feat(provider): add acpx provider for any ACP-compatible coding agent

CHANGELOG.md

@@ -18,6 +18,7 @@
 - Added a Grok CLI provider for review, fix, revalidate, and doctor flows, thanks @ebastos.
 - Added `--since <ref>` on `clawpatch review` and `clawpatch revalidate` to restrict runs to features whose owned or context files changed since the given git ref, thanks @mvanhorn.
 - Added Flask route feature mapping for Python projects, including `web/` source roots, common root entry files, non-list method literals, and Python framework detection.
+- Added the `acpx` provider for routing review, fix, and revalidate through ACP-compatible coding agents, thanks @mvanhorn.
 - Added Next.js route mapping for `src/app` and `src/pages` layouts, thanks @obatried.
 - Added first-pass Python mapping for project metadata, console scripts, source groups, pytest suites, and conservative validation defaults, thanks @xiamx.
 - Added progress output for `clawpatch revalidate`, thanks @twidtwid.

README.md

@@ -93,11 +93,11 @@ working tree during the explicit fix command.
 Supported provider names today:
 
 - `codex`: local Codex CLI
+- `acpx`: any ACP-compatible coding agent (Codex / Claude / Pi / Gemini / ...) via openclaw/acpx
+- `grok`: local Grok Build CLI
 - `mock`: deterministic test provider
 - `mock-fail`: failure test provider
 
-Direct OpenAI, Claude, Gemini, and provider panels are not implemented yet.
-
 ## Commands
 
 - `clawpatch init`: create `.clawpatch/`, detect project basics, write config

docs/providers.md

@@ -14,11 +14,12 @@ clawpatch doctor
 Provider names today:
 
 - `codex`: shells out to `codex exec` (default)
+- `acpx`: routes through any ACP-compatible coding agent via `acpx`
 - `grok`: shells out to the xAI Grok Build CLI in headless mode (`grok --prompt-file`)
 - `mock`: deterministic provider for tests and fixtures
 - `mock-fail`: failure provider for tests
 
-### Codex
+## Codex
 
 Codex invocation:
 
@@ -28,19 +29,58 @@ Codex invocation:
 - output: strict JSON schema via `--output-schema`
 - final message capture: `--output-last-message`
 
-### Grok
+Model selection:
 
-The `grok` provider shells out to the local [Grok Build CLI](https://x.ai/cli)
+```bash
+clawpatch review --model <model>
+CLAWPATCH_MODEL=<model> clawpatch review
+```
+
+## ACPX
+
+The `acpx` provider routes through `acpx <agent> exec`, where `<agent>` is any
+ACP-compatible coding agent.
+
+- review / revalidate: `--approve-reads` plus an explicit read-only prompt directive
+- fix: `--approve-all`
+- output: `--format json --json-strict --suppress-reads`, parsed from known ACP NDJSON envelope kinds
+- tested envelope shape: `acpx@^0.8.0`
+
+Permission caveat: `acpx --approve-all` is not the same as `codex --sandbox
+workspace-write`. Codex's workspace-write mode is an enforced sandbox. ACPX
+approval flags control ACP permission prompts; the underlying agent still has
+whatever filesystem and network access its own runtime grants. For untrusted
+code, run `clawpatch fix --provider acpx` inside an isolated checkout. For
+review and revalidate, strict read-only behavior still depends on the underlying
+agent honoring read-only permissions and the prompt directive.
+
+Agent selection uses `--model` as `<agent>` or `<agent>:<model>`, split on the
+first colon:
+
+- unset: agent `codex`, default model
+- `codex`: agent `codex`, default model
+- `claude`: agent `claude`, default model
+- `claude:sonnet-4-5`: agent `claude`, model `sonnet-4-5`
+- `ollama:llama3:70b`: agent `ollama`, model `llama3:70b`
+
+Migration note: `--provider codex --model gpt-5-codex` is not equivalent to
+`--provider acpx --model gpt-5-codex`; the latter selects an ACP agent named
+`gpt-5-codex`. Use `--provider acpx --model codex:gpt-5-codex`.
+
+## Grok
+
+The `grok` provider shells out to the local [Grok Build CLI](https://x.ai/cli).
 
 Install the Grok CLI:
 
 ```bash
 curl -fsSL https://x.ai/cli/install.sh | bash
 ```
 
-Then ensure `grok --version` works and authenticate using the flow supported by the local Grok CLI.
+Then ensure `grok --version` works and authenticate using the flow supported by
+the local Grok CLI.
 
-**Provider selection:**
+Provider selection:
 
 ```bash
 clawpatch review --provider grok
@@ -49,21 +89,14 @@ clawpatch fix --finding <id> --provider grok --model grok-build
 clawpatch doctor --provider grok
 ```
 
-**Model selection** (works for both `codex` and `grok`):
-
-```bash
-clawpatch review --model <model>
-CLAWPATCH_MODEL=<model> clawpatch review
-```
-
-#### How the grok provider works
+How the Grok provider works:
 
-- **Headless mode**: Uses `--prompt-file` + `--output-format json --always-approve --verbatim --cwd <root>`.
-- **Read-only operations** (`review`, `revalidate`): Adds `--disallowed-tools "search_replace,run_terminal_cmd,Agent"` so the agent cannot modify files or run shell commands.
-- **Write operations** (`fix`): Uses full `--always-approve` so the agent can edit files and run validation commands.
-- **Structured output**: The inner prompt includes clawpatch's JSON schema. The provider parses the text field of the JSON envelope returned by `--output-format json` and validates it with Zod (same schemas used for Codex).
-- **Large prompts**: Review and fix prompts can be tens of kilobytes (they embed the full source of owned and context files). Passing such prompts via `-p`/`--single` is unreliable, so the provider always uses `--prompt-file` (the intended mechanism for automation and large inputs).
+- Headless mode: `--prompt-file` plus `--output-format json --always-approve --verbatim --cwd <root>`
+- Read-only operations: adds `--disallowed-tools "search_replace,run_terminal_cmd,Agent"`
+- Write operations: uses full `--always-approve` so the agent can edit files and run validation commands
+- Structured output: validates the returned JSON against the same Zod schemas used for Codex
+- Large prompts: always uses `--prompt-file` instead of passing prompt text on the command line
 
-Direct OpenAI, Claude, Gemini, local-model, and multi-model panel providers are
-not implemented yet. The `grok` provider uses the local `grok` binary for the
-best integration with the surrounding Grok Build environment.
+Direct OpenAI API, local-model, and multi-model panel providers are not
+implemented yet. The `acpx` provider is the generic route for ACP-compatible
+agents; the `grok` provider is a direct integration for the local Grok CLI.

docs/safety.md

@@ -12,7 +12,11 @@ Current safety rules:
 - `fix` requires explicit `--finding <id>`.
 - `fix` refuses a dirty source worktree by default.
 - `.clawpatch/` state changes are allowed during runs.
-- review and revalidate provider calls use a read-only sandbox.
+- review and revalidate provider calls use a read-only sandbox for the `codex`
+  provider. The `acpx` provider relies on `acpx --approve-reads` plus an explicit
+  read-only prompt directive; underlying agents that bypass ACP permissions (e.g.
+  agents running in their own full-access mode) may not be strictly sandboxed.
+  See docs/providers.md.
 - provider output must pass runtime schema validation.
 - feature locks are stored in feature records and `.clawpatch/locks/`; `status`
   surfaces both, and `clean-locks` clears both.

src/exec.ts

@@ -8,6 +8,19 @@ export async function runCommand(
   cwd: string,
   input?: string,
   options: { trimOutput?: boolean } = {},
+): Promise<CommandResult> {
+  const result = await runCommandRaw(command, cwd, input);
+  return {
+    ...result,
+    stdout: options.trimOutput === false ? result.stdout : trimOutput(result.stdout),
+    stderr: options.trimOutput === false ? result.stderr : trimOutput(result.stderr),
+  };
+}
+
+export async function runCommandRaw(
+  command: string,
+  cwd: string,
+  input?: string,
 ): Promise<CommandResult> {
   const started = Date.now();
   const child = spawn(command, {
@@ -47,8 +60,8 @@ export async function runCommand(
     cwd,
     exitCode,
     durationMs: Date.now() - started,
-    stdout: options.trimOutput === false ? stdout : trimOutput(stdout),
-    stderr: options.trimOutput === false ? stderr : trimOutput(stderr),
+    stdout,
+    stderr,
   };
 }