test(conformance): replay the recording host's optional tools

Hosted Linux on wave ce1ecc8dc (PR CI 36750896459, and both shared-process
passes of 36750913302) drifted in 6 events goldens: the tool_request_snapshot
counted 41 registry-only tools against the recorded 43. Four tools register
only when this host has their backend (code_execution: Python, js_execution:
Node, pandoc_convert: pandoc, image_ocr: Vision/tesseract); the Mac that
recorded the goldens has all four, the runner lacks pandoc and OCR.

dependencies::host_tool_available is now the one probe seam for those four.
Under a conformance replay it answers with recorded_platform.host_tools
(thread-scoped, test builds, unix only), beside os/shell/sandbox posture;
production probes exactly as before. No golden is re-recorded; each case
gains the recorded list, and a case without it fails loud.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: CodeWhale Bot <[email protected]>

crates/tui/src/conformance/events.rs

@@ -288,6 +288,7 @@ fn run_scripted_turn_with_deadline(
     // Held across the whole turn: the engine builds and refreshes its system
     // prompt on this thread's current-thread runtime.
     let _environment = crate::prompts::pin_recorded_environment(&os, &shell);
+    let _host_tools = crate::dependencies::pin_recorded_host_tools(recorded_host_tools(case));
     let record = runtime.block_on(async {
         let config = Config::default();
         let engine_config = EngineConfig {
@@ -355,6 +356,23 @@ fn recorded_environment(case: &Value) -> (String, String) {
     (field("os"), field("shell"))
 }
 
+/// The optional host-backed tools the recording machine had (see
+/// `dependencies::host_tool_available`). They change the registry the golden
+/// snapshot counts; a case without the list fails loud.
+fn recorded_host_tools(case: &Value) -> Vec<String> {
+    case.get("recorded_platform")
+        .and_then(|platform| platform.get("host_tools"))
+        .and_then(Value::as_array)
+        .expect("recorded_platform.host_tools must list the recording host's optional tools")
+        .iter()
+        .map(|tool| {
+            tool.as_str()
+                .expect("recorded_platform.host_tools entries are tool names")
+                .to_string()
+        })
+        .collect()
+}
+
 /// Run one turn and return every engine event up to and including
 /// `TurnComplete`, then everything the engine emits while it shuts down.
 async fn drive_turn(

crates/tui/src/core/engine/tool_catalog.rs

@@ -310,7 +310,9 @@ pub(crate) fn ensure_advanced_tooling(
     // `crate::dependencies::resolve_python_interpreter` for the probe.
     if mode != AppMode::Plan
         && !catalog.iter().any(|t| t.name == CODE_EXECUTION_TOOL_NAME)
-        && crate::dependencies::resolve_python_interpreter().is_some()
+        && crate::dependencies::host_tool_available(CODE_EXECUTION_TOOL_NAME, || {
+            crate::dependencies::resolve_python_interpreter().is_some()
+        })
     {
         catalog.push(Tool {
             tool_type: Some(CODE_EXECUTION_TOOL_TYPE.to_string()),
@@ -341,7 +343,9 @@ pub(crate) fn ensure_advanced_tooling(
     // the tool only when `resolve_node()` succeeds.
     if mode != AppMode::Plan
         && !catalog.iter().any(|t| t.name == JS_EXECUTION_TOOL_NAME)
-        && crate::dependencies::resolve_node().is_some()
+        && crate::dependencies::host_tool_available(JS_EXECUTION_TOOL_NAME, || {
+            crate::dependencies::resolve_node().is_some()
+        })
     {
         let mut tool = crate::tools::js_execution::js_execution_tool_definition();
         tool.defer_loading = Some(should_default_defer_tool(&tool.name, always_load));

crates/tui/src/dependencies.rs

@@ -269,6 +269,46 @@ pub fn resolve_pandoc() -> Option<String> {
         .clone()
 }
 
+/// Whether an optional tool whose backend lives on this host (an interpreter,
+/// a converter, an OCR engine) is available: `probe` decides, except that a
+/// conformance replay answers with the recorded host's set so goldens do not
+/// depend on what the machine running them has installed (test builds only).
+pub(crate) fn host_tool_available(tool: &str, probe: impl FnOnce() -> bool) -> bool {
+    #[cfg(all(test, unix))]
+    if let Some(available) = RECORDED_HOST_TOOLS.with(|cell| {
+        cell.borrow()
+            .as_ref()
+            .map(|tools| tools.iter().any(|name| name == tool))
+    }) {
+        return available;
+    }
+    probe()
+}
+
+#[cfg(all(test, unix))]
+thread_local! {
+    static RECORDED_HOST_TOOLS: std::cell::RefCell<Option<Vec<String>>> =
+        const { std::cell::RefCell::new(None) };
+}
+
+/// Pin [`host_tool_available`] on this thread to a recorded host's tools until
+/// the guard drops.
+#[cfg(all(test, unix))]
+pub(crate) fn pin_recorded_host_tools(tools: Vec<String>) -> RecordedHostToolsGuard {
+    RECORDED_HOST_TOOLS.with(|cell| *cell.borrow_mut() = Some(tools));
+    RecordedHostToolsGuard
+}
+
+#[cfg(all(test, unix))]
+pub(crate) struct RecordedHostToolsGuard;
+
+#[cfg(all(test, unix))]
+impl Drop for RecordedHostToolsGuard {
+    fn drop(&mut self) {
+        RECORDED_HOST_TOOLS.with(|cell| *cell.borrow_mut() = None);
+    }
+}
+
 /// Resolve the Node.js runtime once per process. Used by the
 /// `js_execution` tool to decide whether to advertise itself in
 /// the catalog. Unlike Python, the executable name `node` is the

crates/tui/src/tools/registry.rs

@@ -932,7 +932,9 @@ impl ToolRegistryBuilder {
     /// sees a binary it can't actually use.
     #[must_use]
     pub fn with_pandoc_tools(self) -> Self {
-        if crate::dependencies::resolve_pandoc().is_some() {
+        if crate::dependencies::host_tool_available("pandoc_convert", || {
+            crate::dependencies::resolve_pandoc().is_some()
+        }) {
             use super::pandoc::PandocConvertTool;
             self.with_tool(Arc::new(PandocConvertTool))
         } else {
@@ -945,7 +947,7 @@ impl ToolRegistryBuilder {
     /// Tesseract when installed.
     #[must_use]
     pub fn with_image_ocr_tools(self) -> Self {
-        if super::image_ocr::ocr_available() {
+        if crate::dependencies::host_tool_available("image_ocr", super::image_ocr::ocr_available) {
             use super::image_ocr::ImageOcrTool;
             self.with_tool(Arc::new(ImageOcrTool))
         } else {

crates/tui/tests/fixtures/conformance/README.md

@@ -41,7 +41,10 @@ runner for a fact about the machine, not the Engine. A case that omits the
 field fails loud. The same record carries `os` and `shell` (`macos`, `/bin/bash`: the shell
 dispatcher names the recording process's `$SHELL` path verbatim),
 which the `## Environment` block renders into the frozen prompt prefix and so
-into `prefix_cache_change`'s prefix hash; the harness replays those too.
+into `prefix_cache_change`'s prefix hash; the harness replays those too, and
+`host_tools`: the optional tools whose backend the recording machine had
+(`code_execution`, `js_execution`, `pandoc_convert`, `image_ocr`), which
+change the registry the `tool_request_snapshot` counts.
 Production always probes the host. The per-platform label
 text is owned and tested in `sandbox::policy`. All current cases were
 recorded on macOS. A Linux- or Windows-recorded golden is separate evidence.

crates/tui/tests/fixtures/conformance/events/approval_denied.case.json

@@ -9,7 +9,13 @@
     "os": "macos",
     "shell": "/bin/bash",
     "sandbox_enforcement": "local_os",
-    "no_new_privs_active": null
+    "no_new_privs_active": null,
+    "host_tools": [
+      "code_execution",
+      "image_ocr",
+      "js_execution",
+      "pandoc_convert"
+    ]
   },
   "workspace_files": {
     "README.md": "conformance fixture\n"

crates/tui/tests/fixtures/conformance/events/cancel_midstream.case.json

@@ -9,7 +9,13 @@
     "os": "macos",
     "shell": "/bin/bash",
     "sandbox_enforcement": "local_os",
-    "no_new_privs_active": null
+    "no_new_privs_active": null,
+    "host_tools": [
+      "code_execution",
+      "image_ocr",
+      "js_execution",
+      "pandoc_convert"
+    ]
   },
   "workspace_files": {
     "README.md": "conformance fixture\n",

crates/tui/tests/fixtures/conformance/events/one_tool_call.case.json

@@ -9,7 +9,13 @@
     "os": "macos",
     "shell": "/bin/bash",
     "sandbox_enforcement": "local_os",
-    "no_new_privs_active": null
+    "no_new_privs_active": null,
+    "host_tools": [
+      "code_execution",
+      "image_ocr",
+      "js_execution",
+      "pandoc_convert"
+    ]
   },
   "workspace_files": {
     "README.md": "conformance fixture\n",

crates/tui/tests/fixtures/conformance/events/parallel_tool_calls.case.json

@@ -9,7 +9,13 @@
     "os": "macos",
     "shell": "/bin/bash",
     "sandbox_enforcement": "local_os",
-    "no_new_privs_active": null
+    "no_new_privs_active": null,
+    "host_tools": [
+      "code_execution",
+      "image_ocr",
+      "js_execution",
+      "pandoc_convert"
+    ]
   },
   "workspace_files": {
     "README.md": "conformance fixture\n",

crates/tui/tests/fixtures/conformance/events/plain_answer.case.json

@@ -9,7 +9,13 @@
     "os": "macos",
     "shell": "/bin/bash",
     "sandbox_enforcement": "local_os",
-    "no_new_privs_active": null
+    "no_new_privs_active": null,
+    "host_tools": [
+      "code_execution",
+      "image_ocr",
+      "js_execution",
+      "pandoc_convert"
+    ]
   },
   "workspace_files": {
     "README.md": "conformance fixture\n",

crates/tui/tests/fixtures/conformance/events/provider_error_after_tool_call.case.json

@@ -9,7 +9,13 @@
     "os": "macos",
     "shell": "/bin/bash",
     "sandbox_enforcement": "local_os",
-    "no_new_privs_active": null
+    "no_new_privs_active": null,
+    "host_tools": [
+      "code_execution",
+      "image_ocr",
+      "js_execution",
+      "pandoc_convert"
+    ]
   },
   "workspace_files": {
     "README.md": "conformance fixture\n",

crates/tui/tests/fixtures/conformance/prompt/agent_suggest.case.json

@@ -9,7 +9,13 @@
     "os": "macos",
     "shell": "/bin/bash",
     "sandbox_enforcement": "local_os",
-    "no_new_privs_active": null
+    "no_new_privs_active": null,
+    "host_tools": [
+      "code_execution",
+      "image_ocr",
+      "js_execution",
+      "pandoc_convert"
+    ]
   },
   "user_message": "hello",
   "workspace_files": {

crates/tui/tests/fixtures/conformance/prompt/plan_mode.case.json

@@ -9,7 +9,13 @@
     "os": "macos",
     "shell": "/bin/bash",
     "sandbox_enforcement": "local_os",
-    "no_new_privs_active": null
+    "no_new_privs_active": null,
+    "host_tools": [
+      "code_execution",
+      "image_ocr",
+      "js_execution",
+      "pandoc_convert"
+    ]
   },
   "user_message": "make a plan",
   "workspace_files": {

crates/tui/tests/fixtures/conformance/prompt/trusted_project_instructions.case.json

@@ -9,7 +9,13 @@
     "os": "macos",
     "shell": "/bin/bash",
     "sandbox_enforcement": "local_os",
-    "no_new_privs_active": null
+    "no_new_privs_active": null,
+    "host_tools": [
+      "code_execution",
+      "image_ocr",
+      "js_execution",
+      "pandoc_convert"
+    ]
   },
   "trusted_workspace": true,
   "user_message": "hello",