path: root/tests/test_provider.cpp

#include "provider.h"

#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include "config.h"
#include "dotagents.h"
#include "message.h"
#include "session.h"

static int failures = 0;
static int checks = 0;

#define CHECK(cond, msg)                                                     \
    do {                                                                     \
        checks++;                                                            \
        if (!(cond)) {                                                       \
            failures++;                                                      \
            fprintf(stderr, "FAIL %s:%d: %s\n", __FILE__, __LINE__, (msg)); \
        }                                                                    \
    } while (0)

#define CHECK_CONT(hay, needle, msg)                                         \
    do {                                                                     \
        checks++;                                                            \
        if (!hay || !strstr(hay, needle)) {                                  \
            failures++;                                                      \
            fprintf(stderr, "FAIL %s:%d: %s (want substring `%s`)\n",        \
                    __FILE__, __LINE__, (msg), (needle));                    \
        }                                                                    \
    } while (0)

#define CHECK_NOT_CONT(hay, needle, msg)                                     \
    do {                                                                     \
        checks++;                                                            \
        if (hay && strstr(hay, needle)) {                                    \
            failures++;                                                      \
            fprintf(stderr, "FAIL %s:%d: %s (found unwanted `%s`)\n",        \
                    __FILE__, __LINE__, (msg), (needle));                    \
        }                                                                    \
    } while (0)

static const char *SENTINEL = "SENTINEL-PREAMBLE-XYZ";

static void mock_destroy(llsh_model *m) { (void)m; }
static llsh_model MOCK = {NULL, NULL, mock_destroy};

static session *make_session(const char *provider) {
    config c;
    memset(&c, 0, sizeof c);
    c.provider = provider_parse(provider, NULL);
    c.model = xstrdup("test-model");
    c.preamble = xstrdup(SENTINEL);
    dotagents d;
    memset(&d, 0, sizeof d);
    return session_new_with_model(&MOCK, c, d);
}

/* The preamble must reach the wire for every provider. Before the fix it was
 * only emitted for Anthropic and silently dropped for OpenAI/Ollama. */
static void test_preamble_openai(void) {
    session *s = make_session("openai");
    char *body = provider_preview_request(s, "hello", true);
    CHECK(body != NULL, "openai body built");
    /* OpenAI carries the system prompt as a leading system message. */
    CHECK_CONT(body, "\"role\":\"system\"", "has a system message");
    CHECK_CONT(body, SENTINEL, "system message carries the preamble");
    CHECK_CONT(body, "\"role\":\"user\"", "user message still present");
    CHECK_CONT(body, "hello", "prompt present");
    free(body);
    session_free(s);
}

static void test_preamble_ollama(void) {
    session *s = make_session("ollama");
    char *body = provider_preview_request(s, "hello", true);
    CHECK(body != NULL, "ollama body built");
    /* Ollama carries the system prompt as a top-level `system` field. */
    CHECK_CONT(body, "\"system\":", "has a top-level system field");
    CHECK_CONT(body, SENTINEL, "system field carries the preamble");
    CHECK_CONT(body, "hello", "prompt present");
    free(body);
    session_free(s);
}

static void test_preamble_anthropic(void) {
    session *s = make_session("anthropic");
    char *body = provider_preview_request(s, "hello", true);
    CHECK(body != NULL, "anthropic body built");
    CHECK_CONT(body, "\"system\":", "has a system field");
    CHECK_CONT(body, SENTINEL, "system carries the preamble");
    free(body);
    session_free(s);
}

/* A model name with JSON-special characters must not corrupt the body. */
static void test_model_name_escaped(void) {
    config c;
    memset(&c, 0, sizeof c);
    c.provider = PROV_OPENAI;
    c.model = xstrdup("we\"ird/model");
    c.preamble = xstrdup(SENTINEL);
    dotagents d;
    memset(&d, 0, sizeof d);
    session *s = session_new_with_model(&MOCK, c, d);
    char *body = provider_preview_request(s, "hi", false);
    CHECK(body != NULL, "body built");
    /* The raw `"` in the model name must be escaped to \". */
    CHECK_CONT(body, "we\\\"ird/model", "model name quote is escaped");
    free(body);
    session_free(s);
}

static void test_no_preamble_omits_system(void) {
    config c;
    memset(&c, 0, sizeof c);
    c.provider = PROV_OPENAI;
    c.model = xstrdup("m");
    c.preamble = NULL; /* no preamble */
    dotagents d;
    memset(&d, 0, sizeof d);
    session *s = session_new_with_model(&MOCK, c, d);
    char *body = provider_preview_request(s, "hi", false);
    CHECK(body != NULL, "body built");
    CHECK_NOT_CONT(body, "\"role\":\"system\"", "no system message without preamble");
    free(body);
    session_free(s);
}

/* A turn stopped mid tool-loop leaves an assistant message with more tool
 * calls than the tool-result messages that follow it. The trim must drop the
 * surplus calls so the history stays a valid request. */
static void test_trim_dangling_calls(void) {
    message am = msg_assistant_text("working on it");
    msg_push_assistant_tool_call(am, "read_file", "{\"path\":\"a\"}");
    msg_push_assistant_tool_call(am, "write_file", "{\"path\":\"b\"}");
    message r1 = msg_user_tool_result("call-1", "read_file", "contents");
    message msgs[2] = {am, r1};

    CHECK(am.assistant().parts.size() == 3,
          "assistant has text + 2 calls before trim");
    provider_trim_dangling_calls(msgs, 2);
    int calls = 0;
    for (size_t i = 0; i < msgs[0].assistant().parts.size(); i++)
        if (std::holds_alternative<asst_tool_call_part>(
                msgs[0].assistant().parts[i]))
            calls++;
    CHECK(calls == 1, "surplus call trimmed to match the single result");

    /* Fully-answered calls must be untouched. */
    message am2 = msg_assistant_text("");
    msg_push_assistant_tool_call(am2, "read_file", "{}");
    message ra = msg_user_tool_result("c1", "read_file", "x");
    message rb = msg_user_tool_result("c2", "write_file", "ok");
    msg_push_assistant_tool_call(am2, "write_file", "{}");
    message msgs2[3] = {am2, ra, rb};
    provider_trim_dangling_calls(msgs2, 3);
    int calls2 = 0;
    for (size_t i = 0; i < msgs2[0].assistant().parts.size(); i++)
        if (std::holds_alternative<asst_tool_call_part>(
                msgs2[0].assistant().parts[i]))
            calls2++;
    CHECK(calls2 == 2, "complete call/result pair is preserved");
}

/* Reasoning models return the chain of thought in provider-specific fields.
 * None of the parsers extracted it before the fix, so it was never displayed. */
static void test_reasoning_openai(void) {
    const char *body =
        "{\"choices\":[{\"message\":{\"content\":\"42\","
        "\"reasoning_content\":\"think... think... 42\"}}]}";
    llm_conn conn; memset(&conn, 0, sizeof conn);
    conn.provider = LLM_OPENAI; conn.model = "m";
    llm_completion comp; memset(&comp, 0, sizeof comp);
    CHECK(llm_parse_body(&conn, body, &comp) == 0, "openai body parses");
    CHECK_CONT(comp.text, "42", "answer text extracted");
    CHECK(comp.reasoning != NULL, "reasoning extracted (was NULL before fix)");
    CHECK_CONT(comp.reasoning, "think...", "reasoning content preserved");
    llm_completion_free(&comp);
}

static void test_reasoning_ollama(void) {
    const char *body =
        "{\"message\":{\"content\":\"the answer\","
        "\"thinking\":\"pondering the answer...\"}}";
    llm_conn conn; memset(&conn, 0, sizeof conn);
    conn.provider = LLM_OLLAMA; conn.model = "m";
    llm_completion comp; memset(&comp, 0, sizeof comp);
    CHECK(llm_parse_body(&conn, body, &comp) == 0, "ollama body parses");
    CHECK_CONT(comp.text, "the answer", "answer text extracted");
    CHECK(comp.reasoning != NULL, "reasoning extracted from message.thinking");
    CHECK_CONT(comp.reasoning, "pondering", "reasoning content preserved");
    llm_completion_free(&comp);
}

static void test_reasoning_anthropic(void) {
    const char *body =
        "{\"content\":[{\"type\":\"thinking\",\"thinking\":\"hmm, let me "
        "reason...\"},{\"type\":\"text\",\"text\":\"done\"}]}";
    llm_conn conn; memset(&conn, 0, sizeof conn);
    conn.provider = LLM_ANTHROPIC; conn.model = "m";
    llm_completion comp; memset(&comp, 0, sizeof comp);
    CHECK(llm_parse_body(&conn, body, &comp) == 0, "anthropic body parses");
    CHECK_CONT(comp.text, "done", "answer text extracted");
    CHECK(comp.reasoning != NULL, "reasoning extracted from thinking block");
    CHECK_CONT(comp.reasoning, "reason", "reasoning content preserved");
    llm_completion_free(&comp);
}

/* A plain (non-reasoning) response must not fabricate a reasoning field. */
static void test_no_reasoning_when_absent(void) {
    const char *body =
        "{\"choices\":[{\"message\":{\"content\":\"just an answer\"}}]}";
    llm_conn conn; memset(&conn, 0, sizeof conn);
    conn.provider = LLM_OPENAI; conn.model = "m";
    llm_completion comp; memset(&comp, 0, sizeof comp);
    CHECK(llm_parse_body(&conn, body, &comp) == 0, "plain body parses");
    CHECK_CONT(comp.text, "just an answer", "answer text extracted");
    CHECK(comp.reasoning == NULL, "no reasoning field fabricated");
    llm_completion_free(&comp);
}

/* ------------------------------------------------------------------ */
/* sse_parse_line -- the pure streaming unit (AUDIT phase 2 test seam)  */
/* ------------------------------------------------------------------ */

/* OpenAI: text deltas, two interleaved multi-chunk tool calls,
 * finish_reason, [DONE], usage. */
static void test_sse_openai(void) {
    llm_sse_acc a;
    llm_sse_acc_init(&a, LLM_OPENAI);
    llm_sse_parse_line(LLM_OPENAI, "data: {\"choices\":[{\"delta\":{\"content\":\"Hel\"}}]}", &a);
    llm_sse_parse_line(LLM_OPENAI, "data: {\"choices\":[{\"delta\":{\"content\":\"lo\"}}]}", &a);
    CHECK(a.text.len == 5, "text deltas accumulate");
    CHECK_CONT(a.text.data, "Hello", "text content");
    llm_sse_parse_line(LLM_OPENAI,
                    "data: {\"choices\":[{\"delta\":{\"tool_calls\":["
                    "{\"index\":0,\"id\":\"c1\",\"function\":{\"name\":\"read_file\","
                    "\"arguments\":\"{\\\"pa\"}}]}}]}", &a);
    llm_sse_parse_line(LLM_OPENAI,
                    "data: {\"choices\":[{\"delta\":{\"tool_calls\":["
                    "{\"index\":1,\"id\":\"c2\",\"function\":{\"name\":\"grep\","
                    "\"arguments\":\"{\\\"pat\"}}]}}]}", &a);
    llm_sse_parse_line(LLM_OPENAI,
                    "data: {\"choices\":[{\"delta\":{\"tool_calls\":["
                    "{\"index\":0,\"function\":{\"arguments\":\"th\\\":\\\"a.txt\\\"}\"}}]}}]}",
                    &a);
    llm_sse_parse_line(LLM_OPENAI,
                    "data: {\"choices\":[{\"delta\":{\"tool_calls\":["
                    "{\"index\":1,\"function\":{\"arguments\":\"tern\\\":\\\"x\\\"}\"}}]}}]}",
                    &a);
    CHECK(a.tc_len == 2, "two tool calls indexed");
    CHECK(a.tc_id[0] && strcmp(a.tc_id[0], "c1") == 0, "call 0 id");
    CHECK(a.tc_id[1] && strcmp(a.tc_id[1], "c2") == 0, "call 1 id");
    CHECK(a.tc_name[0] && strcmp(a.tc_name[0], "read_file") == 0, "call 0 name");
    CHECK_CONT(a.tc_args[0].data, "\"path\":\"a.txt\"", "call 0 args assembled");
    CHECK_CONT(a.tc_args[1].data, "\"pattern\":\"x\"", "call 1 args assembled");
    llm_sse_parse_line(LLM_OPENAI,
                    "data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"tool_calls\"}]}",
                    &a);
    CHECK(a.done, "finish_reason marks the stream done");
    llm_sse_parse_line(LLM_OPENAI, "data: [DONE]", &a);
    llm_sse_parse_line(LLM_OPENAI, "data: {\"usage\":{\"prompt_tokens\":42}}", &a);
    CHECK(a.input_tokens == 42, "usage.prompt_tokens captured");
    llm_sse_acc_free(&a);
}

/* Anthropic: event:/data: pairing, thinking + text blocks, a tool_use
 * block with chunked input_json_delta, message_start usage, message_stop. */
static void test_sse_anthropic(void) {
    llm_sse_acc a;
    llm_sse_acc_init(&a, LLM_ANTHROPIC);
    llm_sse_parse_line(LLM_ANTHROPIC, "event: message_start", &a);
    llm_sse_parse_line(LLM_ANTHROPIC,
                    "data: {\"message\":{\"usage\":{\"input_tokens\":7}}}", &a);
    CHECK(a.input_tokens == 7, "usage from message_start");
    llm_sse_parse_line(LLM_ANTHROPIC, "event: content_block_start", &a);
    llm_sse_parse_line(LLM_ANTHROPIC,
                    "data: {\"content_block\":{\"type\":\"thinking\"}}", &a);
    llm_sse_parse_line(LLM_ANTHROPIC, "event: content_block_delta", &a);
    llm_sse_parse_line(LLM_ANTHROPIC,
                    "data: {\"delta\":{\"type\":\"thinking_delta\",\"thinking\":\"hmm \"}}",
                    &a);
    llm_sse_parse_line(LLM_ANTHROPIC, "event: content_block_start", &a);
    llm_sse_parse_line(LLM_ANTHROPIC,
                    "data: {\"content_block\":{\"type\":\"text\"}}", &a);
    llm_sse_parse_line(LLM_ANTHROPIC, "event: content_block_delta", &a);
    llm_sse_parse_line(LLM_ANTHROPIC,
                    "data: {\"delta\":{\"type\":\"text_delta\",\"text\":\"ok\"}}", &a);
    CHECK_CONT(a.reasoning.data, "hmm", "thinking_delta captured");
    CHECK_CONT(a.text.data, "ok", "text_delta captured");
    llm_sse_parse_line(LLM_ANTHROPIC, "event: content_block_start", &a);
    llm_sse_parse_line(LLM_ANTHROPIC,
                    "data: {\"content_block\":{\"type\":\"tool_use\",\"id\":\"t1\","
                    "\"name\":\"write_file\"}}", &a);
    llm_sse_parse_line(LLM_ANTHROPIC, "event: content_block_delta", &a);
    llm_sse_parse_line(LLM_ANTHROPIC,
                    "data: {\"delta\":{\"type\":\"input_json_delta\","
                    "\"partial_json\":\"{\\\"pa\"}}", &a);
    llm_sse_parse_line(LLM_ANTHROPIC,
                    "data: {\"delta\":{\"type\":\"input_json_delta\","
                    "\"partial_json\":\"th\\\":\\\"x\\\"}\"}}", &a);
    CHECK(a.tc_len == 1, "one tool_use block");
    CHECK(a.tc_id[0] && strcmp(a.tc_id[0], "t1") == 0, "tool_use id");
    CHECK(a.tc_name[0] && strcmp(a.tc_name[0], "write_file") == 0, "tool_use name");
    CHECK_CONT(a.tc_args[0].data, "\"path\":\"x\"", "chunked partial_json assembled");
    llm_sse_parse_line(LLM_ANTHROPIC, "event: message_stop", &a);
    llm_sse_parse_line(LLM_ANTHROPIC, "data: [DONE]", &a);
    CHECK(a.done, "message_stop marks done");
    llm_sse_acc_free(&a);
}

/* Ollama NDJSON: content + thinking chunks, prompt_eval_count, tool_calls
 * on the done chunk (object arguments), done:true. */
static void test_sse_ollama(void) {
    llm_sse_acc a;
    llm_sse_acc_init(&a, LLM_OLLAMA);
    llm_sse_parse_line(LLM_OLLAMA, "{\"message\":{\"content\":\"He\"}}", &a);
    llm_sse_parse_line(LLM_OLLAMA, "{\"message\":{\"thinking\":\"think\"}}", &a);
    llm_sse_parse_line(LLM_OLLAMA, "{\"message\":{\"content\":\"llo\"}}", &a);
    CHECK_CONT(a.text.data, "Hello", "content chunks accumulate");
    CHECK_CONT(a.reasoning.data, "think", "thinking chunk captured");
    llm_sse_parse_line(LLM_OLLAMA, "{\"prompt_eval_count\":9}", &a);
    CHECK(a.input_tokens == 9, "prompt_eval_count captured");
    llm_sse_parse_line(LLM_OLLAMA,
                    "{\"message\":{\"tool_calls\":[{\"id\":\"o1\",\"function\":"
                    "{\"name\":\"list_dir\",\"arguments\":{\"path\":\".\"}}}]},"
                    "\"done\":true}", &a);
    CHECK(a.tc_len == 1, "tool call on the done chunk");
    CHECK(a.tc_id[0] && strcmp(a.tc_id[0], "o1") == 0, "tool id");
    CHECK(a.tc_name[0] && strcmp(a.tc_name[0], "list_dir") == 0, "tool name");
    CHECK_CONT(a.tc_args[0].data, "\"path\":\".\"", "object args serialized");
    CHECK(a.done, "done:true marks done");
    llm_sse_acc_free(&a);
}

/* Lines outside the protocol (stray text, bad JSON) are ignored without
 * corrupting the accumulator. */
static void test_sse_noise_lines(void) {
    llm_sse_acc a;
    llm_sse_acc_init(&a, LLM_OPENAI);
    llm_sse_parse_line(LLM_OPENAI, "event: ping", &a); /* not `data:` -> ignored */
    llm_sse_parse_line(LLM_OPENAI, "data: not-json", &a); /* bad JSON -> ignored */
    llm_sse_parse_line(LLM_OPENAI, "data: {\"choices\":[]}", &a); /* empty choices */
    CHECK(a.text.len == 0 && a.tc_len == 0 && !a.done, "noise left acc empty");
    llm_sse_acc_free(&a);

    llm_sse_acc b;
    llm_sse_acc_init(&b, LLM_ANTHROPIC);
    llm_sse_parse_line(LLM_ANTHROPIC, "data: {\"no\":\"event\"}", &b); /* no event set */
    llm_sse_parse_line(LLM_ANTHROPIC, "data: {", &b); /* truncated JSON */
    CHECK(b.text.len == 0 && b.tc_len == 0 && !b.done, "anthropic noise ignored");
    llm_sse_acc_free(&b);

    llm_sse_acc c;
    llm_sse_acc_init(&c, LLM_OLLAMA);
    llm_sse_parse_line(LLM_OLLAMA, "hello", &c); /* not JSON */
    llm_sse_parse_line(LLM_OLLAMA, "{}", &c); /* no message field */
    CHECK(c.text.len == 0 && c.tc_len == 0 && !c.done, "ollama noise ignored");
    llm_sse_acc_free(&c);
}

/* sse_acc_into: unnamed/unargued calls get default ids and `{}` args. */
static void test_sse_acc_into_defaults(void) {
    llm_sse_acc a;
    llm_sse_acc_init(&a, LLM_OLLAMA);
    llm_acc_push_tool(&a, NULL, NULL);
    llm_completion c;
    memset(&c, 0, sizeof c);
    llm_sse_acc_into(&a, &c);
    CHECK(c.tc_len == 1, "one materialized call");
    CHECK(c.tc_id[0] && strstr(c.tc_id[0], "call-1") != NULL, "default id");
    CHECK(c.tc_args[0] && strcmp(c.tc_args[0], "{}") == 0, "default empty args");
    llm_completion_free(&c);
    llm_sse_acc_free(&a);
}

/* sse_acc_into: non-empty text/reasoning transfer ownership to `c` and must
 * survive completion_free + sse_acc_free without a double-free. Regression:
 * the old code freed the stolen buffer right after handing it to `c`,
 * dangling c->text/c->reasoning (crashed on real streamed responses). */
static void test_sse_acc_into_text_reasoning_ownership(void) {
    llm_sse_acc a;
    llm_sse_acc_init(&a, LLM_OPENAI);
    llm_buf_put(&a.text, "the answer is 42");
    llm_buf_put(&a.reasoning, "pondering deeply");
    llm_acc_push_tool(&a, xstrdup("call-9"), xstrdup("read_file"));
    llm_buf_put(&a.tc_args[0], "\"path\":\"AUDIT.md\"");

    llm_completion c;
    memset(&c, 0, sizeof c);
    llm_sse_acc_into(&a, &c);
    CHECK(c.text && strcmp(c.text, "the answer is 42") == 0,
          "text transferred intact");
    CHECK(c.reasoning && strcmp(c.reasoning, "pondering deeply") == 0,
          "reasoning transferred intact");
    CHECK(c.tc_len == 1 && c.tc_args[0] &&
              strcmp(c.tc_args[0], "\"path\":\"AUDIT.md\"") == 0,
           "tool args transferred intact");
    /* Both frees must be safe: `c` owns the strings now, `a` is reset. */
    llm_completion_free(&c);
    llm_sse_acc_free(&a);
}

/* ================= llm_params -> request body =================
 * The `llm` tool threads llm_params (temperature, max_tokens, reasoning
 * effort, thinking budget, response schema) into the wire body. These are
 * pure request-builder tests (no network) per provider. */

static llm_message one_user(const char *text) {
    static llm_part part;
    memset(&part, 0, sizeof part);
    part.kind = LLM_TEXT;
    part.text = text;
    llm_message m;
    memset(&m, 0, sizeof m);
    m.role = LLM_ROLE_USER;
    m.parts = &part;
    m.part_count = 1;
    return m;
}

static void test_params_openai(void) {
    llm_conn c;
    memset(&c, 0, sizeof c);
    c.provider = LLM_OPENAI;
    c.model = "test-model";
    llm_params p;
    llm_params_init(&p);
    p.temperature = 0.5;
    p.max_tokens = 256;
    p.reasoning_effort = "high";
    p.response_schema = "{\"type\":\"object\"}";

    llm_message m = one_user("hi");
    char *body = llm_build_request_ex(&c, NULL, &m, 1, NULL, false, &p);
    CHECK(body != NULL, "openai body built");
    if (body) {
        CHECK_CONT(body, "\"temperature\":0.5", "openai temperature");
        CHECK_CONT(body, "\"max_tokens\":256", "openai max_tokens");
        CHECK_CONT(body, "\"reasoning_effort\":\"high\"",
                   "openai reasoning_effort");
        CHECK_CONT(body, "\"response_format\":{\"type\":\"json_schema\"",
                   "openai response_format");
        CHECK_CONT(body, "\"type\":\"object\"", "openai schema inlined");
        free(body);
    }
}

static void test_params_ollama(void) {
    llm_conn c;
    memset(&c, 0, sizeof c);
    c.provider = LLM_OLLAMA;
    c.model = "ollama-model";
    llm_params p;
    llm_params_init(&p);
    p.temperature = 0.75; /* 3/4: exact in binary, so %.6g is stable */
    p.max_tokens = 128;
    p.reasoning_effort = "medium";
    p.response_schema = "{\"type\":\"object\"}";

    llm_message m = one_user("hi");
    char *body = llm_build_request_ex(&c, "SYS", &m, 1, NULL, false, &p);
    CHECK(body != NULL, "ollama body built");
    if (body) {
        CHECK_CONT(body, "\"system\":\"SYS\"", "ollama top-level system");
        CHECK_CONT(body, "\"options\":{\"temperature\":0.75",
                   "ollama options.temperature");
        CHECK_CONT(body, "\"num_predict\":128", "ollama num_predict");
        CHECK_CONT(body, "\"think\":true", "ollama think enabled");
        CHECK_CONT(body, "\"format\":", "ollama format field");
        CHECK_CONT(body, "\"type\":\"object\"", "ollama schema");
        free(body);
    }
}

static void test_params_anthropic_thinking(void) {
    llm_conn c;
    memset(&c, 0, sizeof c);
    c.provider = LLM_ANTHROPIC;
    c.model = "claude";
    llm_params p;
    llm_params_init(&p);
    p.thinking_budget = 16384;

    llm_message m = one_user("hi");
    char *body = llm_build_request_ex(&c, NULL, &m, 1, NULL, false, &p);
    CHECK(body != NULL, "anthropic body built");
    if (body) {
        CHECK_CONT(body, "\"thinking\":{\"type\":\"enabled\","
                         "\"budget_tokens\":16384}",
                   "anthropic thinking budget");
        /* max_tokens must be bumped above the thinking budget. */
        CHECK_CONT(body, "\"max_tokens\":17408",
                   "anthropic max_tokens > budget");
        free(body);
    }
}

static void test_params_anthropic_effort(void) {
    llm_conn c;
    memset(&c, 0, sizeof c);
    c.provider = LLM_ANTHROPIC;
    c.model = "claude";
    llm_params p;
    llm_params_init(&p);
    p.reasoning_effort = "low"; /* maps to a 1024-token budget */

    llm_message m = one_user("hi");
    char *body = llm_build_request_ex(&c, NULL, &m, 1, NULL, false, &p);
    CHECK(body != NULL, "anthropic body built");
    if (body)
        CHECK_CONT(body, "\"budget_tokens\":1024",
                   "effort 'low' maps to a 1024 budget");
    free(body);
}

static void test_params_anthropic_temperature(void) {
    llm_conn c;
    memset(&c, 0, sizeof c);
    c.provider = LLM_ANTHROPIC;
    c.model = "claude";
    llm_params p;
    llm_params_init(&p);
    p.temperature = 0.25; /* 1/4: exact in binary, no thinking budget */

    llm_message m = one_user("hi");
    char *body = llm_build_request_ex(&c, NULL, &m, 1, NULL, false, &p);
    CHECK(body != NULL, "anthropic body built");
    if (body) {
        CHECK_CONT(body, "\"temperature\":0.25", "anthropic temperature");
        CHECK_NOT_CONT(body, "\"thinking\"",
                       "no thinking block without a budget");
        free(body);
    }
}

static void test_params_anthropic_schema(void) {
    llm_conn c;
    memset(&c, 0, sizeof c);
    c.provider = LLM_ANTHROPIC;
    c.model = "claude";
    llm_params p;
    llm_params_init(&p);
    p.response_schema = "{\"type\":\"object\"}";

    llm_message m = one_user("hi");
    char *body = llm_build_request_ex(&c, "SYS", &m, 1, NULL, false, &p);
    CHECK(body != NULL, "anthropic body built");
    if (body) {
        CHECK_CONT(body, "SYS", "preamble kept");
        CHECK_CONT(body, "JSON Schema", "schema instruction appended");
        /* The schema is JSON-encoded into the system string, so its quotes
         * are escaped in the body. */
        CHECK_CONT(body, "\\\"type\\\"", "schema text embedded (escaped)");
        CHECK_NOT_CONT(body, "response_format",
                       "anthropic has no response_format");
        free(body);
    }
}

static void test_params_unset(void) {
    llm_conn c;
    memset(&c, 0, sizeof c);
    c.provider = LLM_OPENAI;
    c.model = "m";

    llm_message m = one_user("hi");
    char *body = llm_build_request_ex(&c, NULL, &m, 1, NULL, false, NULL);
    CHECK(body != NULL, "body built with NULL params");
    if (body) {
        CHECK_NOT_CONT(body, "temperature", "no temperature when unset");
        CHECK_NOT_CONT(body, "max_tokens", "no max_tokens when unset");
        CHECK_NOT_CONT(body, "response_format",
                       "no response_format when unset");
        free(body);
    }
}

/* ================= `compact` built-in tool =================
 * session_force_compact is the dispatch target for the `compact` tool call. */

static void test_force_compact_short(void) {
    session *s = make_session("openai");
    session_push_history(s, msg_user_text("hello"));
    session_push_history(s, msg_assistant_text("hi"));
    compact_report rep;
    memset(&rep, 0, sizeof rep);
    char *err = NULL;
    int rc = session_force_compact(s, &rep, &err);
    /* user+assistant only: keep-boundary at index 0 -> no pass can run. */
    CHECK(rc == 0, "no pass on a too-short history");
    free(err);
    session_free(s);
}

static void test_force_compact_clears(void) {
    session *s = make_session("openai");
    session_push_history(s, msg_user_text("read the file"));
    session_push_history(s, msg_assistant_tool_call("read_file", "{}"));
    session_push_history(s, msg_user_tool_result("c1", "read_file",
                                                 "a lot of cleared content"));
    session_push_history(s, msg_assistant_text("done reading"));
    session_push_history(s, msg_user_text("now edit it"));

    compact_report rep;
    memset(&rep, 0, sizeof rep);
    char *err = NULL;
    int rc = session_force_compact(s, &rep, &err);
    CHECK(rc == 1, "compaction pass ran");
    CHECK(rep.tier == TIER_FALLBACK_CLEAR,
          "no summarizer -> fallback clear tier");
    CHECK(rep.cleared_tool_results >= 1, "stale tool result cleared");
    free(err);
    session_free(s);
}

int main(void) {
    test_preamble_openai();
    test_preamble_ollama();
    test_preamble_anthropic();
    test_model_name_escaped();
    test_no_preamble_omits_system();
    test_trim_dangling_calls();
    test_reasoning_openai();
    test_reasoning_ollama();
    test_reasoning_anthropic();
    test_no_reasoning_when_absent();
    test_sse_openai();
    test_sse_anthropic();
    test_sse_ollama();
    test_sse_noise_lines();
    test_sse_acc_into_defaults();
    test_sse_acc_into_text_reasoning_ownership();
    test_params_openai();
    test_params_ollama();
    test_params_anthropic_thinking();
    test_params_anthropic_effort();
    test_params_anthropic_temperature();
    test_params_anthropic_schema();
    test_params_unset();
    test_force_compact_short();
    test_force_compact_clears();
    printf("provider: %d checks, %d failures\n", checks, failures);
    return failures ? 1 : 0;
}