Files
cproof/src/ai/ai_client.c
jabber.developer2 f7a818dc1d
All checks were successful
CI Code / Check spelling (pull_request) Successful in 13s
CI Code / Check coding style (pull_request) Successful in 23s
CI Code / Code Coverage (pull_request) Successful in 3m43s
CI Code / Linux (debian) (pull_request) Successful in 5m13s
CI Code / Linux (ubuntu) (pull_request) Successful in 5m20s
CI Code / Linux (arch) (pull_request) Successful in 7m6s
fix(ai): bound response parsing and harden AUTO fallback errors
- bound _parse_responses scans to the output_text part's object and to
  the content array, so a later sibling item's "text" (e.g. a reasoning
  summary) can never be returned as the assistant reply
- do not retry the other flavour on a curl timeout: the request likely
  reached the server and may still be generating, so a re-POST of the
  conversation could trigger a second billed generation
- remember an unparseable 2xx from the first AUTO attempt and surface
  stashed first-attempt errors in both error paths, instead of showing
  only the final attempt's transport or HTTP error
- recognize Ollama's model-not-found wording in _names_model so a model
  typo is not misread as a missing endpoint
- drop the dead, racy provider-lookup fallback in the generic request
  thread: a missing provider ref is a caller bug and now fails loudly
- fix ai_providers_lists_defaults to expect the header the command
  actually prints ("Configured providers:"); the test was broken since
  its introduction but CI never ran it
- add functional test group 5 (AI command surface) to FUNC_TEST_GROUPS
  so the CI parallel target runs it; proftest.c port ranges already
  account for five groups
2026-07-10 13:30:44 +03:00

2317 lines
78 KiB
C

/*
* ai_client.c - AI client for interacting with OpenAI-compatible API providers
*
* Supports multiple providers (OpenAI, Perplexity, etc.) with per-provider
* API keys, custom endpoints, and model selection.
*
* Copyright (C) 2026 CProof Developers
*
* SPDX-License-Identifier: GPL-3.0-or-later WITH OpenSSL-exception
*/
// vim: expandtab:ts=4:sts=4:sw=4
#include "ai_client.h"
#include "common.h"
#include "log.h"
#include "config/preferences.h"
#include "profanity.h"
#include "tools/autocomplete.h"
#include "ui/ui.h"
#include "ui/window_list.h"
#include <curl/curl.h>
#include <glib.h>
#include <pthread.h>
#include <string.h>
/* Global state */
static GHashTable* providers = NULL;
static GHashTable* provider_keys = NULL;
static Autocomplete providers_ac = NULL;
/* ========================================================================
* Forward declarations
* ======================================================================== */
static AIProvider* _ai_add_provider_nopersist(const gchar* name, const gchar* api_url);
static void _ai_load_models_for_provider(AIProvider* provider);
static void _ai_save_models_for_provider(AIProvider* provider);
static const gchar* _find_json_field(const gchar* json, const gchar* field);
static gchar* _extract_json_string(const gchar* json, const gchar* field);
static const gchar* _json_array_end(const gchar* arr);
/* ========================================================================
* Curl helpers
* ======================================================================== */
struct curl_response_t
{
gchar* data;
size_t size;
};
static size_t
_write_callback(void* ptr, size_t size, size_t nmemb, void* userdata)
{
size_t realsize = size * nmemb;
struct curl_response_t* res = (struct curl_response_t*)userdata;
/* Limit response size to 10MB to prevent OOM */
if (res->size + realsize > 10 * 1024 * 1024) {
log_error("[AI-THREAD] Response too large, aborting transfer");
return CURL_WRITEFUNC_ERROR;
}
gchar* new_data = g_realloc(res->data, res->size + realsize + 1);
if (!new_data) {
log_error("[AI-THREAD] Failed to allocate memory for response");
return realsize;
}
res->data = new_data;
memcpy(res->data + res->size, ptr, realsize);
res->size += realsize;
res->data[res->size] = '\0';
return realsize;
}
/* ========================================================================
* Thread-safe UI display helpers
* ======================================================================== */
/**
* Validate aiwin pointer by checking if it still exists in the window list.
* Detects if the original window was freed (TOCTOU protection).
* Returns the validated pointer if valid, NULL if window was closed.
*/
static ProfAiWin*
_aiwin_validate(gpointer user_data)
{
if (!user_data) {
return NULL;
}
gboolean win_exists = wins_ai_exists((ProfAiWin*)user_data);
if (!win_exists) {
log_warning("[AI-THREAD] aiwin=%p no longer exists — window was closed", (void*)user_data);
return NULL;
}
/* Pointer is valid */
return (ProfAiWin*)user_data;
}
/**
* Display an error message in the AI window (thread-safe).
* @param user_data The original user_data pointer (may be NULL)
* @param error_msg The error message to display
*/
static void
_aiwin_display_error(gpointer user_data, const gchar* error_msg)
{
pthread_mutex_lock(&lock);
ProfAiWin* aiwin = _aiwin_validate(user_data);
if (!aiwin) {
log_warning("[AI-THREAD] Cannot display error: aiwin is invalid or window was closed (msg: %s)", error_msg);
pthread_mutex_unlock(&lock);
return;
}
aiwin_display_error(aiwin, error_msg);
pthread_mutex_unlock(&lock);
}
/**
* Display a response message. If aiwin is provided and valid, display in AI window.
* Otherwise, display in console (for commands run from console).
* @param user_data The original user_data pointer (may be NULL)
* @param response The response message to display
*/
static void
_aiwin_display_response(gpointer user_data, const gchar* response)
{
if (!response || strlen(response) == 0) {
return;
}
pthread_mutex_lock(&lock);
ProfAiWin* aiwin = _aiwin_validate(user_data);
if (!aiwin) {
cons_show("%s", response);
pthread_mutex_unlock(&lock);
return;
}
aiwin_display_response(aiwin, response);
pthread_mutex_unlock(&lock);
}
/* ========================================================================
* JSON helpers
* ======================================================================== */
gchar*
ai_json_escape(const gchar* str)
{
if (!str)
return g_strdup("");
GString* result = g_string_new("");
for (const gchar* p = str; *p; p++) {
switch (*p) {
case '"':
g_string_append(result, "\\\"");
break;
case '\\':
g_string_append(result, "\\\\");
break;
case '\b':
g_string_append(result, "\\b");
break;
case '\f':
g_string_append(result, "\\f");
break;
case '\n':
g_string_append(result, "\\n");
break;
case '\r':
g_string_append(result, "\\r");
break;
case '\t':
g_string_append(result, "\\t");
break;
default:
g_string_append_c(result, *p);
break;
}
}
return g_string_free(result, FALSE);
}
/* Keys emitted as fixed payload fields (in either API flavour); custom
* settings must not duplicate them */
static gboolean
_is_reserved_payload_key(const gchar* key)
{
return g_strcmp0(key, "model") == 0 || g_strcmp0(key, "messages") == 0
|| g_strcmp0(key, "input") == 0 || g_strcmp0(key, "stream") == 0;
}
/* RFC 8259 scalars (number/true/false/null) may be emitted bare in a JSON payload */
static gboolean
_is_json_bare_token(const gchar* value)
{
if (g_strcmp0(value, "true") == 0 || g_strcmp0(value, "false") == 0 || g_strcmp0(value, "null") == 0)
return TRUE;
/* Compile the number grammar once; this runs per custom setting per request */
static gsize once = 0;
static GRegex* number_re = NULL;
if (g_once_init_enter(&once)) {
number_re = g_regex_new("^-?(0|[1-9][0-9]*)(\\.[0-9]+)?([eE][+-]?[0-9]+)?$",
G_REGEX_OPTIMIZE, 0, NULL);
g_once_init_leave(&once, 1);
}
return number_re && g_regex_match(number_re, value, 0, NULL);
}
/**
* Convert a 4-hex-digit sequence at @p to a 16-bit code point.
* Returns -1 on invalid hex.
*/
static int
_hex4_to_int(const gchar* p)
{
int val = 0;
for (int i = 0; i < 4; i++) {
char c = p[i];
val <<= 4;
if (c >= '0' && c <= '9')
val |= c - '0';
else if (c >= 'a' && c <= 'f')
val |= c - 'a' + 10;
else if (c >= 'A' && c <= 'F')
val |= c - 'A' + 10;
else
return -1;
}
return val;
}
/**
* Write a UTF-8 encoding of @cp to @w, advancing @w.
* Returns the updated pointer.
*/
static gchar*
_write_utf8(gchar* w, unsigned int cp)
{
if (cp <= 0x7F) {
*w++ = (char)cp;
} else if (cp <= 0x7FF) {
*w++ = (char)(0xC0 | (cp >> 6));
*w++ = (char)(0x80 | (cp & 0x3F));
} else if (cp <= 0xFFFF) {
*w++ = (char)(0xE0 | (cp >> 12));
*w++ = (char)(0x80 | ((cp >> 6) & 0x3F));
*w++ = (char)(0x80 | (cp & 0x3F));
} else if (cp <= 0x10FFFF) {
*w++ = (char)(0xF0 | (cp >> 18));
*w++ = (char)(0x80 | ((cp >> 12) & 0x3F));
*w++ = (char)(0x80 | ((cp >> 6) & 0x3F));
*w++ = (char)(0x80 | (cp & 0x3F));
}
/* Invalid code point: silently skip */
return w;
}
/* Decode JSON string escapes in [start, end) into a freshly allocated buffer.
* Handles \" \\ \/ \n \t \r \b \f and \uXXXX (including surrogate pairs). */
static gchar*
_json_unescape_substring(const gchar* start, const gchar* end)
{
/* Estimate upper bound: each \uXXXX becomes up to 6 UTF-8 bytes. */
gsize max_len = (end - start) * 6 + 1;
gchar* out = g_new0(gchar, max_len);
gchar* w = out;
for (const gchar* in = start; in < end;) {
if (*in == '\\' && in + 1 < end) {
switch (in[1]) {
case '"':
*w++ = '"';
in += 2;
break;
case '\\':
*w++ = '\\';
in += 2;
break;
case '/':
*w++ = '/';
in += 2;
break;
case 'n':
*w++ = '\n';
in += 2;
break;
case 't':
*w++ = '\t';
in += 2;
break;
case 'r':
*w++ = '\r';
in += 2;
break;
case 'b':
*w++ = '\b';
in += 2;
break;
case 'f':
*w++ = '\f';
in += 2;
break;
case 'u':
/* Handle \uXXXX or surrogate pair \uHHHH\uLLLL */
if (in + 6 > end) {
/* Not enough characters — emit literal */
*w++ = *in++;
break;
}
if (in[2] != '\\' || in[3] != 'u') {
/* Single \uXXXX */
int cp = _hex4_to_int(in + 2);
if (cp < 0) {
/* Invalid hex — emit literal backslash */
*w++ = *in++;
} else if (cp < 0xD800 || cp > 0xDFFF) {
/* Not a surrogate — encode directly */
w = _write_utf8(w, (unsigned int)cp);
in += 6;
} else {
/* High or low surrogate — expect pair */
unsigned int high = (unsigned int)cp;
if ((high >= 0xD800 && high <= 0xDBFF) && in + 12 <= end && in[6] == '\\' && in[7] == 'u') {
/* High surrogate followed by \u */
int low = _hex4_to_int(in + 8);
if (low >= 0xDC00 && low <= 0xDFFF) {
/* Valid surrogate pair */
unsigned int cp2 = (0x10000 + ((high & 0x3FF) << 10)) | (low & 0x3FF);
w = _write_utf8(w, cp2);
in += 12;
} else {
/* Invalid low surrogate — emit replacement */
w = _write_utf8(w, 0xFFFD);
in += 6;
}
} else {
/* Orphan high surrogate or lone low surrogate */
w = _write_utf8(w, 0xFFFD);
in += 6;
}
}
} else {
/* \uHHHH\uLLLL — surrogate pair */
int high = _hex4_to_int(in + 2);
if (high < 0) {
*w++ = *in++;
break;
}
int low = _hex4_to_int(in + 8);
if ((high >= 0xD800 && high <= 0xDBFF) && (low >= 0xDC00 && low <= 0xDFFF)) {
unsigned int cp2 = (0x10000 + ((unsigned int)(high & 0x3FF) << 10)) | (unsigned int)(low & 0x3FF);
w = _write_utf8(w, cp2);
in += 12;
} else {
/* Invalid pair — emit replacement for high */
w = _write_utf8(w, 0xFFFD);
in += 6;
}
}
break;
default:
*w++ = *in++;
break;
}
} else {
*w++ = *in++;
}
}
*w = '\0';
/* Realloc to exact size */
return g_realloc(out, w - out + 1);
}
/* Unescape the string value starting at @val (a pointer to its opening quote).
* Treats any \X (X != \0) inside the string body as an opaque escape so an
* embedded \" cannot be misread as the closing quote. */
static gchar*
_extract_json_string_at(const gchar* val)
{
if (!val || *val != '"')
return NULL;
const gchar* start = val + 1;
for (const gchar* p = start; *p; p++) {
if (*p == '\\' && *(p + 1) != '\0') {
p++; /* skip the escaped character; loop ++ advances past it */
continue;
}
if (*p == '"') {
return _json_unescape_substring(start, p);
}
}
return NULL;
}
/* Locate "field":"...value..." in json and return the unescaped value.
* Skips JSON whitespace around the colon. @end (optional) bounds the search:
* a value starting at or past it is not extracted. */
static gchar*
_extract_json_string_within(const gchar* json, const gchar* field, const gchar* end)
{
const gchar* val = _find_json_field(json, field);
if (!val || (end && val >= end))
return NULL;
return _extract_json_string_at(val);
}
static gchar*
_extract_json_string(const gchar* json, const gchar* field)
{
return _extract_json_string_within(json, field, NULL);
}
/* ========================================================================
* API type helpers
* ======================================================================== */
gboolean
ai_api_type_from_string(const gchar* str, AIApiType* type)
{
if (!str || !type)
return FALSE;
if (g_strcmp0(str, "auto") == 0) {
*type = AI_API_TYPE_AUTO;
} else if (g_strcmp0(str, "responses") == 0) {
*type = AI_API_TYPE_RESPONSES;
} else if (g_strcmp0(str, "chat-completions") == 0) {
*type = AI_API_TYPE_CHAT_COMPLETIONS;
} else {
return FALSE;
}
return TRUE;
}
const gchar*
ai_api_type_to_string(AIApiType type)
{
switch (type) {
case AI_API_TYPE_RESPONSES:
return "responses";
case AI_API_TYPE_CHAT_COMPLETIONS:
return "chat-completions";
default:
return "auto";
}
}
static const gchar*
_api_type_endpoint(AIApiType type)
{
return type == AI_API_TYPE_CHAT_COMPLETIONS ? "v1/chat/completions" : "v1/responses";
}
static AIApiType
_other_flavour(AIApiType type)
{
return type == AI_API_TYPE_CHAT_COMPLETIONS ? AI_API_TYPE_RESPONSES : AI_API_TYPE_CHAT_COMPLETIONS;
}
static gboolean
_names_model(const gchar* s)
{
if (!s)
return FALSE;
/* "does not exist"/"not found" alone also appear in route-missing bodies;
* require them to talk about a model (covers Ollama's 'model "x" not
* found, try pulling it first') */
return strstr(s, "model_not_found") || strstr(s, "invalid_model")
|| strstr(s, "unknown_model")
|| (strstr(s, "model") && (strstr(s, "does not exist") || strstr(s, "not found")));
}
/* A missing endpoint answers 404/405/501. A wrong model can also 404/405, but
* the error names the model (OpenAI code model_not_found, vLLM message "The
* model `x` does not exist"), so treat the endpoint as present then. Prefer
* the structured error fields; fall back to the raw body for servers that
* send no JSON error envelope. */
static gboolean
_body_names_model_error(const gchar* body)
{
if (!body)
return FALSE;
/* Only trust a structured error envelope. Matching the raw body also
* fires on route-missing 404 pages that merely mention models (e.g. an
* HTML 404 listing /v1/models), which would wrongly suppress the fallback */
const gchar* err = strstr(body, "\"error\"");
if (!err)
return FALSE;
auto_gchar gchar* code = _extract_json_string(err, "code");
auto_gchar gchar* type = _extract_json_string(err, "type");
auto_gchar gchar* message = _extract_json_string(err, "message");
return _names_model(code) || _names_model(type) || _names_model(message);
}
static gboolean
_is_endpoint_missing(long http_code, const gchar* body)
{
if (http_code != 404 && http_code != 405 && http_code != 501)
return FALSE;
return !_body_names_model_error(body);
}
/* During an AUTO probe, move on to the other flavour when this endpoint is
* missing or rejects the payload shape (400/422) — but not on auth/rate-limit
* or model errors, which the other endpoint would fail the same way. */
static gboolean
_should_try_other_flavour(long http_code, const gchar* body)
{
if (_is_endpoint_missing(http_code, body))
return TRUE;
if (http_code != 400 && http_code != 422)
return FALSE;
return !_body_names_model_error(body);
}
/* Copy the provider URL under settings_lock, slash-terminated; @epoch_out
* (optional) captures api_epoch in the same critical section so a flavour
* stamped later can be tied to the URL it was probed against */
static gchar*
_provider_base_url(AIProvider* provider, gint* epoch_out)
{
pthread_mutex_lock(&provider->settings_lock);
gchar* base = g_strdup_printf("%s%s", provider->api_url,
g_str_has_suffix(provider->api_url, "/") ? "" : "/");
if (epoch_out)
*epoch_out = g_atomic_int_get(&provider->api_epoch);
pthread_mutex_unlock(&provider->settings_lock);
return base;
}
/* Single writer for resolved_api_type: takes settings_lock and, when
* @expected_epoch >= 0, drops the write if the epoch moved (the URL or
* api-type changed while the request was in flight) */
static void
_provider_set_resolved(AIProvider* provider, AIApiType type, gint expected_epoch)
{
pthread_mutex_lock(&provider->settings_lock);
if (expected_epoch < 0 || g_atomic_int_get(&provider->api_epoch) == expected_epoch)
g_atomic_int_set(&provider->resolved_api_type, type);
pthread_mutex_unlock(&provider->settings_lock);
}
/* ========================================================================
* Provider Management
* ======================================================================== */
static AIProvider*
ai_provider_new(const gchar* name, const gchar* api_url)
{
AIProvider* provider = g_new0(AIProvider, 1);
provider->name = g_strdup(name);
provider->api_url = g_strdup(api_url ? api_url : "");
provider->project_id = NULL;
provider->default_model = NULL;
provider->api_type = AI_API_TYPE_AUTO;
provider->resolved_api_type = AI_API_TYPE_AUTO;
provider->api_epoch = 0;
provider->settings = g_hash_table_new_full(g_str_hash, g_str_equal, g_free, g_free);
provider->models = NULL;
provider->models_fresh = FALSE;
provider->ref_count = 1;
pthread_mutex_init(&provider->settings_lock, NULL);
return provider;
}
AIProvider*
ai_provider_ref(AIProvider* provider)
{
if (provider) {
g_atomic_int_inc(&provider->ref_count);
}
return provider;
}
void
ai_provider_unref(AIProvider* provider)
{
if (!provider)
return;
if (!g_atomic_int_dec_and_test(&provider->ref_count)) {
return;
}
g_free(provider->name);
g_free(provider->api_url);
g_free(provider->project_id);
g_free(provider->default_model);
g_hash_table_destroy(provider->settings);
pthread_mutex_destroy(&provider->settings_lock);
GList* curr = provider->models;
while (curr) {
g_free(curr->data);
curr = g_list_next(curr);
}
g_list_free(provider->models);
g_free(provider);
}
static void
ai_load_keys(void)
{
if (!provider_keys) {
provider_keys = g_hash_table_new_full(g_str_hash, g_str_equal, g_free, g_free);
}
GList* tokens = prefs_ai_list_tokens();
if (!tokens) {
return;
}
GList* curr = tokens;
while (curr) {
gchar* provider = (gchar*)curr->data;
gchar* key = prefs_ai_get_token(provider);
if (key && strlen(key) > 0) {
g_hash_table_insert(provider_keys, g_strdup(provider), g_strdup(key));
}
g_free(key);
curr = g_list_next(curr);
}
prefs_free_ai_tokens(tokens);
log_info("Loaded %d saved API keys from config", g_hash_table_size(provider_keys));
}
void
ai_client_init(void)
{
if (providers)
return; /* Already initialized */
curl_global_init(CURL_GLOBAL_ALL);
/* Create hash tables */
providers = g_hash_table_new_full(g_str_hash, g_str_equal, g_free, (GDestroyNotify)ai_provider_unref);
provider_keys = g_hash_table_new_full(g_str_hash, g_str_equal, g_free, g_free);
/* Create autocomplete for provider names */
providers_ac = autocomplete_new();
/* Get all providers (defaults if nothing configured, otherwise config values) */
GList* all_providers = prefs_ai_get_providers();
for (GList* curr = all_providers; curr; curr = g_list_next(curr)) {
/* List contains pairs: name, url, name, url, ... */
gchar* name = (gchar*)curr->data;
curr = g_list_next(curr);
if (!curr)
break;
gchar* url = (gchar*)curr->data;
_ai_add_provider_nopersist(name, url);
}
prefs_free_ai_providers(all_providers);
/* Load saved API keys from config */
ai_load_keys();
/* Load cached models, persisted default-model and custom settings for
* all providers. */
GList* provider_list = ai_list_providers();
for (GList* curr = provider_list; curr; curr = g_list_next(curr)) {
AIProvider* provider = (AIProvider*)curr->data;
_ai_load_models_for_provider(provider);
auto_gchar gchar* dm = prefs_ai_get_default_model(provider->name);
if (dm && dm[0] != '\0') {
g_free(provider->default_model);
provider->default_model = g_strdup(dm);
}
auto_gchar gchar* at = prefs_ai_get_api_type(provider->name);
AIApiType api_type;
if (at && ai_api_type_from_string(at, &api_type)) {
provider->api_type = api_type;
}
GList* settings = prefs_ai_list_settings(provider->name);
for (GList* s = settings; s; s = g_list_next(s)) {
const gchar* name = (const gchar*)s->data;
auto_gchar gchar* val = prefs_ai_get_setting(provider->name, name);
if (val) {
if (!provider->settings) {
provider->settings = g_hash_table_new_full(g_str_hash, g_str_equal, g_free, g_free);
}
g_hash_table_insert(provider->settings, g_strdup(name), g_strdup(val));
}
}
prefs_free_ai_settings(settings);
}
g_list_free(provider_list);
log_info("AI client initialized with providers from preferences");
}
void
ai_client_shutdown(void)
{
if (!providers)
return;
g_hash_table_destroy(providers);
g_hash_table_destroy(provider_keys);
providers = NULL;
provider_keys = NULL;
if (providers_ac) {
autocomplete_free(providers_ac);
providers_ac = NULL;
}
curl_global_cleanup();
log_info("AI client shutdown");
}
AIProvider*
ai_get_provider(const gchar* name)
{
if (!name || !providers)
return NULL;
return g_hash_table_lookup(providers, name);
}
/* Internal helper to add provider without persisting to config.
* Used for default providers and loading from config on startup. */
static AIProvider*
_ai_add_provider_nopersist(const gchar* name, const gchar* api_url)
{
/* Check if provider already exists */
AIProvider* existing = g_hash_table_lookup(providers, name);
if (existing) {
pthread_mutex_lock(&existing->settings_lock);
g_free(existing->api_url);
existing->api_url = g_strdup(api_url);
g_atomic_int_inc(&existing->api_epoch);
/* New URL may serve a different API flavour; reset under the lock so a
* request thread cannot re-stamp the stale flavour for the new URL */
g_atomic_int_set(&existing->resolved_api_type, AI_API_TYPE_AUTO);
pthread_mutex_unlock(&existing->settings_lock);
return existing; /* borrowed, same as the create path below */
}
/* Create new provider (ref_count=1 owned by hash table) */
AIProvider* provider = ai_provider_new(name, api_url);
g_hash_table_insert(providers, g_strdup(name), provider);
/* Sync autocomplete */
autocomplete_add(providers_ac, name);
return provider;
}
AIProvider*
ai_add_provider(const gchar* name, const gchar* api_url)
{
if (!name || !api_url)
return NULL;
if (!providers) {
ai_client_init();
}
gboolean existed = g_hash_table_lookup(providers, name) != NULL;
/* Persist first, then reuse the internal helper for both the update and
* create paths (it handles the locked URL swap or the fresh insert) */
prefs_ai_set_provider(name, api_url);
AIProvider* provider = _ai_add_provider_nopersist(name, api_url);
if (existed) {
log_info("Updated provider: %s", name);
} else {
log_info("Added provider: %s (URL: %s)", name, api_url);
}
return provider; /* Caller gets non-owning pointer; hash table owns ref */
}
gboolean
ai_remove_provider(const gchar* name)
{
if (!name || !providers)
return FALSE;
/* Wipe the whole [ai/<name>] subsection from config — url, key,
* default_model, models, custom settings — in one call. */
prefs_ai_remove_section(name);
/* Sync autocomplete before removing */
autocomplete_remove(providers_ac, name);
return g_hash_table_remove(providers, name);
}
GList*
ai_list_providers(void)
{
if (!providers)
return NULL;
GList* result = NULL;
GHashTableIter iter;
gpointer key, value;
g_hash_table_iter_init(&iter, providers);
while (g_hash_table_iter_next(&iter, &key, &value)) {
result = g_list_append(result, value);
}
return result;
}
/* ========================================================================
* Provider autocomplete state
* ======================================================================== */
gchar*
ai_providers_find(const char* const search_str, gboolean previous, void* context)
{
if (!providers_ac) {
log_debug("[AI-AC] Uninitialized call to ai_providers_find");
return NULL;
}
return autocomplete_complete(providers_ac, search_str, FALSE, previous);
}
/* Reset the provider autocomplete state.
* Called from cmd_ac_reset() to clear autocomplete state when the user presses Ctrl+C. */
void
ai_providers_reset_ac(void)
{
if (!providers_ac) {
return;
}
autocomplete_reset(providers_ac);
}
gchar*
ai_get_provider_key(const gchar* provider_name)
{
if (!provider_name || !provider_keys)
return NULL;
gchar* key = g_hash_table_lookup(provider_keys, provider_name);
return g_strdup(key);
}
void
ai_set_provider_key(const gchar* provider_name, const gchar* api_key)
{
if (!provider_name)
return;
if (!provider_keys) {
provider_keys = g_hash_table_new_full(g_str_hash, g_str_equal, g_free, g_free);
}
if (api_key) {
g_hash_table_insert(provider_keys, g_strdup(provider_name), g_strdup(api_key));
/* Persist to config file */
prefs_ai_set_token(provider_name, api_key);
log_info("API key set for provider: %s", provider_name);
} else {
g_hash_table_remove(provider_keys, provider_name);
/* Remove from config file */
prefs_ai_remove_token(provider_name);
log_info("API key removed for provider: %s", provider_name);
}
}
/* ========================================================================
* Provider default model and settings
* ======================================================================== */
void
ai_set_provider_default_model(const gchar* provider_name, const gchar* model)
{
if (!provider_name || !model)
return;
AIProvider* provider = ai_get_provider(provider_name);
if (!provider) {
log_warning("Provider '%s' not found for default model setting", provider_name);
return;
}
g_free(provider->default_model);
provider->default_model = g_strdup(model);
prefs_ai_set_default_model(provider_name, model);
log_info("Default model for provider '%s' set to: %s", provider_name, model);
}
const gchar*
ai_get_provider_default_model(const gchar* provider_name)
{
if (!provider_name)
return NULL;
AIProvider* provider = ai_get_provider(provider_name);
if (!provider)
return NULL;
return provider->default_model;
}
gboolean
ai_set_provider_api_type(const gchar* provider_name, const gchar* api_type)
{
AIApiType type;
if (!provider_name || !ai_api_type_from_string(api_type, &type))
return FALSE;
AIProvider* provider = ai_get_provider(provider_name);
if (!provider) {
log_warning("Provider '%s' not found for API type setting", provider_name);
return FALSE;
}
/* Same protocol as a URL swap: bump the epoch under the lock so an
* in-flight request cannot re-stamp resolved_api_type after this reset */
pthread_mutex_lock(&provider->settings_lock);
g_atomic_int_set(&provider->api_type, type);
g_atomic_int_inc(&provider->api_epoch);
g_atomic_int_set(&provider->resolved_api_type, AI_API_TYPE_AUTO);
pthread_mutex_unlock(&provider->settings_lock);
if (type == AI_API_TYPE_AUTO) {
prefs_ai_remove_api_type(provider_name);
} else {
prefs_ai_set_api_type(provider_name, api_type);
}
log_info("API type for provider '%s' set to: %s", provider_name, ai_api_type_to_string(type));
return TRUE;
}
AIApiType
ai_get_provider_api_type(const gchar* provider_name)
{
AIProvider* provider = provider_name ? ai_get_provider(provider_name) : NULL;
if (!provider)
return AI_API_TYPE_AUTO;
return g_atomic_int_get(&provider->api_type);
}
gboolean
ai_set_provider_setting(const gchar* provider_name, const gchar* setting, const gchar* value)
{
if (!provider_name || !setting)
return FALSE;
/* Reject setting a reserved key, but still allow removing one so a stale
* entry persisted before the key was reserved can be cleared */
if (value && _is_reserved_payload_key(setting)) {
log_warning("Setting '%s' for provider '%s' rejected: reserved payload key", setting, provider_name);
return FALSE;
}
AIProvider* provider = ai_get_provider(provider_name);
if (!provider) {
log_warning("Provider '%s' not found for setting '%s'", provider_name, setting);
return FALSE;
}
/* settings is iterated by the request thread; mutate under settings_lock */
pthread_mutex_lock(&provider->settings_lock);
if (!provider->settings) {
provider->settings = g_hash_table_new_full(g_str_hash, g_str_equal, g_free, g_free);
}
if (value) {
g_hash_table_insert(provider->settings, g_strdup(setting), g_strdup(value));
} else {
g_hash_table_remove(provider->settings, setting);
}
pthread_mutex_unlock(&provider->settings_lock);
if (value) {
prefs_ai_set_setting(provider_name, setting, value);
log_info("Setting '%s' for provider '%s' set to: %s", setting, provider_name, value);
} else {
prefs_ai_remove_setting(provider_name, setting);
log_info("Setting '%s' removed for provider '%s'", setting, provider_name);
}
return TRUE;
}
gchar*
ai_get_provider_setting(const gchar* provider_name, const gchar* setting)
{
if (!provider_name || !setting)
return NULL;
AIProvider* provider = ai_get_provider(provider_name);
if (!provider || !provider->settings)
return NULL;
pthread_mutex_lock(&provider->settings_lock);
gchar* value = g_strdup(g_hash_table_lookup(provider->settings, setting));
pthread_mutex_unlock(&provider->settings_lock);
return value;
}
GList*
ai_get_provider_setting_keys(const gchar* provider_name)
{
AIProvider* provider = provider_name ? ai_get_provider(provider_name) : NULL;
if (!provider || !provider->settings)
return NULL;
GList* keys = NULL;
pthread_mutex_lock(&provider->settings_lock);
GHashTableIter iter;
gpointer key, value;
g_hash_table_iter_init(&iter, provider->settings);
while (g_hash_table_iter_next(&iter, &key, &value)) {
keys = g_list_prepend(keys, g_strdup((gchar*)key));
}
pthread_mutex_unlock(&provider->settings_lock);
return keys;
}
/* ========================================================================
* Generic HTTP request handling
* ======================================================================== */
typedef void (*ai_response_handler_t)(AIProvider* provider, const gchar* response, gpointer user_data);
typedef struct
{
gchar* provider_name;
AIProvider* provider; /* owned ref taken on the main thread; required */
gpointer user_data;
gchar* request_url;
ai_response_handler_t handler;
} ai_request_t;
/**
* Build curl headers with auth.
* Returns headers list (caller must free with curl_slist_free_all).
*/
static struct curl_slist*
_build_curl_headers(AIProvider* provider, const gchar* api_key)
{
struct curl_slist* headers = NULL;
auto_gchar gchar* auth_header = g_strdup_printf("Authorization: Bearer %s", api_key);
headers = curl_slist_append(headers, "Content-Type: application/json");
headers = curl_slist_append(headers, auth_header);
return headers;
}
/**
* Execute a curl request and call handler on success.
* Returns CURLcode.
*/
static CURLcode
_curl_exec_and_handle(CURL* curl, const gchar* url, struct curl_slist* headers,
struct curl_response_t* response, ai_response_handler_t handler,
AIProvider* provider, gpointer user_data)
{
curl_easy_setopt(curl, CURLOPT_URL, url);
curl_easy_setopt(curl, CURLOPT_HTTPHEADER, headers);
curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, _write_callback);
curl_easy_setopt(curl, CURLOPT_WRITEDATA, response);
curl_easy_setopt(curl, CURLOPT_TIMEOUT, 30L);
CURLcode res = curl_easy_perform(curl);
long http_code = 0;
curl_easy_getinfo(curl, CURLINFO_RESPONSE_CODE, &http_code);
if (res != CURLE_OK) {
auto_gchar gchar* error_msg = g_strdup(curl_easy_strerror(res));
log_error("Request failed for '%s': %s", provider->name, error_msg);
_aiwin_display_error(user_data, error_msg);
} else if (http_code >= 400) {
auto_gchar gchar* parsed = ai_parse_error_message(response->data);
auto_gchar gchar* error_msg = parsed
? g_strdup_printf("HTTP %ld: %s", http_code, parsed)
: g_strdup_printf("HTTP %ld: %s", http_code,
response->data && strlen(response->data) > 0 ? response->data : "Unknown error");
log_error("Request HTTP error for '%s': %s", provider->name, error_msg);
_aiwin_display_error(user_data, error_msg);
} else {
handler(provider, response->data, user_data);
}
return res;
}
static gpointer
_ai_generic_request_thread(gpointer data)
{
ai_request_t* req = (ai_request_t*)data;
/* The ref taken on the main thread keeps the provider alive for the
* duration of this request; a thread-side lookup instead would race
* against /ai remove provider, so its absence is a caller bug */
AIProvider* provider = g_steal_pointer(&req->provider);
if (!provider) {
log_error("BUG: request for '%s' spawned without a provider ref", req->provider_name);
g_free(req->provider_name);
g_free(req->request_url);
g_free(req);
return NULL;
}
CURL* curl = curl_easy_init();
if (!curl) {
log_error("Failed to initialize curl for %s", req->provider_name);
ai_provider_unref(provider);
g_free(req->provider_name);
g_free(req->request_url);
g_free(req);
return NULL;
}
auto_gchar gchar* api_key = ai_get_provider_key(req->provider_name);
if (!api_key || strlen(api_key) == 0) {
auto_gchar gchar* error_msg = g_strdup_printf("No API key set for provider '%s'.", req->provider_name);
_aiwin_display_error(req->user_data, error_msg);
curl_easy_cleanup(curl);
ai_provider_unref(provider);
g_free(req->provider_name);
g_free(req->request_url);
g_free(req);
return NULL;
}
struct curl_slist* headers = _build_curl_headers(provider, api_key);
struct curl_response_t response;
response.data = g_new0(gchar, 1);
response.size = 0;
_curl_exec_and_handle(curl, req->request_url, headers, &response, req->handler, provider, req->user_data);
curl_slist_free_all(headers);
curl_easy_cleanup(curl);
g_free(response.data);
/* Release the main-thread ref — provider may now be freed. */
ai_provider_unref(provider);
g_free(req->provider_name);
g_free(req->request_url);
g_free(req);
return NULL;
}
/* ========================================================================
* Model fetching
* ======================================================================== */
gboolean
ai_models_are_fresh(const gchar* provider_name)
{
if (!provider_name)
return FALSE;
AIProvider* provider = ai_get_provider(provider_name);
if (!provider)
return FALSE;
return provider->models_fresh;
}
static gboolean
_model_exists(AIProvider* provider, const gchar* model)
{
GList* curr = provider->models;
while (curr) {
if (g_strcmp0(curr->data, model) == 0) {
return TRUE;
}
curr = g_list_next(curr);
}
return FALSE;
}
static void
_add_model_if_new(AIProvider* provider, const gchar* model)
{
if (!_model_exists(provider, model)) {
provider->models = g_list_append(provider->models, g_strdup(model));
}
}
/* ========================================================================
* Model cache persistence
* ======================================================================== */
/**
* Load cached models for a provider from preferences.
*/
static void
_ai_load_models_for_provider(AIProvider* provider)
{
if (!provider)
return;
GList* models = prefs_ai_get_models(provider->name);
if (!models) {
log_info("No cached models for provider '%s'", provider->name);
return;
}
/* Add loaded models to provider */
GList* curr = models;
while (curr) {
_add_model_if_new(provider, curr->data);
curr = g_list_next(curr);
}
prefs_free_ai_models(models);
if (provider->models) {
provider->models_fresh = TRUE;
log_info("Loaded %d cached models for provider '%s'", g_list_length(provider->models), provider->name);
}
}
/**
* Save models for a provider to preferences.
*/
static void
_ai_save_models_for_provider(AIProvider* provider)
{
if (!provider || !provider->models)
return;
/* Build array of model strings */
gint count = g_list_length(provider->models);
gchar** models = g_new0(gchar*, count);
GList* curr = provider->models;
gint i = 0;
while (curr) {
models[i++] = curr->data;
curr = g_list_next(curr);
}
prefs_ai_set_models(provider->name, (const gchar* const*)models, count);
g_free(models);
log_debug("Saved %d models for provider '%s' to prefs", count, provider->name);
}
/**
* Find a JSON field value, handling both string and non-string values.
* For string values: "field":"value" -> returns pointer to "value"
* For array values: "field":[ -> returns pointer to [
* For object values: "field":{ -> returns pointer to {
* Handles optional whitespace after colon.
* Returns pointer to the value start, or NULL if not found.
*/
static const gchar*
_find_json_field(const gchar* json, const gchar* field)
{
if (!json || !field) {
return NULL;
}
auto_gchar gchar* key = g_strdup_printf("\"%s\"", field);
gsize keylen = strlen(key);
/* An occurrence not followed by ':' is a string value equal to the field
* name (e.g. "text" as the value of "type"), not the key — skip it and
* look for the next one instead of giving up */
for (const gchar* search = json; (search = strstr(search, key)) != NULL; search += keylen) {
const gchar* pos = search + keylen;
while (*pos == ' ' || *pos == '\t' || *pos == '\n' || *pos == '\r') {
pos++;
}
if (*pos != ':') {
continue;
}
pos++; /* skip colon */
while (*pos == ' ' || *pos == '\t' || *pos == '\n' || *pos == '\r') {
pos++;
}
return pos;
}
return NULL;
}
/**
* Parse models from OpenAI-compatible API response format.
* Expected format: {"object":"list","data":[{"id":"model1","object":"model",...},...]}
* Extracts only the "id" field from each model object.
*/
static void
_parse_openai_models(AIProvider* provider, const gchar* json)
{
if (!json || !provider) {
return;
}
/* Find the "data" array start */
const gchar* data_start = _find_json_field(json, "data");
if (!data_start || *data_start != '[') {
log_debug("No data array found in API response");
return;
}
/* Find all {"id":"..."} objects within the data array */
const gchar* obj_start = data_start;
while (obj_start) {
/* Find next "id" field (handle both {"id": and { "id": formats) */
obj_start = strstr(obj_start, "\"id\"");
if (!obj_start) {
break;
}
/* Verify this "id" is within an object (preceded by { or ,) */
const gchar* before_id = obj_start - 1;
while (before_id >= data_start && (*before_id == ' ' || *before_id == '\t' || *before_id == '\n' || *before_id == '\r')) {
before_id--;
}
if (before_id < data_start || (*before_id != '{' && *before_id != ',')) {
/* This "id" is not at the start of an object in the array, skip it */
obj_start = obj_start + 2; /* move past this "id" */
continue;
}
/* Find the colon after "id" */
const gchar* id_key = obj_start + 4; /* skip "id" */
/* Skip whitespace */
while (*id_key == ' ' || *id_key == '\t' || *id_key == '\n' || *id_key == '\r') {
id_key++;
}
if (*id_key != ':') {
obj_start = id_key;
continue;
}
id_key++; /* skip : */
/* Skip whitespace */
while (*id_key == ' ' || *id_key == '\t' || *id_key == '\n' || *id_key == '\r') {
id_key++;
}
if (*id_key != '"') {
/* "id" value is not a string, skip */
obj_start = id_key;
continue;
}
id_key++; /* skip opening quote */
/* Find the closing quote (handling escaped quotes) */
const gchar* id_end = id_key;
while (*id_end && *id_end != '"') {
if (*id_end == '\\' && *(id_end + 1)) {
id_end += 2; /* skip escaped character */
} else {
id_end++;
}
}
if (!*id_end) {
break; /* No closing quote found */
}
/* Extract the model ID */
auto_gchar gchar* model = g_strndup(id_key, id_end - id_key);
_add_model_if_new(provider, model);
/* Move past this object to find the next one */
obj_start = id_end + 1;
/* Safety: stop if we've gone past the data array */
if (obj_start - json > 5 * 1024 * 1024) {
break; /* 5MB safety limit */
}
}
}
static void
_parse_array_models(AIProvider* provider, const gchar* json)
{
/* Format: ["model1", "model2", ...] */
const gchar* arr_start = strchr(json, '[');
if (!arr_start) {
return;
}
const gchar* model_start = strchr(arr_start, '"');
while (model_start) {
const gchar* model_end = strchr(model_start + 1, '"');
if (!model_end) {
break;
}
auto_gchar gchar* model = g_strndup(model_start + 1, model_end - model_start - 1);
_add_model_if_new(provider, model);
model_start = strchr(model_end + 1, '"');
}
}
static void
_parse_and_cache_models(AIProvider* provider, const gchar* response_json)
{
if (!provider || !response_json) {
return;
}
/* Try OpenAI format first: {"data": [{"id": "model1"}, ...]} */
_parse_openai_models(provider, response_json);
/* If no models found, try simple array format: ["model1", ...] */
if (!provider->models) {
_parse_array_models(provider, response_json);
}
provider->models_fresh = TRUE;
log_info("Cached %d models for provider '%s'", g_list_length(provider->models), provider->name);
/* Persist models to preferences */
_ai_save_models_for_provider(provider);
}
/**
* Public API: Parse model IDs from JSON response.
* Used by tests and internal model fetching.
*/
void
ai_parse_models_from_json(AIProvider* provider, const gchar* json)
{
if (!provider || !json) {
return;
}
/* Clear existing models before parsing */
GList* curr = provider->models;
while (curr) {
g_free(curr->data);
curr = g_list_next(curr);
}
g_list_free(provider->models);
provider->models = NULL;
/* Parse using the same logic as the internal handler */
_parse_and_cache_models(provider, json);
}
static void
_models_response_handler(AIProvider* provider, const gchar* response, gpointer user_data)
{
_parse_and_cache_models(provider, response);
gint count = g_list_length(provider->models);
GString* msg = g_string_new(NULL);
g_string_append_printf(msg, "Cached %d models for provider '%s':", count, provider->name);
for (GList* curr = provider->models; curr; curr = g_list_next(curr)) {
g_string_append_printf(msg, "\n %s", (const gchar*)curr->data);
}
if (count > 0) {
g_string_append_printf(msg, "\nUse '/ai start %s <model>' to start a chat.", provider->name);
}
auto_gchar gchar* out = g_string_free(msg, FALSE);
_aiwin_display_response(user_data, out);
}
static gpointer
_models_fetch_thread(gpointer data)
{
ai_request_t* req = (ai_request_t*)data;
/* req->provider carries a ref taken on the main thread, so the provider
* cannot be freed under us; the locked snapshot guards the URL swap */
auto_gchar gchar* base_url = _provider_base_url(req->provider, NULL);
req->request_url = g_strdup_printf("%sv1/models", base_url);
req->handler = _models_response_handler;
return _ai_generic_request_thread(req);
}
gboolean
ai_fetch_models(const gchar* provider_name, gpointer user_data)
{
if (!provider_name)
return FALSE;
AIProvider* provider = ai_get_provider(provider_name);
if (!provider) {
log_error("Provider '%s' not found for model fetch", provider_name);
return FALSE;
}
/* Always fetch fresh models from API. Ref the provider on this thread,
* while it is guaranteed alive, so /ai remove provider cannot free it
* under the worker; the worker's request struct owns the ref. */
ai_request_t* req = g_new0(ai_request_t, 1);
req->provider_name = g_strdup(provider_name);
req->provider = ai_provider_ref(provider);
req->user_data = user_data;
GThread* thread = g_thread_new("ai_models_fetch", _models_fetch_thread, req);
if (!thread) {
ai_provider_unref(req->provider);
g_free(req->provider_name);
g_free(req);
log_error("Failed to create thread for model fetch");
return FALSE;
}
g_thread_unref(thread);
cons_show("Fetching fresh models for provider '%s'...", provider_name);
return TRUE;
}
/* ========================================================================
* Session Management
* ======================================================================== */
AISession*
ai_session_create(const gchar* provider_name, const gchar* model)
{
if (!provider_name || !model)
return NULL;
AIProvider* provider = ai_get_provider(provider_name);
if (!provider) {
log_error("Provider not found: %s", provider_name);
return NULL;
}
AISession* session = g_new0(AISession, 1);
if (pthread_mutex_init(&session->lock, NULL) != 0) {
log_error("Failed to initialize session mutex");
g_free(session);
return NULL;
}
session->provider_name = g_strdup(provider_name);
session->provider = ai_provider_ref(provider);
session->model = g_strdup(model);
session->api_key = ai_get_provider_key(provider_name);
session->history = NULL;
session->ref_count = 1;
log_info("AI session created: %s/%s", provider_name, model);
return session;
}
AISession*
ai_session_ref(AISession* session)
{
if (session) {
g_atomic_int_inc(&session->ref_count);
}
return session;
}
void
ai_session_unref(AISession* session)
{
if (!session)
return;
if (!g_atomic_int_dec_and_test(&session->ref_count)) {
return;
}
g_free(session->provider_name);
ai_provider_unref(session->provider);
g_free(session->model);
g_free(session->api_key);
GList* curr = session->history;
while (curr) {
AIMessage* msg = curr->data;
g_free(msg->role);
g_free(msg->content);
g_free(msg);
curr = g_list_next(curr);
}
g_list_free(session->history);
pthread_mutex_destroy(&session->lock);
g_free(session);
log_debug("AI session destroyed");
}
void
ai_session_add_message(AISession* session, const gchar* role, const gchar* content)
{
if (!session || !role || !content)
return;
pthread_mutex_lock(&session->lock);
AIMessage* msg = g_new0(AIMessage, 1);
msg->role = g_strdup(role);
msg->content = g_strdup(content);
session->history = g_list_append(session->history, msg);
pthread_mutex_unlock(&session->lock);
log_debug("Added %s message to session (total: %d)", role, g_list_length(session->history));
}
void
ai_session_clear_history(AISession* session)
{
if (!session)
return;
pthread_mutex_lock(&session->lock);
GList* curr = session->history;
while (curr) {
AIMessage* msg = curr->data;
g_free(msg->role);
g_free(msg->content);
g_free(msg);
curr = g_list_next(curr);
}
g_list_free(session->history);
session->history = NULL;
pthread_mutex_unlock(&session->lock);
log_info("AI session history cleared");
}
const gchar*
ai_session_get_model(AISession* session)
{
if (!session)
return NULL;
return session->model;
}
void
ai_session_set_model(AISession* session, const gchar* model)
{
if (!session || !model)
return;
pthread_mutex_lock(&session->lock);
g_free(session->model);
session->model = g_strdup(model);
pthread_mutex_unlock(&session->lock);
log_info("Session model changed to: %s", model);
}
void
ai_session_switch(AISession* session, const gchar* provider_name,
const gchar* model, gchar* api_key)
{
if (!session || !provider_name || !model || !api_key)
return;
AIProvider* provider = ai_get_provider(provider_name);
if (!provider) {
log_warning("Provider '%s' not found for session switch", provider_name);
g_free(api_key);
return;
}
pthread_mutex_lock(&session->lock);
g_free(session->provider_name);
session->provider_name = g_strdup(provider_name);
ai_provider_unref(session->provider);
session->provider = ai_provider_ref(provider);
g_free(session->model);
session->model = g_strdup(model);
g_free(session->api_key);
session->api_key = g_strdup(api_key);
pthread_mutex_unlock(&session->lock);
g_free(api_key);
log_info("Session switched to %s/%s", provider_name, model);
}
/* ========================================================================
* API Request Handling
* ======================================================================== */
/**
* Serialize history plus the new user prompt as a JSON message-object list
* (without the enclosing brackets); the fragment is shared by both API
* flavours. Must be called with the session lock held.
*
* @param history GList of AIMessage* (must be held under session lock)
* @param prompt The new user prompt to append
* @return Newly allocated fragment (caller must free)
*/
static gchar*
_serialize_message_list(GList* history, const gchar* prompt)
{
GString* messages_json = g_string_new("");
for (GList* curr = history; curr; curr = g_list_next(curr)) {
AIMessage* msg = curr->data;
auto_gchar gchar* escaped_content = ai_json_escape(msg->content);
auto_gchar gchar* escaped_role = ai_json_escape(msg->role);
if (messages_json->len > 0) {
g_string_append_c(messages_json, ',');
}
g_string_append_printf(messages_json, "{\"role\":\"%s\",\"content\":\"%s\"}",
escaped_role, escaped_content);
}
/* Add the new user message */
auto_gchar gchar* escaped_prompt = ai_json_escape(prompt);
if (messages_json->len > 0) {
g_string_append_c(messages_json, ',');
}
g_string_append_printf(messages_json, "{\"role\":\"user\",\"content\":\"%s\"}", escaped_prompt);
return g_string_free(messages_json, FALSE);
}
/**
* Serialize custom provider settings as a JSON key-value fragment: scalar
* values (number/true/false/null) are emitted bare, anything else as a JSON
* string; reserved keys (model/messages/input/stream) are skipped.
*
* @param provider The AI provider (may be NULL)
* @param store_set Output: set TRUE when a custom "store" value is present
* @return Newly allocated fragment, empty string when there are no settings
*/
static gchar*
_serialize_custom_settings(AIProvider* provider, gboolean* store_set)
{
GString* custom_json = g_string_new("");
*store_set = FALSE;
if (provider) {
pthread_mutex_lock(&provider->settings_lock);
if (provider->settings) {
GHashTableIter iter;
gpointer key, value;
g_hash_table_iter_init(&iter, provider->settings);
while (g_hash_table_iter_next(&iter, &key, &value)) {
if (_is_reserved_payload_key((gchar*)key)) {
continue; /* fixed payload fields stay authoritative */
}
if (g_strcmp0((gchar*)key, "store") == 0) {
*store_set = TRUE; /* suppresses the fixed "store":false */
}
auto_gchar gchar* escaped_key = ai_json_escape((gchar*)key);
if (custom_json->len > 0) {
g_string_append_c(custom_json, ',');
}
if (_is_json_bare_token((gchar*)value)) {
g_string_append_printf(custom_json, "\"%s\": %s", escaped_key, (gchar*)value);
} else {
/* non-scalar values must be quoted, else the payload is invalid JSON */
auto_gchar gchar* escaped_val = ai_json_escape((gchar*)value);
g_string_append_printf(custom_json, "\"%s\": \"%s\"", escaped_key, escaped_val);
}
}
}
pthread_mutex_unlock(&provider->settings_lock);
}
return g_string_free(custom_json, FALSE);
}
/**
* Assemble the payload for one API flavour from pre-serialized fragments:
* model, message list, custom settings, then fixed keys. Chat completions
* uses the "messages" list key; responses uses "input" plus "store":false
* unless a custom "store" setting overrides it.
*
* @param model The model identifier
* @param messages_json Fragment from _serialize_message_list()
* @param custom_json Fragment from _serialize_custom_settings()
* @param store_set Whether custom_json already carries a "store" value
* @param api_type The API flavour to build for (must not be AUTO)
* @return Newly allocated JSON string (caller must free)
*/
static gchar*
_assemble_json_payload(const gchar* model, const gchar* messages_json, const gchar* custom_json, gboolean store_set, AIApiType api_type)
{
auto_gchar gchar* escaped_model = ai_json_escape(model);
const gchar* list_key = api_type == AI_API_TYPE_CHAT_COMPLETIONS ? "messages" : "input";
const gchar* fixed_tail = (api_type == AI_API_TYPE_CHAT_COMPLETIONS || store_set)
? "\"stream\":false"
: "\"stream\":false,\"store\":false";
return g_strdup_printf("{\"model\":\"%s\",\"%s\":[%s],%s%s%s}",
escaped_model, list_key, messages_json,
custom_json, *custom_json ? "," : "", fixed_tail);
}
/* Chat completions: {"choices":[{"message":{"content":"..."}}]}. Anchor on
* "message" so a non-string "content" ordered earlier in the choice (e.g.
* "logprobs":{"content":[...]}) is not mistaken for the reply. */
static gchar*
_parse_chat_completions(const gchar* json)
{
const gchar* choices = strstr(json, "\"choices\"");
if (!choices)
return NULL;
/* Bound the search to the choices array so a "message"/"content" in a
* sibling object (e.g. a top-level "warning") is not taken as the reply */
const gchar* arr = strchr(choices, '[');
const gchar* end = arr ? _json_array_end(arr) : NULL;
const gchar* message = strstr(choices, "\"message\"");
if (message && end && message >= end)
message = NULL;
const gchar* anchor = message ? message : choices;
const gchar* cpos = _find_json_field(anchor, "content");
if (!cpos || (end && cpos >= end))
return NULL;
return _extract_json_string(anchor, "content");
}
/* TRUE when @from and @to belong to the same JSON object: a balanced nested
* sub-object between them is fine, but a brace that exits the enclosing object
* (net depth < 0) or leaves us inside a deeper one (net depth != 0) is not */
static gboolean
_same_json_object(const gchar* from, const gchar* to)
{
gboolean in_string = FALSE;
gint depth = 0;
for (const gchar* p = from; p < to && *p; p++) {
if (in_string) {
if (*p == '\\' && *(p + 1) != '\0')
p++;
else if (*p == '"')
in_string = FALSE;
} else if (*p == '"') {
in_string = TRUE;
} else if (*p == '{') {
depth++;
} else if (*p == '}') {
if (depth == 0)
return FALSE;
depth--;
}
}
return depth == 0;
}
/* Pointer to the '}' closing the object that encloses @pos (which must sit
* outside a string literal, e.g. at a key's opening quote), or NULL if
* unterminated. Braces inside string literals are ignored. */
static const gchar*
_json_object_end(const gchar* pos)
{
gboolean in_string = FALSE;
gint depth = 0;
for (const gchar* p = pos; *p; p++) {
if (in_string) {
if (*p == '\\' && *(p + 1) != '\0')
p++;
else if (*p == '"')
in_string = FALSE;
} else if (*p == '"') {
in_string = TRUE;
} else if (*p == '{') {
depth++;
} else if (*p == '}') {
if (depth == 0)
return p;
depth--;
}
}
return NULL;
}
/* Pointer just past the ']' matching the '[' at @arr, or NULL if unterminated.
* Brackets inside string literals are ignored. */
static const gchar*
_json_array_end(const gchar* arr)
{
gboolean in_string = FALSE;
gint depth = 0;
for (const gchar* p = arr; *p; p++) {
if (in_string) {
if (*p == '\\' && *(p + 1) != '\0')
p++;
else if (*p == '"')
in_string = FALSE;
} else if (*p == '"') {
in_string = TRUE;
} else if (*p == '[') {
depth++;
} else if (*p == ']') {
if (--depth == 0)
return p + 1;
}
}
return NULL;
}
/* Responses: {"output":[...,{"content":[{"type":"output_text","text":"..."}]}]}.
* "content" is an array here, so extract the inner "text" of the output_text
* part; reasoning items carry their own "text" (summary) ahead of it. */
static gchar*
_parse_responses(const gchar* json)
{
const gchar* output = strstr(json, "\"output\"");
if (!output)
return NULL;
const gchar* part = strstr(output, "\"output_text\"");
if (part) {
/* Bound the forward scan to the part's own object so a "text" from a
* later sibling item (e.g. a following reasoning summary) can never
* be returned as the reply */
const gchar* part_end = _json_object_end(part);
gchar* text = _extract_json_string_within(part, "text", part_end);
if (text)
return text;
/* "text" may precede "type":"output_text" in the part: take the
* nearest preceding "text" that shares the part's object, so an
* earlier item's text (a reasoning summary) is never returned; skip
* matches where extraction fails (e.g. a string value equal to "text") */
gsize window = part - output;
while (window > 0) {
const gchar* key = g_strrstr_len(output, window, "\"text\"");
if (!key)
break;
if (_same_json_object(key, part)) {
text = _extract_json_string_within(key, "text", part_end);
if (text)
return text;
}
window = key - output;
}
return NULL;
}
/* No output_text tag (truncated reply, or a server tagging the part
* "type":"text"): anchor strictly on the message "content" array. A
* reasoning-only body carries "summary", not "content", so returning NULL
* here surfaces a parse error instead of leaking the reasoning text */
const gchar* content_arr = strstr(output, "\"content\"");
if (!content_arr)
return NULL;
/* Bound the scan to the content array itself: an empty or text-less array
* must fail the parse rather than pick up a "text" from a later item */
const gchar* cval = _find_json_field(content_arr, "content");
if (!cval || *cval != '[')
return NULL;
return _extract_json_string_within(cval, "text", _json_array_end(cval));
}
/* Extract assistant content, trying the known request flavour's extractor
* first (AUTO means chat-completions first), then the other flavour, then a
* bare top-level "content" string. Each extractor runs at most once. */
gchar*
ai_parse_response_typed(AIApiType api_type, const gchar* response_json)
{
if (!response_json || strlen(response_json) == 0)
return NULL;
gchar* out;
if (api_type == AI_API_TYPE_RESPONSES) {
out = _parse_responses(response_json);
if (!out)
out = _parse_chat_completions(response_json);
} else {
out = _parse_chat_completions(response_json);
if (!out)
out = _parse_responses(response_json);
}
if (out)
return out;
/* No recognizable envelope: accept a bare top-level "content" string */
return _extract_json_string(response_json, "content");
}
gchar*
ai_parse_response(const gchar* response_json)
{
return ai_parse_response_typed(AI_API_TYPE_AUTO, response_json);
}
/* Extract the human-readable error.message from a provider error envelope:
* {"error":{"message":"...","type":"...","code":...}}
* Scoped to the substring after "error": so unrelated top-level "message"
* fields do not get picked up. */
gchar*
ai_parse_error_message(const gchar* error_json)
{
if (!error_json || strlen(error_json) == 0)
return NULL;
const gchar* error_obj = strstr(error_json, "\"error\":");
if (!error_obj)
return NULL;
return _extract_json_string(error_obj, "message");
}
static gpointer
_ai_request_thread(gpointer data)
{
log_debug("[AI-THREAD] Starting AI request thread");
/* Data is an array: [0]=session, [1]=prompt, [2]=user_data */
gpointer* args = (gpointer*)data;
AISession* session = (AISession*)args[0];
auto_gchar gchar* prompt = args[1];
gpointer user_data = args[2];
/* =====================================================================
* CRITICAL: Copy all session state under lock to prevent UAF races with
* cmd_ai_switch() and ai_session_clear_history() which mutate session
* fields on the main thread without holding a session ref.
*
* The worker holds a ref to the session (via ai_session_ref in
* ai_send_prompt), which prevents the session struct from being freed,
* but does NOT prevent field mutations (g_free/g_strdup of strings,
* list modifications). We must snapshot all fields atomically.
* ===================================================================== */
pthread_mutex_lock(&session->lock);
auto_gchar gchar* local_provider_name = g_strdup(session->provider_name);
AIProvider* local_provider = ai_provider_ref(session->provider);
auto_gchar gchar* local_model = g_strdup(session->model);
auto_gchar gchar* local_api_key = g_strdup(session->api_key ? session->api_key : "");
pthread_mutex_unlock(&session->lock);
log_debug("[AI-THREAD] Session: %s/%s", local_provider_name, local_model);
log_debug("[AI-THREAD] API key length: %zu", strlen(local_api_key));
/* Check for API key first */
if (!local_api_key || strlen(local_api_key) == 0) {
auto_gchar gchar* error_msg = g_strdup_printf("No API key set for provider '%s'. Use '/ai set token %s <key>' to configure.",
local_provider_name, local_provider_name);
log_error("AI request failed for %s/%s: %s", local_provider_name, local_model, error_msg);
_aiwin_display_error(user_data, error_msg);
ai_provider_unref(local_provider);
ai_session_unref(session);
g_free(args);
return NULL;
}
CURL* curl = curl_easy_init();
log_debug("[AI-THREAD] Curl initialized: %s", curl ? "OK" : "FAILED");
if (!curl) {
log_error("AI request failed for %s/%s: Failed to initialize curl",
local_provider_name, local_model);
_aiwin_display_error(user_data, "Failed to initialize curl.");
ai_provider_unref(local_provider);
ai_session_unref(session);
g_free(args);
return NULL;
}
/* Serialize the payload fragments once from a single history snapshot
* (under lock); per-flavour envelopes are assembled cheaply per attempt */
log_debug("[AI-THREAD] Building JSON payload...");
pthread_mutex_lock(&session->lock);
auto_gchar gchar* messages_json = _serialize_message_list(session->history, prompt);
/* Add user message to history now (it's in JSON but not yet in history) */
AIMessage* user_msg = g_new0(AIMessage, 1);
user_msg->role = g_strdup("user");
user_msg->content = g_strdup(prompt);
session->history = g_list_append(session->history, user_msg);
pthread_mutex_unlock(&session->lock);
log_debug("[AI-THREAD] Added user message to history");
gboolean store_set = FALSE;
auto_gchar gchar* custom_json = _serialize_custom_settings(local_provider, &store_set);
/* Build headers using shared helper */
struct curl_slist* headers = _build_curl_headers(local_provider, local_api_key);
/* Response buffer (allocated per attempt inside the loop) */
struct curl_response_t response;
response.data = NULL;
response.size = 0;
/* Snapshot the URL together with the epoch, so a flavour stamped after
* this request cannot be attributed to a URL it was not probed against */
gint api_epoch = 0;
auto_gchar gchar* base_url = _provider_base_url(local_provider, &api_epoch);
/* Flavour(s) to try: explicit setting wins; a previously detected flavour
* is a hint that keeps the other flavour as a fallback attempt (so a
* backend change self-corrects in-request); else probe responses first. */
AIApiType configured = g_atomic_int_get(&local_provider->api_type);
AIApiType attempts[2];
gint n_attempts = 0;
if (configured != AI_API_TYPE_AUTO) {
attempts[n_attempts++] = configured;
} else {
AIApiType resolved = g_atomic_int_get(&local_provider->resolved_api_type);
AIApiType first = resolved != AI_API_TYPE_AUTO ? resolved : AI_API_TYPE_RESPONSES;
attempts[n_attempts++] = first;
attempts[n_attempts++] = _other_flavour(first);
}
log_debug("[AI-THREAD] API URL: %s", base_url);
log_debug("[AI-THREAD] Model: %s", local_model);
/* Per-request options shared by all attempts */
curl_easy_setopt(curl, CURLOPT_HTTPHEADER, headers);
curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, _write_callback);
curl_easy_setopt(curl, CURLOPT_WRITEDATA, &response);
curl_easy_setopt(curl, CURLOPT_TIMEOUT, 60L);
CURLcode res = CURLE_OK;
long http_code = 0;
AIApiType used_type = attempts[0];
/* A payload rejection (400/422) that triggered the fallback is the
* actionable error; keep it in case the other flavour also fails */
auto_gchar gchar* first_reject_body = NULL;
long first_reject_code = 0;
/* Same for a 2xx we could not parse: the user must learn their endpoint
* answered, even when the fallback attempt fails with its own error */
auto_gchar gchar* first_unparsable_note = NULL;
auto_gchar gchar* content = NULL;
gboolean first_endpoint_missing = FALSE;
for (gint i = 0; i < n_attempts; i++) {
used_type = attempts[i];
auto_gchar gchar* request_url = g_strdup_printf("%s%s", base_url, _api_type_endpoint(used_type));
auto_gchar gchar* json_payload = _assemble_json_payload(local_model, messages_json, custom_json, store_set, used_type);
log_debug("[AI-THREAD] API Request URL: %s", request_url);
log_debug("[AI-THREAD] JSON payload: %s", json_payload);
/* Reset the response buffer for this attempt */
g_free(response.data);
response.data = g_new0(gchar, 1);
response.size = 0;
curl_easy_setopt(curl, CURLOPT_URL, request_url);
curl_easy_setopt(curl, CURLOPT_COPYPOSTFIELDS, json_payload);
res = curl_easy_perform(curl);
log_debug("[AI-THREAD] curl_easy_perform completed, res: %d", res);
http_code = 0;
curl_easy_getinfo(curl, CURLINFO_RESPONSE_CODE, &http_code);
log_debug("[AI-THREAD] HTTP response code: %ld", http_code);
if (i == 0 && res == CURLE_OK && _is_endpoint_missing(http_code, response.data))
first_endpoint_missing = TRUE;
gboolean more = (i + 1 < n_attempts);
/* Endpoint missing or payload-shape rejection: fall back to the other flavour */
if (res == CURLE_OK && more && _should_try_other_flavour(http_code, response.data)) {
log_info("[AI-THREAD] %s endpoint rejected request for '%s' (HTTP %ld), falling back to %s",
ai_api_type_to_string(used_type), local_provider_name, http_code,
ai_api_type_to_string(attempts[i + 1]));
log_debug("[AI-THREAD] Rejected %s response body: %s",
ai_api_type_to_string(used_type), response.data);
if (!first_reject_body && !_is_endpoint_missing(http_code, response.data)) {
first_reject_body = g_strdup(response.data);
first_reject_code = http_code;
}
continue;
}
/* Connectivity failure (reset/DNS/refused) on this flavour's endpoint:
* the other flavour may still be reachable, so try it before failing.
* A timeout is excluded: the request likely reached the server and may
* still be generating (and billing) — re-POSTing the conversation to
* the other endpoint could trigger a second generation */
if (res != CURLE_OK && res != CURLE_OPERATION_TIMEDOUT && more) {
log_info("[AI-THREAD] %s request for '%s' failed (%s), trying %s",
ai_api_type_to_string(used_type), local_provider_name,
curl_easy_strerror(res), ai_api_type_to_string(attempts[i + 1]));
continue;
}
/* Parse a success body here so a 2xx we cannot parse during an AUTO
* probe (e.g. a proxy 200 page) falls back to the other flavour
* instead of wedging on this endpoint */
if (res == CURLE_OK && http_code < 400) {
g_free(content);
content = ai_parse_response_typed(used_type, response.data);
if (!content && more && configured == AI_API_TYPE_AUTO
&& http_code >= 200 && http_code < 300) {
log_info("[AI-THREAD] %s returned an unparsable 2xx for '%s', trying %s",
ai_api_type_to_string(used_type), local_provider_name,
ai_api_type_to_string(attempts[i + 1]));
if (!first_unparsable_note)
first_unparsable_note = g_strdup_printf("%s endpoint answered HTTP %ld but the body could not be parsed",
ai_api_type_to_string(used_type), http_code);
continue;
}
}
break;
}
auto_gchar gchar* response_data = g_steal_pointer(&response.data);
if (configured == AI_API_TYPE_AUTO && res == CURLE_OK) {
if (_is_endpoint_missing(http_code, response_data)) {
/* The flavour we ended on is not served; re-probe on the next request */
_provider_set_resolved(local_provider, AI_API_TYPE_AUTO, api_epoch);
} else if (first_endpoint_missing) {
/* The first (hint/default) flavour's endpoint is gone but this one
* answered; prefer it as the hint so later requests stop probing
* the dead endpoint, even if this request itself failed (e.g. auth) */
_provider_set_resolved(local_provider, used_type, api_epoch);
}
}
if (res != CURLE_OK) {
/* The stashed first-attempt failure is usually the actionable part:
* that endpoint did answer, this one did not */
auto_gchar gchar* first_note = NULL;
if (first_reject_body) {
auto_gchar gchar* parsed = ai_parse_error_message(first_reject_body);
first_note = g_strdup_printf("%s endpoint rejected the request with HTTP %ld: %s",
ai_api_type_to_string(attempts[0]), first_reject_code,
parsed ? parsed : first_reject_body);
} else if (first_unparsable_note) {
first_note = g_strdup(first_unparsable_note);
}
auto_gchar gchar* error_msg = first_note
? g_strdup_printf("%s (first attempt: %s)", curl_easy_strerror(res), first_note)
: g_strdup(curl_easy_strerror(res));
log_error("AI request failed for %s/%s: %s", local_provider_name, local_model, error_msg);
_aiwin_display_error(user_data, error_msg);
} else if (http_code >= 400) {
/* The first attempt's 400/422 is usually the actionable error, but an
* auth/rate-limit/server failure on the second attempt (401/403/429/5xx)
* is more actionable — surface that rather than the stashed payload error */
gboolean second_more_actionable = http_code == 401 || http_code == 403
|| http_code == 429 || http_code >= 500;
gboolean use_first = first_reject_body && !second_more_actionable;
const gchar* err_body = use_first ? first_reject_body : response_data;
long err_code = use_first ? first_reject_code : http_code;
log_debug("[AI-THREAD] HTTP error response body (%zu bytes): %s",
response.size, response_data);
/* Try to extract the actual error message from the JSON response */
auto_gchar gchar* parsed_error = ai_parse_error_message(err_body);
auto_gchar gchar* error_msg = parsed_error ? g_strdup_printf("HTTP %ld: %s", err_code, parsed_error) : g_strdup_printf("HTTP %ld: %s", err_code, err_body && strlen(err_body) > 0 ? err_body : "Unknown error");
if (first_unparsable_note) {
gchar* with_note = g_strdup_printf("%s (first attempt: %s)", error_msg, first_unparsable_note);
g_free(error_msg);
error_msg = with_note;
}
log_error("AI request failed for %s/%s: %s", local_provider_name, local_model, error_msg);
_aiwin_display_error(user_data, error_msg);
} else {
log_debug("[AI-THREAD] Raw API response (%zu bytes): %s", response.size, response_data);
if (content) {
if (configured == AI_API_TYPE_AUTO && http_code >= 200 && http_code < 300) {
/* Remember the flavour that worked so later requests try it
* first; the epoch check discards the stamp if the URL or
* api-type changed while this request was in flight */
_provider_set_resolved(local_provider, used_type, api_epoch);
}
/* Add assistant response to history (under lock) */
pthread_mutex_lock(&session->lock);
AIMessage* assistant_msg = g_new0(AIMessage, 1);
assistant_msg->role = g_strdup("assistant");
assistant_msg->content = g_strdup(content);
session->history = g_list_append(session->history, assistant_msg);
pthread_mutex_unlock(&session->lock);
_aiwin_display_response(user_data, content);
} else {
/* A 2xx we cannot parse means the cached flavour may be wrong; drop
* it so the next request re-probes instead of wedging on it */
if (configured == AI_API_TYPE_AUTO)
_provider_set_resolved(local_provider, AI_API_TYPE_AUTO, api_epoch);
log_error("AI response parse failed for %s/%s: %.200s...",
local_provider_name, local_model, response_data);
_aiwin_display_error(user_data, "Failed to parse AI response.");
}
}
curl_slist_free_all(headers);
curl_easy_cleanup(curl);
ai_provider_unref(local_provider);
ai_session_unref(session);
g_free(args);
return NULL;
}
gboolean
ai_send_prompt(AISession* session, const gchar* prompt, gpointer user_data)
{
log_debug("[AI-PROMPT] ENTER: session=%p, prompt='%s', user_data=%p",
(void*)session, prompt, user_data);
if (!session || !prompt) {
log_error("[AI-PROMPT] FAIL: invalid session or prompt");
return FALSE;
}
/* Prepare thread arguments: [0]=session, [1]=prompt, [2]=user_data */
gpointer* args = g_new0(gpointer, 3);
args[0] = ai_session_ref(session);
args[1] = g_strdup(prompt);
args[2] = user_data;
log_debug("[AI-PROMPT] Prepared args, creating thread...");
GThread* thread = g_thread_new("ai_request", _ai_request_thread, args);
if (!thread) {
g_free(args[1]);
ai_session_unref(session);
g_free(args);
log_error("[AI-PROMPT] FAIL: g_thread_new returned NULL");
return FALSE;
}
log_debug("[AI-PROMPT] Thread created successfully: %p", (void*)thread);
g_thread_unref(thread);
return TRUE;
}