fix(providers): retry the local probe after a startup miss

src/providers/local.c

@@ -1,6 +1,7 @@
 /**
  * @file local.c
  * @brief OpenAI-compatible Chat Completions client for llama-server; init probe marks unreachable without failing startup.
+ * A later chat probes again so a server that was down only at startup can still serve.
  */
 
 #define _POSIX_C_SOURCE 200809L
@@ -184,6 +185,14 @@ static int local_chat(const provider_message_t *messages, size_t message_count,
 		provider_set_error(response, "Local provider not initialized");
 		return -1;
 	}
+	/* Init probes once, often before llama-server has loaded the model.
+	 * Default fallback order puts local last, so router recovery (which
+	 * only probes backends before the active index) never clears the flag. */
+	if (s_local_unavailable) {
+		const char *endpoint = config_provider_local_endpoint(s_local_cfg);
+		if (endpoint && endpoint[0] != '\0' && provider_local_endpoint_reachable(endpoint))
+			provider_local_recovery_probe_succeeded();
+	}
 	if (s_local_unavailable) {
 		provider_set_error(response, "Local inference server unreachable");
 		return -1;

src/providers/provider.h

@@ -134,7 +134,7 @@ void provider_openai_set_live_config(const config_t *cfg);
 const provider_t *provider_local_get(void);
 /** After SIGHUP reload: refresh live config reads without re-init (PRD §9 Q4). */
 void provider_local_set_live_config(const config_t *cfg);
-/** Lightweight GET probe via `/v1/models` or `/health`; used by init and router recovery. */
+/** Lightweight GET probe via `/v1/models` or `/health`; used by init, chat retry, and router recovery. */
 int provider_local_endpoint_reachable(const char *endpoint);
 /** After a successful recovery GET probe on local; clears init-time unreachable flag. */
 void provider_local_recovery_probe_succeeded(void);

tests/test_local_provider.c

@@ -75,6 +75,52 @@ static int test_init_unreachable_sets_unavailable_then_chat_errors(void)
 	return 0;
 }
 
+/**
+ * Init probe failure must not stick for the process lifetime. llama-server
+ * often is not listening yet at startup; the next chat has to probe again.
+ */
+static int test_chat_reprobes_when_server_comes_up(void)
+{
+	FILE *f = fopen(TMP_CFG, "w");
+	config_t *cfg = NULL;
+	char errbuf[256];
+	const provider_t *p;
+	provider_message_t msg = { .role = "user", .content = "ping", .tool_calls = NULL,
+	                           .tool_calls_count = 0, .tool_use_id = NULL };
+	provider_response_t response = {0};
+	ASSERT(f);
+	local_provider_test_reset();
+	unsetenv("SHELLCLAW_FALLBACK_CHAIN");
+	unsetenv("SHELLCLAW_LOCAL_ENDPOINT");
+	unsetenv("SHELLCLAW_LOCAL_MODEL");
+	fprintf(f, "[agent]\nmodel = \"test-model\"\n");
+	fprintf(f, "[providers.local]\n");
+	fprintf(f, "endpoint = \"http://127.0.0.1:1/v1/chat/completions\"\nmodel = \"tiny\"\n");
+	fclose(f);
+	ASSERT(config_load(TMP_CFG, &cfg, errbuf, sizeof(errbuf)) == 0);
+	p = provider_local_get();
+	ASSERT(p->init(cfg) == 0);
+#ifdef SHELLCLAW_TEST
+	ASSERT(local_provider_is_unavailable_for_test() != 0);
+#endif
+	local_provider_test_set_skip_probe(1);
+	local_provider_test_set_http_response(
+		"{\"choices\":[{\"message\":{\"role\":\"assistant\",\"content\":\"back online\"}}]}");
+	ASSERT(p->chat(&msg, 1, NULL, 0, &response) == 0);
+	ASSERT(response.error == 0);
+	ASSERT(response.content != NULL);
+	ASSERT(strstr(response.content, "back online") != NULL);
+#ifdef SHELLCLAW_TEST
+	ASSERT(local_provider_is_unavailable_for_test() == 0);
+#endif
+	p->cleanup();
+	provider_response_clear(&response);
+	local_provider_test_reset();
+	config_free(cfg);
+	remove(TMP_CFG);
+	return 0;
+}
+
 static int test_chat_mock_http_happy_path(void)
 {
 	FILE *f = fopen(TMP_CFG, "w");
@@ -113,6 +159,7 @@ int main(void)
 	RUN(test_local_provider_vtable());
 	RUN(test_init_requires_config());
 	RUN(test_init_unreachable_sets_unavailable_then_chat_errors());
+	RUN(test_chat_reprobes_when_server_comes_up());
 	RUN(test_chat_mock_http_happy_path());
 	curl_global_cleanup();
 	printf("test_local_provider: all tests passed\n");