compatible","name":"Acme","options":{"baseURL":"http://127.0.0.1:8788/v1","apiKey":"x"},"compatibility":{"reasoningField":"reasoning_content"},"models":{"glm-5.2":{"name":"GLM-5.2","limit":{"context":1000000,"output":64000}}}}}}'
   run "V2 shape (providers/package/settings)" '{"providers":{"acme":{"name":"Acme","package":"@opencode/ai/providers/openai-compatible","settings":{"baseURL":"http://127.0.0.1:8788/v1","apiKey":"x"},"compatibility":{"reasoningField":"reasoning_content"},"models":{"glm-5.2":{"name":"GLM-5.2","limit":{"context":1000000,"output":64000}}}}}}'

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

   added

   reproduced

   on Oct 15, 2026
3. Finn763 commented on Oct 15, 2026

   

   Finn763

   on Oct 15, 2026

   Contributor

   More actions

   Taking this one. The legacy (provider/npm/options) shape is validated by the v1 provider schema, which has no compatibility field, so the key is silently dropped. I'll extend the v1 schema to preserve it (same field definition as the v2 shape) and pin it with a config-load test.
4. Finn763 commented on Oct 15, 2026

   

   Finn763

   on Oct 15, 2026

   Contributor

   More actions

   PR is up: #55703 — the v1 provider schema now accepts the optional compatibility block (same ProviderCompatibility schema the v2 shape uses) and migrateStandardProvider carries it through, with the legacy setCacheKey derivation merged underneath an explicit block. Two normalization cases added; both are red on the base (the block was dropped) and green after.
5. opencode-agent commented on Oct 16, 2026

   

   opencode-agentBot

   on Oct 16, 2026 – with OpenCode Agent

   Contributor

   More actions

   Closing this as not planned. The config mixes the two shapes: compatibility is a V2 providers field and has never been part of the V1 provider schema, so it isn't recognized inside a V1 provider entry.

   Use one shape or the other:

   • V2 (recommended): providers + package + settings + compatibility, as in your second example.
   • V1: set the reasoning field per model with interleaved, which is migrated to compatibility.reasoningField:

   "models": {
     "glm-5.2": { "interleaved": { "field": "reasoning_content" } }
   }
6. opencode-agent

   closed this as not plannedon Oct 16, 2026