[heydryft]

Summary

Add an optional Runcrate memory provider preset using the existing OpenAI Chat Completions transport, following the merged presets in #420 and #432. Runcrate serves open models (DeepSeek, GLM, Qwen, Kimi, gpt-oss and others) over an OpenAI-compatible API.

• memoryProvider: "runcrate": endpoint default https://api.runcrate.ai/v1, model default deepseek-ai/DeepSeek-V4.1-Flash, RUNCRATE_API_KEY fallback, explicit env:// key configuration, and a distinct runcrate session tag.
• Factory, configuration validation, readiness detection, example/template, and all README languages updated.
• Tests cover authentication, request serialization (tool_choice: "required"), tool response extraction, overrides, environment-key resolution, and rejection of remote-provider changes from project configuration.
• No defaults change for existing users. Users opt in with their own key (created at https://www.runcrate.ai/dashboard/api-keys?utm_source=opencode-mem&utm_medium=directory&utm_campaign=gpu_pricing); the docs name the destination for conversation data (https://api.runcrate.ai).

Disclosure: I work on Runcrate.

Test plan

• Exercised locally: real calls through RuncrateProvider.executeToolCall (config built by buildMemoryProviderConfig({ memoryProvider: "runcrate", memoryApiKey }), forced tool_choice: "required") returned save_memories tool data with success: true on deepseek-ai/DeepSeek-V4.1-Flash (the default), zai-org/GLM-5.3-Flash and Qwen/Qwen3.8-Flash. Example result: {"memories":["Prefers bun over npm as the package manager","Prefers tabs over spaces for indentation"]}. AIProviderFactory.createProvider("runcrate", ...) returned a RuncrateProvider.
• No known limitations beyond those of other OpenAI-compatible presets.

Checklist

• Branch is based on main (do not open PRs from commits already on main)
• bun test passes: 975 pass, 4 skip, 0 fail across 133 files (after bun run build, which the plugin-loader contract tests need)
• bun run typecheck passes (included in check)
• bun run check passes (format + lint + typecheck)
• Docs / README updated when behavior or contribution process changes

This PR was prepared with AI assistance: an AI coding agent wrote the change and ran the checks above.

[EyJunge1]

@heydryft Pls fix the action issues

[heydryft]

@EyJunge1 Looked into it. The four failures are not from this change. They all fail in npm install of the packed tarball with ETARGET No matching version found for @redis/json@6.3.1.

That package comes in transitively (@opencode/plugin > @opencode/util > @effect/platform-node > redis@6.3.1). redis@6.3.1 was published to npm at 09:18:49Z, but its @redis/json@6.3.1 dependency only at 09:23:16Z. The Linux and macOS jobs ran npm install between 09:22:09Z and 09:22:58Z, inside that gap. The Windows job is slower and installed after 09:23:16Z, which is why it passed. This PR does not touch package.json or bun.lock.

I re-ran the whole package smoke step locally on macOS (arm64, Node 24) against this branch: bun run build, npm pack, npm install --ignore-scripts of the tarball, then native-deps-smoke.mjs, verify-libsql-vector.mjs and smoke-test.mjs. All pass now that @redis/json@6.3.1 is on the registry.

I pushed an empty commit (55ac05f) to trigger a fresh run, since I can't re-run jobs here. The new Quality and Platform Package Smoke runs are waiting for workflow approval. Re-running the failed jobs of run 40820080040 would work too.