CST-agent/tests/test_artifacts.c

/* deliver_file: a finished file handed to the person, not to the model.
 *
 * The three things that must hold:
 *   - the stored copy is the file itself (any bytes, its own extension and a
 *     mime taken from the name), not the text reduction an attachment gets;
 *   - it carries `for_user`, so no request ever contains it -- and a document
 *     attachment on the same session still travels;
 *   - the tool is in the engine's registry, takes a workspace path and refuses
 *     a directory or a path that leaves the workspace.
 */
#define _POSIX_C_SOURCE 200809L

#include "cJSON.h"
#include "cst_agent.h"
#include "cst_attach.h"
#include "cst_test_util.h"
#include "cst_tools.h"
#include "os_compat.h"

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

static int nfail;
static char g_ws[1024];

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

static void wfile(const char *rel, const void *data, size_t n) {
  char p[1200];
  snprintf(p, sizeof(p), "%s/%s", g_ws, rel);
  FILE *f = fopen(p, "wb");
  if (!f) {
    fprintf(stderr, "cannot write %s\n", p);
    exit(1);
  }
  if (n && fwrite(data, 1, n, f) != n) {
    fprintf(stderr, "short write %s\n", p);
    exit(1);
  }
  fclose(f);
}

static void full_path(char *out, size_t n, const char *rel) {
  snprintf(out, n, "%s/%s", g_ws, rel);
}

/* name, then key/value pairs, NULL-terminated (the ctl_tool_call_t a turn
 * builds from a model's call). */
static ctl_tool_call_t mk(const char *name, ...) {
  ctl_tool_call_t c;
  memset(&c, 0, sizeof(c));
  snprintf(c.id, sizeof(c.id), "t");
  snprintf(c.name, sizeof(c.name), "%s", name);
  va_list ap;
  va_start(ap, name);
  const char *k;
  while ((k = va_arg(ap, const char *)) != NULL) {
    const char *v = va_arg(ap, const char *);
    ctl_arg_t *a = &c.args[c.nargs++];
    snprintf(a->key, sizeof(a->key), "%s", k);
    a->value = (char *)v;
    a->value_len = strlen(v);
  }
  va_end(ap);
  return c;
}

static void test_registry(void) {
  const cst_tool_def_t *t = cst_tool_find("deliver_file");
  EXPECT(t != NULL, "deliver_file is in the registry");
  if (!t)
    return;
  char why[160] = "";
  EXPECT(cst_tool_available(t, why, sizeof(why)) == 0,
         "deliver_file is available here");
  /* It copies out of the workspace and never writes it: a read. */
  EXPECT(strcmp(cst_tool_perm_name(t), "read") == 0,
         "deliver_file is a read-policy tool");
  EXPECT(t->owner == NULL, "deliver_file is an engine tool");
}

static void test_import(const char *session) {
  const char *md = "# Report\n\n- one\n- two\n";
  wfile("out/report.md", md, strlen(md));
  char src[1200];
  full_path(src, sizeof(src), "out/report.md");
  char err[256] = "";
  cst_attachment_t *a =
      cst_attach_import_artifact(session, src, NULL, err, sizeof(err));
  EXPECT(a != NULL, "a markdown file is delivered");
  if (!a) {
    fprintf(stderr, "  (%s)\n", err);
    return;
  }
  EXPECT(strcmp(a->mime, "text/markdown") == 0, "mime comes from the name");
  EXPECT(strcmp(a->name, "report.md") == 0, "display name is the file name");
  EXPECT(a->for_user == 1, "marked for the user");
  EXPECT(!cst_attach_is_text(a), "a delivered file is not inlined as text");
  EXPECT(!cst_attach_is_image(a), "and it is not an image part either");
  EXPECT(strncmp(a->rel, "art-", 4) == 0, "stored under an artifact id");
  size_t rl = strlen(a->rel);
  EXPECT(rl > 4 && strcmp(a->rel + rl - 3, ".md") == 0,
         "the stored copy keeps the extension");
  EXPECT(a->bytes == (long long)strlen(md), "size is the file's size");

  size_t n = 0;
  char *back = cst_attach_read(session, a, &n, err, sizeof(err));
  EXPECT(back && n == strlen(md) && memcmp(back, md, n) == 0,
         "the payload is byte for byte");
  free(back);

  /* The copy lives beside the session, not in the workspace. */
  char stored[1200];
  EXPECT(cst_attach_full_path(session, a, stored, sizeof(stored)) == 0,
         "stored path resolves");
  os_stat_t st;
  EXPECT(os_lstat(stored, &st) == 0 && st.size == (long long)strlen(md),
         "the stored file is on disk");

  /* A title wins over the file name, and its extension decides the type. */
  cst_attachment_t *b = cst_attach_import_artifact(session, src, "summary.html",
                                                   err, sizeof(err));
  EXPECT(b && strcmp(b->name, "summary.html") == 0, "title is the display name");
  EXPECT(b && strcmp(b->mime, "text/html") == 0, "the title's extension decides");
  cst_attach_free(b);
  cst_attach_free(a);
}

static void test_any_bytes(const char *session) {
  /* A zip is bytes, not text: an attachment would refuse it (the content
   * decides), a delivered file is copied as it is. */
  char zip[512];
  for (size_t i = 0; i < sizeof(zip); i++)
    zip[i] = (char)(i % 251);
  memcpy(zip, "PK\003\004", 4);
  wfile("out/build.zip", zip, sizeof(zip));
  char src[1200];
  full_path(src, sizeof(src), "out/build.zip");

  char err[256] = "";
  cst_attachment_t *a =
      cst_attach_import_artifact(session, src, NULL, err, sizeof(err));
  EXPECT(a != NULL, "a binary file is delivered");
  EXPECT(a && strcmp(a->mime, "application/zip") == 0, "zip mime");
  if (a) {
    size_t n = 0;
    char *back = cst_attach_read(session, a, &n, err, sizeof(err));
    EXPECT(back && n == sizeof(zip) && memcmp(back, zip, n) == 0,
           "binary payload survives");
    free(back);
    cst_attach_free(a);
  }

  /* The same file through the attachment path: refused, as before. */
  cst_attachment_t *as_attach =
      cst_attach_import_file(session, src, err, sizeof(err));
  EXPECT(as_attach == NULL, "an attachment still refuses binary garbage");
  cst_attach_free(as_attach);

  /* Bigger than an attachment may be (16 MB) but inside the artifact cap: a
   * report is allowed to be large. */
  char big[1200];
  full_path(big, sizeof(big), "out/big.log");
  FILE *f = fopen(big, "wb");
  if (f) {
    char line[64];
    memset(line, 'x', sizeof(line) - 1);
    line[sizeof(line) - 1] = '\n';
    /* 17 MB: past the 16 MB attachment cap, well inside the artifact one. */
    for (int i = 0; i < 17 * 1024 * 1024 / (int)sizeof(line); i++)
      fwrite(line, 1, sizeof(line), f);
    fclose(f);
  }
  os_stat_t st;
  EXPECT(os_lstat(big, &st) == 0 && st.size > CST_ATTACH_MAX_BYTES,
         "the scratch file is larger than the attachment cap");
  cst_attachment_t *large =
      cst_attach_import_artifact(session, big, NULL, err, sizeof(err));
  EXPECT(large != NULL, "a file past the attachment cap is delivered");
  EXPECT(large && large->bytes == (long long)st.size, "all of it is stored");
  EXPECT(large && strcmp(large->mime, "text/plain") == 0, "log mime");
  cst_attach_free(large);
  cst_attachment_t *small =
      cst_attach_import_file(session, big, err, sizeof(err));
  EXPECT(small == NULL, "the attachment path still caps at 16 MB");
  cst_attach_free(small);
}

static void test_json(const char *session) {
  char src[1200];
  full_path(src, sizeof(src), "out/report.md");
  char err[256] = "";
  cst_attachment_t *a =
      cst_attach_import_artifact(session, src, NULL, err, sizeof(err));
  EXPECT(a != NULL, "artifact for the json round trip");
  if (!a)
    return;
  char *json = cst_attachments_to_json(a, 1);
  EXPECT(json && strstr(json, "\"for_user\":1") != NULL,
         "the JSONL says it is the user's");
  cJSON *parsed = json ? cJSON_Parse(json) : NULL;
  int n = 0;
  cst_attachment_t *back = parsed ? cst_attachments_from_json(parsed, &n) : NULL;
  EXPECT(back && n == 1, "one entry parses back");
  if (back) {
    EXPECT(back[0].for_user == 1, "for_user survives the JSONL");
    EXPECT(strcmp(back[0].mime, "text/markdown") == 0, "mime survives");
    EXPECT(strcmp(back[0].rel, a->rel) == 0, "stored name survives");
    cst_attachments_free(back, n);
  }
  cJSON_Delete(parsed);
  free(json);
  cst_attach_free(a);
}

/* The session file is what a replayed transcript is read from: a delivered
 * file that loses `for_user` on the way through it would be handed back to the
 * model (and never shown as a card) on the next start. */
static void test_session_file(void) {
  char src[1200], path[1200];
  full_path(src, sizeof(src), "out/report.md");
  snprintf(path, sizeof(path), "%s/delivered.jsonl", g_ws);
  char err[256] = "";
  cst_attachment_t *a =
      cst_attach_import_artifact(path, src, NULL, err, sizeof(err));
  EXPECT(a != NULL, "the file is stored for a session message");
  cst_session_t *s = cst_session_open(path);
  EXPECT(s != NULL, "the session opens");
  if (!s) {
    cst_attach_free(a);
    return;
  }
  cst_message_t m;
  memset(&m, 0, sizeof(m));
  m.role = CST_ROLE_TOOL;
  m.content = (char *)"delivered report.md";
  snprintf(m.tool_call_id, sizeof(m.tool_call_id), "c1");
  snprintf(m.tool_name, sizeof(m.tool_name), "deliver_file");
  m.attach = a;
  m.nattach = a ? 1 : 0;
  EXPECT(cst_session_append(s, &m) == 0, "the tool result is appended");
  cst_session_close(s);
  cst_attach_free(a);

  /* Read it back the way cst_agent_open() does. */
  s = cst_session_open(path);
  EXPECT(s && cst_session_count(s) == 1, "the message is in the file");
  const cst_message_t *back = s ? cst_session_get(s, 0) : NULL;
  EXPECT(back && back->nattach == 1, "its attachment comes back");
  EXPECT(back && back->attach && back->attach[0].for_user == 1,
         "and is still the user's, not the model's");
  EXPECT(back && back->attach && !cst_attach_is_text(&back->attach[0]),
         "so it is never inlined");
  EXPECT(back && back->attach && !cst_attach_is_image(&back->attach[0]),
         "and it is not an image part either");
  cst_session_close(s);
}

/* The tool through the registry, the way a turn calls it. */
static void test_tool(const char *session) {
  cst_agent_config_t cfg;
  memset(&cfg, 0, sizeof(cfg));
  snprintf(cfg.workspace_root, sizeof(cfg.workspace_root), "%s", g_ws);
  snprintf(cfg.session_path, sizeof(cfg.session_path), "%s", session);
  cst_agent_t *ag = cst_agent_create(&cfg);
  EXPECT(ag != NULL, "agent for the tool call");
  if (!ag)
    return;

  cst_tool_ctx_t ctx;
  memset(&ctx, 0, sizeof(ctx));
  ctx.workspace_root = g_ws;
  ctx.agent = ag;

  cst_tool_result_t r;
  memset(&r, 0, sizeof(r));
  ctl_tool_call_t c = mk("deliver_file", "path", "out/report.md", NULL);
  cst_tool_run(&ctx, &c, &r); /* by name, through the registry */
  EXPECT(r.exit_code == 0, "deliver_file runs");
  EXPECT(r.out && strstr(r.out, "delivered report.md") != NULL,
         "the result names the file");
  EXPECT(r.nattach == 1 && r.attach && r.attach[0].for_user == 1,
         "the tool hands the file to the front end");
  EXPECT(r.attach && !cst_attach_is_text(&r.attach[0]),
         "and never to a request");
  cst_tool_result_free(&r);

  memset(&r, 0, sizeof(r));
  c = mk("deliver_file", "path", "out", NULL); /* a directory */
  cst_tool_run(&ctx, &c, &r);
  EXPECT(r.exit_code == 1 && r.out && strstr(r.out, "is a directory") != NULL,
         "a directory is refused");
  cst_tool_result_free(&r);

  memset(&r, 0, sizeof(r));
  c = mk("deliver_file", "path", "../outside.txt", NULL);
  cst_tool_run(&ctx, &c, &r);
  EXPECT(r.exit_code == 1 && r.out &&
             strstr(r.out, "outside workspace") != NULL,
         "a path outside the workspace is refused");
  cst_tool_result_free(&r);

  memset(&r, 0, sizeof(r));
  c = mk("deliver_file", NULL);
  cst_tool_run(&ctx, &c, &r);
  EXPECT(r.exit_code == 1 && r.out && strstr(r.out, "path") != NULL,
         "a missing path is refused");
  cst_tool_result_free(&r);

  cst_agent_destroy(ag);
}

/* What actually goes out: the delivered file is not in the body, a document
 * attachment on the same transcript still is. */
static void test_wire(const char *session) {
  char md_src[1200], txt_src[1200];
  full_path(md_src, sizeof(md_src), "out/report.md");
  const char *note = "the build log lives in out/build.log\n";
  wfile("notes.txt", note, strlen(note));
  full_path(txt_src, sizeof(txt_src), "notes.txt");

  char err[256] = "";
  cst_attachment_t *art =
      cst_attach_import_artifact(session, md_src, NULL, err, sizeof(err));
  cst_attachment_t *doc =
      cst_attach_import_file(session, txt_src, err, sizeof(err));
  EXPECT(art && doc, "both a delivered file and a document");

  cst_message_t msgs[4];
  memset(msgs, 0, sizeof(msgs));
  msgs[0].role = CST_ROLE_USER;
  msgs[0].content = (char *)"what did you build?";
  msgs[1].role = CST_ROLE_ASSISTANT;
  msgs[1].tool_calls_json = (char *)"[{\"id\":\"c1\",\"type\":\"function\","
                                      "\"function\":{\"name\":\"deliver_file\","
                                      "\"arguments\":\"{}\"}}]";
  /* Two tool results of one turn: the first delivered a file for the person,
   * the second came back with a document the model is meant to read. The
   * delivered file must add nothing (not even a "not sent" note); the
   * document must still be inlined. */
  msgs[2].role = CST_ROLE_TOOL;
  msgs[2].content = (char *)"the finished report is stored for the user";
  snprintf(msgs[2].tool_call_id, sizeof(msgs[2].tool_call_id), "c1");
  snprintf(msgs[2].tool_name, sizeof(msgs[2].tool_name), "deliver_file");
  msgs[2].attach = art;
  msgs[2].nattach = art ? 1 : 0;
  msgs[3].role = CST_ROLE_TOOL;
  msgs[3].content = (char *)"read the build log";
  snprintf(msgs[3].tool_call_id, sizeof(msgs[3].tool_call_id), "c2");
  snprintf(msgs[3].tool_name, sizeof(msgs[3].tool_name), "read_file");
  msgs[3].attach = doc;
  msgs[3].nattach = doc ? 1 : 0;

  cst_llm_request_t req;
  memset(&req, 0, sizeof(req));
  req.model = "test-model";
  req.msgs = msgs;
  req.nmsgs = 4;
  req.session_path = session; /* a document is read back from the store */
  char *body = cst_llm_build_body(&req);
  EXPECT(body != NULL, "body built");
  if (body) {
    EXPECT(art && strstr(body, art->rel) == NULL,
           "the delivered file's stored name is not in the request");
    EXPECT(strstr(body, "for_user") == NULL,
           "and nothing marks it as an attachment either");
    EXPECT(strstr(body, "the build log lives in") != NULL,
           "a document attachment still travels as text");
    cJSON *j = cJSON_Parse(body);
    const cJSON *m = j ? cJSON_GetObjectItemCaseSensitive(j, "messages") : NULL;
    EXPECT(m && cJSON_GetArraySize(m) == 5,
           "four messages plus the one deferred document message");
    cJSON_Delete(j);
    free(body);
  }
  cst_attachments_free(art, art ? 1 : 0);
  cst_attachments_free(doc, doc ? 1 : 0);
}

int main(void) {
  if (test_tmpdir(g_ws, sizeof(g_ws), "cst-artifacts") != 0) {
    perror("tmpdir");
    return 1;
  }
  char mk1[1200];
  snprintf(mk1, sizeof(mk1), "%s/out", g_ws);
  test_mkdir(mk1, 0755);
  char session[1200];
  snprintf(session, sizeof(session), "%s/s.jsonl", g_ws);

  test_registry();
  test_import(session);
  test_any_bytes(session);
  test_json(session);
  test_session_file();
  test_tool(session);
  test_wire(session);

  test_rmrf(g_ws);
  if (nfail) {
    fprintf(stderr, "%d check(s) failed\n", nfail);
    return 1;
  }
  printf("test_artifacts: OK\n");
  return 0;
}