openclaw · steipete · Jun 15, 2026 · Jun 15, 2026
@@ -2,6 +2,7 @@
 
 ## 0.6.1 - Unreleased
 
+- Removed the direct MiniMax HTTP provider and its transport dependency; provider integrations are now explicitly limited to coding harnesses and agent CLIs.
 - Added uv workspace member mapping with repository-relative paths and member-local test commands while preserving mixed root source and test groups, thanks @srnm.
 
 ## 0.6.0 - 2026-06-11

diff --git a/README.md b/README.md
@@ -129,7 +129,6 @@ Supported provider names today:
 - `claude`: local Claude Code CLI in print mode
 - `cursor`: local Cursor Agent CLI (experimental; `doctor` is enabled by default)
 - `grok`: local Grok Build CLI
-- `minimax`: MiniMax OpenAI-compatible HTTP API; supports `map`, `review`, and `revalidate`, but not `fix`
 - `opencode`: local OpenCode CLI
 - `pi`: local Pi coding agent in print mode
 - `mock`: deterministic test provider
@@ -230,4 +229,6 @@ to features so runs can resume and be audited.
 - Provider output is parsed through strict schemas.
 - Symlinked directories and generated build output are skipped during mapping.
 
-See `docs/spec.md` for the longer product and implementation spec.
+See [`VISION.md`](VISION.md) for project scope and `docs/spec.md` for the
+longer product and implementation spec. Provider integrations are limited to
+coding harnesses and agent CLIs; direct model API providers are out of scope.
diff --git a/VISION.md b/VISION.md
@@ -0,0 +1,69 @@
+# Clawpatch Vision
+
+Clawpatch is an automated code-review and repair tool built on coding
+harnesses. It maps repositories into bounded features, asks a coding agent to
+review or repair them, validates the result, and preserves an auditable record.
+
+Project overview: [`README.md`](README.md)
+Provider details: [`docs/providers.md`](docs/providers.md)
+
+## Core Boundary
+
+Clawpatch integrates coding harnesses and agent CLIs, not model APIs.
+
+A provider adapter belongs in Clawpatch when it launches an installed coding
+harness that owns model transport, authentication, tool use, and coding-agent
+behavior. Current examples include Codex, ACPX, Claude Code, Cursor Agent, Grok
+Build, OpenCode, and Pi.
+
+Clawpatch will not add direct model inference integrations. This excludes:
+
+- HTTP clients or SDKs for chat, responses, or model inference APIs
+- Clawpatch-owned provider API keys, base URLs, request envelopes, or billing behavior
+- model-specific retry, quota, streaming, or structured-output transports
+- read-only chat providers that cannot operate as coding harnesses
+
+If a model provider needs support, add it to a coding harness or ACP adapter
+first, then integrate that harness through its stable CLI or protocol. Model
+selection and API authentication remain the harness's responsibility.
+Adapters may pass documented authentication environment variables through to
+the harness; they must not use those credentials to call model APIs themselves.
+
+This boundary does not prohibit network access for non-model infrastructure
+such as package registries. It specifically keeps model transport out of
+Clawpatch.
+
+## Principles
+
+### 1. Harnesses own agent execution
+
+Clawpatch should coordinate coding agents, not reimplement their model clients.
+Harnesses own authentication, sessions, tools, permissions, model protocols,
+and provider-specific compatibility.
+
+### 2. Review and repair are one workflow
+
+Provider integrations should fit the full coding lifecycle: inspect a checkout,
+produce structured findings, plan or apply explicit fixes, and revalidate.
+Partial support needs a strong reason and must still come from a coding
+harness, not a bare inference endpoint.
+
+### 3. Safety stays explicit
+
+Review and revalidation should be read-only. Repair remains an explicit command
+with clean-worktree checks, bounded permissions, validation, and an audit trail.
+Clawpatch should expose the harness's security boundary honestly rather than
+claiming stronger isolation than the harness provides.
+
+### 4. Keep provider code small
+
+An adapter should translate Clawpatch's stable inputs and schemas to a harness's
+documented CLI or protocol. Provider-specific transport stacks and credential
+systems are out of scope.
+
+## Contribution Guardrail
+
+Pull requests and issues proposing direct model API providers are out of scope
+and will be closed. Proposals for new coding harnesses should document the
+harness command or protocol, authentication ownership, read/write permission
+model, structured-output path, timeout behavior, and real review/repair proof.
diff --git a/docs/index.md b/docs/index.md
@@ -45,6 +45,7 @@ stderr so pipes stay parseable.
 - **Fixing findings.** [Patching](patching.md) documents the explicit fix workflow and validation steps.
 - **Reading reports.** [Reporting](reporting.md) shows how to generate Markdown reports and filter by severity.
 - **Configuring providers.** [Providers](providers.md) lists supported backends and future provider integration plans.
+- **Understanding project scope.** [Vision](https://github.com/openclaw/clawpatch/blob/main/VISION.md) defines the coding-harness-only provider boundary.
 
 ## All features
 
@@ -65,4 +66,4 @@ stderr so pipes stay parseable.
 
 ## Project
 
-Active development; the [changelog](https://github.com/openclaw/clawpatch/blob/main/CHANGELOG.md) tracks recent releases. Goals and implementation details in [spec.md](spec.md). Released under the [MIT license](https://github.com/openclaw/clawpatch/blob/main/LICENSE).
+Active development; the [changelog](https://github.com/openclaw/clawpatch/blob/main/CHANGELOG.md) tracks recent releases. Product scope lives in [VISION.md](https://github.com/openclaw/clawpatch/blob/main/VISION.md), with implementation details in [spec.md](spec.md). Released under the [MIT license](https://github.com/openclaw/clawpatch/blob/main/LICENSE).
@@ -7,6 +7,9 @@ description: "AI provider configuration and model selection"
 
 The default provider is the local Codex CLI.
 
+Clawpatch integrates coding harnesses and agent CLIs only. Direct model API
+providers are out of scope; see the project [vision](../VISION.md).
+
 ```bash
 clawpatch doctor
 ```
@@ -17,7 +20,6 @@ Provider names today:
 - `acpx`: routes through any ACP-compatible coding agent via `acpx`
 - `claude`: shells out to Claude Code in print mode (`claude -p`)
 - `grok`: shells out to the xAI Grok Build CLI in headless mode (`grok --prompt-file`)
-- `minimax`: calls the MiniMax OpenAI-compatible HTTP API directly
 - `opencode`: shells out to `opencode run --format json`
 - `pi`: shells out to `pi -p` (non-interactive print mode)
 - `cursor`: shells out to `cursor-agent -p --output-format json`
@@ -350,55 +352,3 @@ implementation uses `--trust` for the explicit trusted-workspace path and never
 uses `--force` or `--yolo`. Complete HITL verification before promoting this to
 default provider support, especially for ambient rules, MCP configuration,
 temporary prompt file handling, timeout behavior, and any claimed read-only mode.
-
-## MiniMax
-
-The `minimax` provider calls the MiniMax OpenAI-compatible HTTP API directly. It
-requires a Token Plan API key in `MINIMAX_API_KEY` and defaults to
-`https://api.minimax.io/v1`.
-
-```bash
-export MINIMAX_API_KEY=sk-cp-...
-clawpatch doctor --provider minimax
-clawpatch review --provider minimax
-clawpatch review --provider minimax --model MiniMax-M2.7-highspeed
-```
-
-How the MiniMax provider works:
-
-- Endpoint: `POST ${MINIMAX_BASE_URL:-https://api.minimax.io/v1}/chat/completions`
-  with `Authorization: Bearer ${MINIMAX_API_KEY}`. `MINIMAX_BASE_URL` is trimmed,
-  normalized, and must use `https` unless it targets loopback HTTP for local
-  development; the bearer token is sent to that configured endpoint.
-- Operations: `map`, `review`, and `revalidate` are supported. `fix` is not
-  supported because the chat completions API cannot edit the worktree; it fails
-  before checking credentials or making network calls with `unsupported-provider`
-  and exit code 2.
-- Structured output: MiniMax M-series Chat Completions do not document OpenAI
-  `response_format.json_schema` support. Clawpatch therefore sends the provider
-  schema in the prompt, disables M3 thinking with `thinking: {type: "disabled"}`,
-  asks for `reasoning_split`, extracts the returned JSON, and validates it
-  locally with the same Zod schemas used by other providers.
-- Model selection: `--model <name>` sets the request `model`; otherwise
-  `MINIMAX_MODEL` is used, then `MiniMax-M3`.
-- HTTP failures: `401` and `403` map to exit code 4, `429` maps to exit code 5,
-  and other non-2xx statuses map to exit code 1. Embedded MiniMax auth,
-  rate-limit, balance, and usage-limit status codes are mapped the same way.
-  Error bodies are reduced to safe status/type/code signals and byte counts.
-- Timeout: 30 minutes by default for provider calls, override with
-  `CLAWPATCH_MINIMAX_TIMEOUT_MS` or `CLAWPATCH_PROVIDER_TIMEOUT_MS`. The `/models`
-  doctor probe uses a 30-second timeout. Clawpatch uses a custom undici
-  dispatcher and an operation abort signal so socket headers/body timeouts and
-  full response-body reads track the configured timeout.
-- Bounds: request bodies over 64 MiB and responses over 10 MiB fail before
-  parsing; error bodies are capped separately.
-
-Permission caveat: MiniMax is a remote API provider. Review inputs are sent to
-the configured MiniMax endpoint, and read-only behavior depends on the remote
-model following the prompt. For untrusted code, run clawpatch in an isolated
-checkout and avoid sending secret-bearing files.
-
-Direct local-model and multi-model panel providers are not implemented yet. The
-`acpx` provider is the generic route for ACP-compatible agents; the `grok`,
-`opencode`, `pi`, and `cursor` providers are direct integrations for local CLIs;
-the `minimax` provider is a direct integration for the MiniMax HTTP API.
diff --git a/docs/spec.md b/docs/spec.md
@@ -421,8 +421,8 @@ Initial config:
     ".clawpatch/**"
   ],
   "provider": {
-    "name": "openai",
-    "model": "gpt-5.2"
+    "name": "codex",
+    "model": null
   },
   "commands": {
     "typecheck": null,
@@ -906,7 +906,6 @@ Implemented providers:
 - `claude`: Claude Code CLI in print mode.
 - `cursor`: experimental Cursor Agent CLI integration.
 - `grok`: Grok Build CLI.
-- `minimax`: MiniMax OpenAI-compatible HTTP API for map, review, and revalidate.
 - `opencode`: OpenCode CLI.
 - `pi`: pi coding agent.
 - `mock` / `mock-fail`: deterministic test providers.

@@ -9,6 +9,7 @@
   "files": [
     "dist",
     "README.md",
+    "VISION.md",
     "LICENSE"
   ],
   "type": "module",
@@ -30,7 +31,6 @@
     "crabbox:warmup": "crabbox warmup"
   },
   "dependencies": {
-    "undici": "^6.26.0",
     "zod": "^4.4.3"
   },
   "devDependencies": {

diff --git a/src/package-smoke.test.ts b/src/package-smoke.test.ts
@@ -28,14 +28,14 @@ async function packageSmokeTesting(): Promise<PackageSmokeTesting["packageSmokeT
 describe("package smoke harness", () => {
   it("installs the packed clawpatch artifact with packed runtime dependencies", async () => {
     const smoke = await packageSmokeTesting();
-    const dependencySource = "/repo/node_modules/.pnpm/undici@6.26.0/node_modules/undici";
+    const dependencySource = "/repo/node_modules/.pnpm/zod@4.4.3/node_modules/zod";
     const clawpatchTarball = "/tmp/clawpatch-0.5.1.tgz";
-    const dependencyTarball = "/tmp/undici-6.26.0.tgz";
+    const dependencyTarball = "/tmp/zod-4.4.3.tgz";
 
     const dependencyNames = smoke.runtimeDependencyNames({
-      dependencies: { undici: "^6.26.0", zod: "^4.4.3" },
+      dependencies: { zod: "^4.4.3" },
     });
-    expect(dependencyNames).toEqual(["undici", "zod"]);
+    expect(dependencyNames).toEqual(["zod"]);
 
     const packArgs = smoke.packDependencyArgs({
       dependencyPath: dependencySource,