sovereign_browser/src/agent_loop.c

/*
 * agent_loop.c — ReAct tool-call loop on a background GThread
 *
 * Implements the standard ReAct loop:
 *   1. Load chat history from agent_chat_store_get_messages().
 *   2. Prepend a system prompt message (must be first in the array).
 *   3. Call the LLM via agent_llm_chat().
 *   4. If the response contains tool_calls, dispatch each:
 *        - Browser tools → hop to the GTK main thread via g_idle_add() +
 *          GAsyncQueue (WebKitGTK is not thread-safe).
 *        - Filesystem/shell tools → run directly on this thread via
 *          agent_fs_tools_dispatch().
 *   5. Append tool results to the chat store.
 *   6. Repeat until no more tool_calls, the iteration cap is reached, or
 *      the cancel flag is set.
 *   7. Persist the final assistant message and update session status.
 *
 * See plans/embedded-agent.md for the full concurrency model.
 */

#include "agent_loop.h"
#include "agent_llm.h"
#include "agent_chat_store.h"
#include "agent_fs_tools.h"
#include "agent_tools.h"
#include "agent_skills.h"
#include "agent_server.h"
#include "settings.h"
#include "cjson/cJSON.h"

#include <glib.h>
#include <string.h>

/* ── Context ────────────────────────────────────────────────────────── */

typedef struct {
	/* Atomic flags */
	volatile gint cancel_flag;
	volatile gint running;       /* gboolean as gint */

	/* Status (protected by status_lock) */
	GMutex status_lock;
	agent_loop_state_t state;
	int iteration;
	char *current_tool;    /* tool name being executed */
	char *last_message;    /* last assistant text */
	char *error;           /* error message */

	/* Thread handle */
	GThread *thread;
} agent_loop_ctx_t;

static agent_loop_ctx_t g_ctx = {0};

/* ── Status helpers ─────────────────────────────────────────────────── */

/* Map a state enum to its string name for JSON events. */
static const char *
state_name(agent_loop_state_t s)
{
	switch (s) {
		case AGENT_LOOP_IDLE:      return "idle";
		case AGENT_LOOP_THINKING:  return "thinking";
		case AGENT_LOOP_TOOL_CALL: return "tool_call";
		case AGENT_LOOP_COMPLETE:  return "complete";
		case AGENT_LOOP_ERROR:     return "error";
		case AGENT_LOOP_CANCELLED: return "cancelled";
	}
	return "idle";
}

/* Emit the current status as a WebSocket push event so the chat UI
 * can update instantly without polling. Must be called WITHOUT the
 * mutex held (it takes it itself). */
static void
emit_status_event(void)
{
	g_mutex_lock(&g_ctx.status_lock);
	cJSON *data = cJSON_CreateObject();
	cJSON_AddStringToObject(data, "state", state_name(g_ctx.state));
	cJSON_AddNumberToObject(data, "iteration", g_ctx.iteration);
	if (g_ctx.current_tool)
		cJSON_AddStringToObject(data, "current_tool", g_ctx.current_tool);
	else
		cJSON_AddNullToObject(data, "current_tool");
	if (g_ctx.last_message)
		cJSON_AddStringToObject(data, "last_message", g_ctx.last_message);
	else
		cJSON_AddNullToObject(data, "last_message");
	if (g_ctx.error)
		cJSON_AddStringToObject(data, "error", g_ctx.error);
	else
		cJSON_AddNullToObject(data, "error");
	g_mutex_unlock(&g_ctx.status_lock);

	/* agent_server_emit_event takes ownership of data. */
	agent_server_emit_event("agent_status", data);
}

static void
set_status(agent_loop_state_t state, int iter,
           const char *tool, const char *msg, const char *err)
{
	g_mutex_lock(&g_ctx.status_lock);
	g_ctx.state = state;
	g_ctx.iteration = iter;
	if (tool) {
		g_free(g_ctx.current_tool);
		g_ctx.current_tool = g_strdup(tool);
	} else {
		/* Clear tool name when not in a tool-call state. */
		g_free(g_ctx.current_tool);
		g_ctx.current_tool = NULL;
	}
	if (msg) {
		g_free(g_ctx.last_message);
		g_ctx.last_message = g_strdup(msg);
	}
	/* Only set the error field when err is non-NULL. When err is NULL
	 * and we're transitioning to a non-error state, clear any stale
	 * error so the UI doesn't show an old error message. */
	if (err) {
		g_free(g_ctx.error);
		g_ctx.error = g_strdup(err);
	} else if (state != AGENT_LOOP_ERROR) {
		g_free(g_ctx.error);
		g_ctx.error = NULL;
	}
	g_mutex_unlock(&g_ctx.status_lock);
	emit_status_event();
}

static void
set_state(agent_loop_state_t state)
{
	g_mutex_lock(&g_ctx.status_lock);
	g_ctx.state = state;
	/* Clear stale error when transitioning to a non-error state. */
	if (state != AGENT_LOOP_ERROR) {
		g_free(g_ctx.error);
		g_ctx.error = NULL;
	}
	/* Clear tool name when leaving the tool_call state. */
	if (state != AGENT_LOOP_TOOL_CALL) {
		g_free(g_ctx.current_tool);
		g_ctx.current_tool = NULL;
	}
	g_mutex_unlock(&g_ctx.status_lock);
	emit_status_event();
}

static void
set_error(const char *msg)
{
	g_mutex_lock(&g_ctx.status_lock);
	g_ctx.state = AGENT_LOOP_ERROR;
	g_free(g_ctx.error);
	g_ctx.error = g_strdup(msg);
	g_mutex_unlock(&g_ctx.status_lock);
	/* Emit a status event so the UI learns about the error immediately.
	 * Without this, the chat page stays stuck on "Thinking..." because
	 * the WebSocket push never fires and the polling fallback may have
	 * been stopped when the WebSocket connected. */
	emit_status_event();
}

/* ── Build messages array with system prompt first ──────────────────── */

/*
 * Build the messages array for the LLM: a system message first, then
 * all messages from the chat store. Returns a newly-allocated cJSON
 * array (caller frees with cJSON_Delete), or NULL on error.
 */
static cJSON *
build_messages_with_system(const char *system_prompt)
{
	cJSON *history = agent_chat_store_get_messages();
	if (history == NULL)
		return NULL;

	cJSON *messages = cJSON_CreateArray();
	if (messages == NULL) {
		cJSON_Delete(history);
		return NULL;
	}

	/* System prompt must be the first message */
	cJSON *sys_msg = cJSON_CreateObject();
	cJSON_AddStringToObject(sys_msg, "role", "system");
	cJSON_AddStringToObject(sys_msg, "content", system_prompt);
	cJSON_AddItemToArray(messages, sys_msg);

	/* Copy all history messages into the new array */
	cJSON *msg;
	cJSON_ArrayForEach(msg, history) {
		cJSON_AddItemToArray(messages, cJSON_Duplicate(msg, 1));
	}

	cJSON_Delete(history);
	return messages;
}

/* ── Background thread ──────────────────────────────────────────────── */

static gpointer
agent_loop_thread(gpointer data)
{
	char *user_message = (char *)data;
	g_print("[agent-loop] Thread started, message: %s\n", user_message);

	/* 1. Add user message to chat store */
	agent_chat_store_add_user_message(user_message);
	g_free(user_message);
	/* Notify UI that a message was added. */
	cJSON *umsg = cJSON_CreateObject();
	cJSON_AddStringToObject(umsg, "role", "user");
	agent_server_emit_event("agent_message", umsg);

	/* 2. Get provider settings (copy — they could change while running) */
	const browser_settings_t *s = settings_get();
	char base_url[512], api_key[512], model[128];
	g_strlcpy(base_url, s->agent_llm_base_url, sizeof(base_url));
	g_strlcpy(api_key, s->agent_llm_api_key, sizeof(api_key));
	g_strlcpy(model, s->agent_llm_model, sizeof(model));
	g_print("[agent-loop] base_url=%s model=%s api_key=%s max_iter=%d\n",
	        base_url, model, api_key[0] ? "(set)" : "(empty)", s->agent_max_iterations);

	/* Build the system prompt. agent_skills_build_system_prompt() now
	 * always returns a non-NULL string: it starts with the Sovereign
	 * Browser Skill template (from settings) as the base, then
	 * appends any selected Nostr skills' templates. */
	char *system_prompt = agent_skills_build_system_prompt();
	if (system_prompt == NULL) {
		/* Defensive fallback — should never happen. */
		system_prompt = g_strdup(SETTINGS_AGENT_SYSTEM_PROMPT_DEFAULT);
	}
	g_print("[agent-loop] Using system prompt (%zu bytes)\n",
	        strlen(system_prompt));

	int max_iter = s->agent_max_iterations;
	if (max_iter <= 0)
		max_iter = SETTINGS_AGENT_MAX_ITERATIONS_DEFAULT;

	/* Empty API key → pass NULL (local servers like Ollama don't need one) */
	const char *key_arg = (api_key[0] != '\0') ? api_key : NULL;

	/* 3. Build OpenAI tools array (built once, reused each iteration) */
	cJSON *tools = agent_llm_build_openai_tools();

	/* 4. ReAct loop */
	for (int iter = 0; iter < max_iter; iter++) {
		/* Check cancel */
		if (g_atomic_int_get(&g_ctx.cancel_flag)) {
			set_state(AGENT_LOOP_CANCELLED);
			break;
		}

		/* Update status: thinking */
		set_status(AGENT_LOOP_THINKING, iter, NULL, NULL, NULL);

		/* Build messages: system prompt first, then chat history */
		cJSON *messages = build_messages_with_system(system_prompt);
		if (messages == NULL) {
			set_error("Failed to load messages");
			break;
		}

		/* Call LLM (blocking HTTP on this thread) */
		g_print("[agent-loop] iter %d: calling LLM...\n", iter);
		agent_llm_response_t *resp = agent_llm_chat(base_url, key_arg,
		                                            model, messages, tools);
		cJSON_Delete(messages);

		if (resp == NULL) {
			g_print("[agent-loop] iter %d: LLM call returned NULL\n", iter);
			set_error("LLM API call failed");
			break;
		}
		g_print("[agent-loop] iter %d: LLM responded, finish=%s, content=%s, tool_calls=%d\n",
		        iter, resp->finish_reason ? resp->finish_reason : "(null)",
		        resp->content ? resp->content : "(null)",
		        resp->tool_calls ? cJSON_GetArraySize(resp->tool_calls) : 0);

		/* Persist assistant message */
		char *tool_calls_str = (resp->tool_calls != NULL)
			? cJSON_PrintUnformatted(resp->tool_calls) : NULL;
		agent_chat_store_add_assistant_message(resp->content, tool_calls_str);
		g_free(tool_calls_str);
		/* Notify UI that a message was added. */
		cJSON *amsg = cJSON_CreateObject();
		cJSON_AddStringToObject(amsg, "role", "assistant");
		agent_server_emit_event("agent_message", amsg);

		/* Update last_message status */
		set_status(AGENT_LOOP_THINKING, iter, NULL,
		           resp->content ? resp->content : "", NULL);

		/* If no tool_calls, we're done */
		if (resp->tool_calls == NULL ||
		    cJSON_GetArraySize(resp->tool_calls) == 0) {
			agent_llm_response_free(resp);
			set_state(AGENT_LOOP_COMPLETE);
			goto done;
		}

		/* Dispatch each tool call */
		cJSON *tc;
		cJSON_ArrayForEach(tc, resp->tool_calls) {
			if (g_atomic_int_get(&g_ctx.cancel_flag))
				break;

			cJSON *fn = cJSON_GetObjectItem(tc, "function");
			const char *tool_name = fn
				? cJSON_GetStringValue(cJSON_GetObjectItem(fn, "name"))
				: NULL;
			const char *args_str = fn
				? cJSON_GetStringValue(cJSON_GetObjectItem(fn, "arguments"))
				: NULL;
			const char *tc_id = cJSON_GetStringValue(cJSON_GetObjectItem(tc, "id"));

			cJSON *args = (args_str != NULL)
				? cJSON_Parse(args_str) : cJSON_CreateObject();
			if (args == NULL)
				args = cJSON_CreateObject();

			set_status(AGENT_LOOP_TOOL_CALL, iter,
			           tool_name ? tool_name : "?", NULL, NULL);
			g_print("[agent-loop] iter %d: tool %s (args: %s)\n",
			        iter, tool_name ? tool_name : "?",
			        args_str ? args_str : "(null)");

			/* Dispatch: fs/shell tools and browser tools both run on
			 * this background thread. The browser tools use the same
			 * sync JS eval path (conn=NULL) as the MCP HTTP handler,
			 * which calls gtk_main_iteration() to pump the main loop
			 * while waiting for the async JS result. This is safe to
			 * call from a background thread — the main loop keeps
			 * running and processes the JS eval callback. */
			cJSON *result;
			if (tool_name != NULL && agent_fs_is_tool(tool_name)) {
				result = agent_fs_tools_dispatch(tool_name, args);
			} else {
				/* Build the request JSON and dispatch directly (same
				 * as the MCP HTTP handler does from the libsoup thread). */
				cJSON *request = cJSON_CreateObject();
				cJSON_AddStringToObject(request, "tool",
				                        tool_name ? tool_name : "");
				cJSON_AddItemToObject(request, "params", args);
				args = NULL;  /* transferred to request */
				result = agent_tools_dispatch(request, NULL);
				cJSON_Delete(request);
			}
			if (args) cJSON_Delete(args);

			/* Get result JSON string */
			char *result_str = (result != NULL)
				? cJSON_PrintUnformatted(result) : g_strdup("{}");
			cJSON_Delete(result);

			/* Persist tool result */
			agent_chat_store_add_tool_result(
				tc_id ? tc_id : "", result_str ? result_str : "{}");
			g_free(result_str);
			/* Notify UI that a tool result was added. */
			cJSON *tmsg = cJSON_CreateObject();
			cJSON_AddStringToObject(tmsg, "role", "tool");
			agent_server_emit_event("agent_message", tmsg);
		}

		agent_llm_response_free(resp);
	}

	/* If we exited the loop without completing or erroring, we hit the
	 * iteration cap — treat as complete (the LLM was still working). */
	{
		g_mutex_lock(&g_ctx.status_lock);
		if (g_ctx.state != AGENT_LOOP_CANCELLED &&
		    g_ctx.state != AGENT_LOOP_ERROR) {
			g_ctx.state = AGENT_LOOP_COMPLETE;
		}
		g_mutex_unlock(&g_ctx.status_lock);
	}

done:
	g_free(system_prompt);
	cJSON_Delete(tools);
	g_atomic_int_set(&g_ctx.running, 0);
	return NULL;
}

/* ── Public API ─────────────────────────────────────────────────────── */

/*
 * Lazily initialize the mutex on first use. g_mutex_init() is idempotent
 * enough for our purposes — we guard with a g_once_init_enter block so
 * it only happens once.
 */
static void
ensure_mutex_init(void)
{
	static gsize initialized = 0;
	if (g_once_init_enter(&initialized)) {
		g_mutex_init(&g_ctx.status_lock);
		g_ctx.state = AGENT_LOOP_IDLE;
		g_once_init_leave(&initialized, 1);
	}
}

int
agent_loop_run(const char *user_message)
{
	g_print("[agent-loop] agent_loop_run called: %s\n",
	        user_message ? user_message : "(null)");
	if (user_message == NULL)
		return -1;

	ensure_mutex_init();

	if (g_atomic_int_get(&g_ctx.running)) {
		g_print("[agent-loop] already running, rejecting\n");
		return -1;  /* already running */
	}

	/* Check that a provider is configured (base URL + model).
	 * An empty API key is allowed (local servers like Ollama). */
	const browser_settings_t *s = settings_get();
	if (s->agent_llm_base_url[0] == '\0' || s->agent_llm_model[0] == '\0') {
		g_print("[agent-loop] no provider configured: base_url=%s model=%s\n",
		        s->agent_llm_base_url, s->agent_llm_model);
		set_error("No LLM provider configured (set base URL and model on "
		          "sovereign://agents)");
		return -1;
	}

	/* Reset cancel flag */
	g_atomic_int_set(&g_ctx.cancel_flag, 0);

	/* Reset status */
	set_status(AGENT_LOOP_THINKING, 0, NULL, NULL, NULL);

	g_atomic_int_set(&g_ctx.running, 1);

	/* Spawn background thread (takes ownership of the duplicated string) */
	g_ctx.thread = g_thread_new("agent-loop", agent_loop_thread,
	                            g_strdup(user_message));
	return 0;
}

void
agent_loop_cancel(void)
{
	g_atomic_int_set(&g_ctx.cancel_flag, 1);
}

gboolean
agent_loop_is_running(void)
{
	return g_atomic_int_get(&g_ctx.running) != 0;
}

void
agent_loop_get_status(agent_loop_state_t *state_out,
                      int *iteration_out,
                      char **current_tool_out,
                      char **last_message_out,
                      char **error_out)
{
	ensure_mutex_init();

	g_mutex_lock(&g_ctx.status_lock);

	if (state_out)
		*state_out = g_ctx.state;
	if (iteration_out)
		*iteration_out = g_ctx.iteration;
	if (current_tool_out)
		*current_tool_out = g_strdup(g_ctx.current_tool);
	if (last_message_out)
		*last_message_out = g_strdup(g_ctx.last_message);
	if (error_out)
		*error_out = g_strdup(g_ctx.error);

	g_mutex_unlock(&g_ctx.status_lock);
}