@wombatfish/opencode-toolcall-repair

OpenCode plugin that fixes the finite set of tool-call wire-format violations open models emit (DeepSeek, Qwen, GLM, Kimi, MiniMax, Granite…) — the "harness gap," not a model gap.

The model's reasoning is fine; it just gets the wire shape wrong: sends null for an optional field instead of omitting it, a bare "foo" where the schema wants ["foo"], or a stringified '{"k":"v"}' where it wants an object. OpenCode's AI SDK Zod-validates tool args and rejects these before they execute, so the call dies (routed to opencode's invalid tool).

This plugin runs a tiny in-process OpenAI-compatible proxy that repairs those args before opencode validates them.

opencode → http://localhost:11435/v1 (this proxy) → http://localhost:11434 (Ollama)
                         │ per request: read tools[] JSON schemas
                         │ per response: repair tool_calls[].function.arguments

What it repairs

Schema-driven, from the request's tools[].function.parameters:

| # | Violation | Fix | |---|-----------|-----| | 1 | null on a non-required field | omit the field | | 2 | bare string / "[\"a\",\"b\"]" where schema wants array | ["..."] / parsed array | | 3 | '{"k":"v"}' where schema wants object | parsed object |

Tool-name case is not repaired here — opencode already does that in experimental_repairToolCall.

Out of scope (reasoning failures no format fix can solve): wrong tool, wrong order, hallucinated parameter, context-boundary loss.

Install (per workstation)

1. Add the plugin to opencode.json:

   { "plugin": ["@wombatfish/opencode-toolcall-repair"] }

   OpenCode auto-installs it (Bun). For local dev before publishing, point at the source: "plugin": ["file:///D:/Projects/opencode-toolcall-repair"].

2. Add a provider that routes through the proxy, and point your open models at it:

   {
     "provider": {
       "ollama-repair": {
         "npm": "@ai-sdk/openai-compatible",
         "name": "Ollama (repaired)",
         "options": { "baseURL": "http://localhost:11435/v1" },
         "models": {
           "deepseek-v4-pro:cloud": { "name": "DeepSeek V4 Pro (repaired)" }
         }
       }
     }
   }

That's it. Select a model under Ollama (repaired); keep the direct Ollama provider alongside for A/B.

How it runs

The plugin spawns the proxy as a separate detached bun process on load (port-probe first, so a second opencode session reuses the running daemon instead of double-spawning). It must be a separate process, not in-process: opencode is the HTTP client, and an in-process Bun.serve makes it the server too — the streaming chat/completions response deadlocks on the shared event loop (verified). A separate daemon also owns its listen socket cleanly.

Requires bun on PATH (opencode's own ecosystem dependency). Daemon log: <tmpdir>/opencode-toolcall-repair.log.

Because it speaks the stable OpenAI HTTP wire format, the package has no ai/AI-SDK version coupling — it survives opencode upgrades.

Config (env)

| Var | Default | Meaning | |-----|---------|---------| | REPAIR_PROXY_PORT | 11435 | proxy listen port (must match the provider baseURL) | | REPAIR_UPSTREAM | http://localhost:11434 | upstream origin (Ollama) |

Develop

bun install
bun test                 # unit (repair core + SSE) + integration (proxy vs mock upstream)
bun run start            # run the proxy standalone

Publish

Tag vX.Y.Z → GitHub Actions publishes to npm via trusted publishing (OIDC). Configure the trusted publisher for this package on npmjs.com first (no long-lived token).

License

MIT