ci: stabilize live harness shards and MCP App E2E policy

.config/nextest.toml

@@ -20,14 +20,50 @@ platform = 'cfg(windows)'
 filter = 'package(labby-codemode) & test(git::provider::tests::)'
 threads-required = 'num-test-threads'
 
-# Shared lifecycle harness tests are compiled into several integration binaries.
-# Windows CI observed four concurrent instances exhausting their unchanged 45s
-# readiness deadline during database initialization; later instances finished in
-# 2-7s. Reserve slots for these cases to avoid competing daemon startups. Their
-# own multi-process scenarios still run concurrently inside the test.
+# Shared lifecycle harness tests are compiled into many integration binaries.
+# They launch real daemons, process groups, listeners, and cleanup guardians.
+# Windows first exposed concurrent copies exhausting readiness budgets; hosted
+# Linux shards and macOS reproduction showed the same interference class.
+# Reserve every nextest slot for each live_labby case on every platform while
+# preserving each case's intended internal multi-process concurrency.
 [[profile.ci.overrides]]
-platform = 'cfg(windows)'
-filter = 'package(labby) & test(live_labby::tests::)'
+filter = 'package(labby) & test(live_labby::)'
+threads-required = 'num-test-threads'
+
+# Integration binaries that launch real Labby daemons, browser/process fixtures,
+# or the shared lifecycle harness must not overlap on one runner. Their product
+# deadlines are intentional oracles; runner contention must not consume them.
+# Reserve every nextest slot for each test in these binaries instead of inflating
+# deadlines. The target shards still execute on separate CI runners in parallel.
+[[profile.ci.overrides]]
+filter = '''package(labby) & (
+    binary(=browser_runtime_e2e)
+    | binary(=codemode_qualification)
+    | binary(=e2e_coverage_report)
+    | binary(=gateway_auto_reconnect)
+    | binary(=lifecycle_conformance)
+    | binary(=live_api_actions)
+    | binary(=live_browser_bridge)
+    | binary(=live_cli_actions)
+    | binary(=live_http_ipv6)
+    | binary(=live_http_observability)
+    | binary(=live_http_routes)
+    | binary(=live_integration_identity)
+    | binary(=live_mcp_actions)
+    | binary(=live_process_harness)
+    | binary(=live_protected_routes)
+    | binary(=live_restart_persistence)
+    | binary(=live_surface_parity)
+    | binary(=mcp_apps_host_qualification)
+    | binary(=mcp_primitives_qualification)
+    | binary(=mcp_spec_http_contracts)
+    | binary(=mcp_spec_wire_compliance)
+    | binary(=mcp_tools_transport_qualification)
+    | binary(=oauth_qualification)
+    | binary(=proxy_qualification)
+    | binary(=skills_mcp_e2e)
+    | binary(=webmcp_browser_qualification)
+)'''
 threads-required = 'num-test-threads'
 
 # These three process-inventory controls hit their unchanged 3s deadlines when

crates/labby-runtime/src/agent_runtime.rs

@@ -34,7 +34,7 @@ pub struct AgentResourceBounds {
 impl AgentResourceBounds {
     pub fn validate(self) -> Result<Self, AgentRuntimeError> {
         if self.max_runtime_millis == 0
-            || self.max_runtime_millis > 86_400_000
+            || self.max_runtime_millis > AGENT_MAX_RUNTIME_MILLIS
             || self.max_output_bytes == 0
             || self.max_output_bytes > 64 * 1024 * 1024
             || self.max_external_effects > 10_000
@@ -619,6 +619,25 @@ mod tests {
         );
     }
 
+    #[test]
+    fn resource_bounds_cannot_outlive_the_authority_lease_contract() {
+        let valid = AgentResourceBounds {
+            max_runtime_millis: AGENT_MAX_RUNTIME_MILLIS,
+            max_output_bytes: 1,
+            max_external_effects: 0,
+        };
+        assert_eq!(valid.validate().unwrap(), valid);
+
+        let overlong = AgentResourceBounds {
+            max_runtime_millis: AGENT_MAX_RUNTIME_MILLIS + 1,
+            ..valid
+        };
+        assert_eq!(
+            overlong.validate().unwrap_err(),
+            AgentRuntimeError::InvalidBounds
+        );
+    }
+
     #[tokio::test]
     async fn runtime_bound_cancels_an_overrunning_executor_and_requests_cleanup() {
         let initial = epochs(1);

crates/labby-runtime/src/gateway_config.rs

@@ -111,6 +111,7 @@ fn default_mcp_scopes() -> Vec<String> {
 pub struct McpAppsConfig {
     /// Attach MCP App metadata to the always-available `mcp_app` control tool and advertise its UI resources.
     /// The control tool itself remains available when this is false.
+    /// Fresh installs expose the manager UI by default.
     #[serde(default = "default_true")]
     pub manager: bool,
     /// Advertise the synthetic Add Server app tool and its UI resources.

crates/labby/src/mcp/handlers_resources.rs

@@ -1584,8 +1584,9 @@ impl LabMcpServer {
         // Branch 0: MCP Apps UI resources. This must precede all lab://
         // fallbacks so ui:// has its own exact lookup semantics.
         //
-        // The `mcp_app` control tool is always locally available, but its own
-        // Labby-owned UI is opt-in like every other Labby-owned MCP App.
+        // The `mcp_app` control tool is always locally available. Its UI follows
+        // `mcp_apps.manager`: fresh installs enable it, while an operator can
+        // disable the resource surface without removing the control tool.
         #[cfg(feature = "gateway")]
         if uri.starts_with(MCP_APPS_APP_URI) {
             if !self.mcp_apps_config().await.manager {

crates/labby/tests/action_matrix_completeness.rs

@@ -1,22 +1,26 @@
 #![allow(clippy::panic)]
 
-#[path = "support/lib.rs"]
-mod support;
+#[allow(dead_code)]
+#[path = "support/action_matrix.rs"]
+mod action_matrix;
+#[allow(dead_code)]
+#[path = "support/authority_matrix.rs"]
+mod authority_matrix;
 
 use std::collections::{BTreeMap, BTreeSet};
 
-use labby_primitives::access::{Capability, CapabilitySchemaVersion, RoleTemplate};
-use serde_json::Value;
-use support::action_matrix::{
+use action_matrix::{
     CatalogAction, EXPECTED_ACTIONS, EXPECTED_API_ACTIONS, EXPECTED_CLI_ACTIONS,
     EXPECTED_MCP_ACTIONS, EXPECTED_SHARED_CLI_MCP_API_ACTIONS, EXPECTED_WEB_ACTIONS, EvidenceLevel,
     PersistenceClass, ScenarioKind, ScenarioOwner, Surface, catalog_map, intent_map,
     intent_map_from, intents, validate_intent_shape,
 };
-use support::authority_matrix::{
+use authority_matrix::{
     DEPOT_OPERATIONS, OperationClass, OwnerKind, ResourceFamily, classify_labby,
     depot_fixture_operation_names, depot_snapshot_operation_names, duplicate_depot_operations,
 };
+use labby_primitives::access::{Capability, CapabilitySchemaVersion, RoleTemplate};
+use serde_json::Value;
 
 const ACTION_CATALOG: &str = include_str!("../../../docs/generated/action-catalog.json");
 const AUTHORITY_MATRIX: &str =
@@ -994,7 +998,7 @@ services = ["stash"]
 
 #[test]
 fn generated_outcomes_cannot_downgrade_required_intent() {
-    use support::action_matrix::{CaseOutcome, OutcomeStatus, outcome_satisfies};
+    use action_matrix::{CaseOutcome, OutcomeStatus, outcome_satisfies};
     let intent = intents().iter().find(|intent| intent.required).unwrap();
     let skipped = CaseOutcome {
         key: intent.key(),

crates/labby/tests/ci_changed_paths.rs

@@ -412,6 +412,15 @@ fn rust_manifests_lockfiles_and_toolchains_run_full_tests() {
     }
 }
 
+#[test]
+fn nextest_policy_changes_run_the_full_rust_test_path() {
+    let out = classify("pull_request", &[".config/nextest.toml"]);
+    assert_eq!(out["rust_compile"], "true");
+    assert_eq!(out["rust_test"], "true");
+    assert_eq!(out["docker"], "true");
+    assert_eq!(out["release"], "true");
+}
+
 #[test]
 fn frontend_changes_enable_web_release_and_container_without_rust_tests() {
     let out = classify("pull_request", &["apps/gateway-admin/app/page.tsx"]);
@@ -1210,6 +1219,74 @@ fn ci_gate_aggregates_every_non_advisory_job() {
     }
 }
 
+/// Process-owning integration fixtures are compiled into many test binaries.
+/// The CI profile must serialize both duplicated harness-unit cases and every
+/// integration binary that launches real process state.
+#[test]
+fn nextest_ci_isolates_process_harnesses_and_timing_oracles() {
+    let root = repo_root();
+    let nextest =
+        fs::read_to_string(root.join(".config/nextest.toml")).expect("read .config/nextest.toml");
+    let lifecycle_filter = "filter = 'package(labby) & test(live_labby::)'";
+    assert!(
+        nextest.contains(lifecycle_filter),
+        "CI must serialize every embedded live_labby harness copy"
+    );
+    let lifecycle = nextest
+        .split(lifecycle_filter)
+        .nth(1)
+        .and_then(|section| section.split("[[").next())
+        .expect("live_labby override");
+    assert!(
+        lifecycle.contains("threads-required = 'num-test-threads'"),
+        "live_labby override must reserve every nextest slot"
+    );
+    assert!(
+        !lifecycle.contains("platform ="),
+        "process-harness isolation must apply on Linux, macOS, and Windows"
+    );
+
+    let process_override = nextest
+        .split("filter = '''package(labby) & (")
+        .nth(1)
+        .and_then(|section| section.split("[[").next())
+        .expect("process-owning integration override");
+    assert!(
+        process_override.contains("threads-required = 'num-test-threads'"),
+        "process-owning integration tests must reserve every nextest slot"
+    );
+    let test_dir = root.join("crates/labby/tests");
+    let mut process_binaries = Vec::new();
+    for entry in fs::read_dir(&test_dir).expect("read integration test directory") {
+        let path = entry.expect("test entry").path();
+        if path.extension().and_then(|value| value.to_str()) != Some("rs") {
+            continue;
+        }
+        let source = fs::read_to_string(&path).expect("read integration test source");
+        let path_marker = ["support", "/live_labby.rs"].concat();
+        let type_marker = ["support::", "LiveLabby"].concat();
+        let module_marker = ["support::", "live_labby"].concat();
+        if source.contains(&path_marker)
+            || source.contains(&type_marker)
+            || source.contains(&module_marker)
+        {
+            process_binaries.push(
+                path.file_stem()
+                    .and_then(|value| value.to_str())
+                    .expect("UTF-8 test name")
+                    .to_owned(),
+            );
+        }
+    }
+    assert!(!process_binaries.is_empty());
+    for binary in process_binaries {
+        assert!(
+            process_override.contains(&format!("binary(={binary})")),
+            "process-owning integration binary {binary} must be isolated in nextest CI"
+        );
+    }
+}
+
 /// The merge gate must finish in about ten minutes. That budget is kept by
 /// fanning the long serial suites out across matrix shards, by moving the
 /// slow non-gating suites (coverage, the gateway-slice re-run, doctests) off
@@ -1261,6 +1338,7 @@ fn merge_gate_shards_heavy_suites_to_stay_under_ten_minutes() {
         "--partition \"hash:${index}/${unit_shards}\"",
         "for file in crates/labby/tests/*.rs",
         "cargo nextest run -p labby",
+        "not test(live_labby::) | binary(=live_process_harness)",
         "--exclude labby",
         "cargo test --doc --workspace --all-features --locked",
     ] {

crates/labby/tests/mcp_apps_host_qualification.rs

@@ -157,9 +157,23 @@ async fn q4_real_resources_render_in_distinct_openai_and_anthropic_emulators() {
             .await
             .expect("real Labby MCP process");
 
-    let initially_hidden = runner.read_resource("ui://lab/apps/manage").await;
+    // Fresh installs intentionally expose Labby-owned MCP Apps by default.
+    // Exercise the policy transition explicitly so this live oracle verifies
+    // that direct resources/read cannot bypass a disabled surface.
+    let initially_disabled = runner
+        .call_raw(
+            "mcp_app",
+            serde_json::json!({"action":"disable", "params":{"target":"manager"}}),
+        )
+        .await
+        .expect("disable manager through real tools/call");
+    assert_ne!(
+        initially_disabled.is_error,
+        Some(true),
+        "initial disable failed: {initially_disabled:?}"
+    );
     assert!(
-        initially_hidden.is_err(),
+        runner.read_resource("ui://lab/apps/manage").await.is_err(),
         "disabled app resource must not bypass policy"
     );
 
@@ -263,10 +277,12 @@ async fn q4_real_resources_render_in_distinct_openai_and_anthropic_emulators() {
         Some(true),
         "disable failed: {disabled:?}"
     );
-    assert!(
-        runner.read_resource(mcp_uri).await.is_err(),
-        "revoked resource remained readable on the same authenticated session"
-    );
+    for uri in [mcp_uri, openai_uri] {
+        assert!(
+            runner.read_resource(uri).await.is_err(),
+            "revoked resource {uri} remained readable on the same authenticated session"
+        );
+    }
 
     let cleanup = runner.finish().await;
     assert!(

crates/labby/tests/support/mcp_apps_host_qualification/runner.mjs

@@ -121,21 +121,38 @@ function listen() {
 		server.listen(0, "127.0.0.1", () => ok(server.address().port));
 	});
 }
-async function wait(p, s) {
+async function waitManagerSwitch(p, checked) {
+	const expected = String(checked);
 	try {
 		await p
-			.locator("#status")
-			.filter({ hasText: s })
+			.locator(
+				`button[data-target=manager][aria-checked="${expected}"]:not([disabled])`,
+			)
 			.waitFor({ timeout: 5000 });
 	} catch (e) {
+		const button = p.locator("button[data-target=manager]");
 		throw new Error(
-			`${e.message}; actual status=${await p
+			`${e.message}; actual checked=${await button
+				.getAttribute("aria-checked")
+				.catch(() => "missing")}; summary=${await p
+				.locator("#summary")
+				.textContent()
+				.catch(() => "missing")}; status=${await p
 				.locator("#status")
 				.textContent()
 				.catch(() => "missing")}`,
 		);
 	}
 }
+async function setManager(enabled) {
+	await call({
+		name: "mcp_app",
+		arguments: {
+			action: enabled ? "enable" : "disable",
+			params: { target: "manager" },
+		},
+	});
+}
 async function csp(p, port) {
 	const got = await p.evaluate(async (port) => {
 		const violations = [];
@@ -176,16 +193,10 @@ async function csp(p, port) {
 	return got;
 }
 async function exercise(host, p, port) {
-	await wait(p, "up to date");
-	const policy = await csp(p, port),
-		on = await p
-			.locator("button[data-target=manager]")
-			.getAttribute("aria-checked");
+	await waitManagerSwitch(p, true);
+	const policy = await csp(p, port);
 	await p.locator("button[data-target=manager]").click();
-	await wait(
-		p,
-		on === "true" ? "Disabled MCP Apps Manager" : "Enabled MCP Apps Manager",
-	);
+	await waitManagerSwitch(p, false);
 	const auth = await p.evaluate(() =>
 		invalidMcpCall({
 			name: "mcp_app",
@@ -214,6 +225,7 @@ try {
 	await ctx.exposeFunction("invalidMcpCall", (x) =>
 		call(x, "invalid-q4-token"),
 	);
+	await setManager(true);
 	const o = await ctx.newPage();
 	await o.addInitScript(
 		() =>
@@ -242,6 +254,7 @@ try {
 		message: "injected OpenAI transport failure",
 	});
 	events.find((event) => event.host === "openai-emulator").type_error_calls = 1;
+	await setManager(true);
 	const h = await ctx.newPage();
 	await h.goto(`http://127.0.0.1:${port}/anthropic`);
 	await h.waitForTimeout(250);

docs/dev/TESTING.md

@@ -252,6 +252,15 @@ Preferred runner:
 - use `cargo test` only when nextest is unavailable or you need a narrow one-off command that nextest does not cover cleanly
 - for this repo, `cargo nextest run --manifest-path crates/labby/Cargo.toml --all-features` is the standard full-crate verification command
 
+CI splits `crates/labby/tests/*.rs` across `labby-int-*` target shards. Many of
+those binaries embed the same `live_labby` process-supervision harness. Shards
+de-duplicate those internal support tests and execute them once through the
+`live_process_harness` target. The CI nextest profile also reserves all local
+test slots for process-owning integration binaries, so separate daemon/browser
+fixtures cannot consume one another's readiness, cleanup, or timeout budgets.
+This preserves the product deadlines the tests prove without trading flakiness
+for inflated timeouts.
+
 If tests were not run, say so explicitly.
 
 ## Command Guidance