dreadnode 2.0.29__tar.gz → 2.0.30__tar.gz

{dreadnode-2.0.29 → dreadnode-2.0.30}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dreadnode
-Version: 2.0.29
+Version: 2.0.30
 Summary: Dreadnode SDK
 Project-URL: Homepage, https://dreadnode.io
 Project-URL: Documentation, https://docs.dreadnode.io

{dreadnode-2.0.29 → dreadnode-2.0.30}/dreadnode/airt/assessment.py

@@ -155,6 +155,11 @@ class Assessment:
         self._tracing_enabled: bool = False
         self._auto_registered: bool = False
         self._context_token: contextvars.Token[Assessment | None] | None = None
+        # Terminal-failure latch. Once fail() runs, auto-finalization
+        # (_finalize/complete/atexit) must NOT flip the status back to
+        # "completed" — otherwise a failure caught inside the `async with`
+        # block is masked by the clean-exit auto-complete (ENG-6822).
+        self._failed: bool = False
 
     async def __aenter__(self) -> Assessment:
         self._context_token = _current_assessment.set(self)
@@ -507,7 +512,7 @@ class Assessment:
             return
 
         api, profile = ctx
-        status = "completed" if self._attack_results else "failed"
+        status = "completed" if (self._attack_results and not self._failed) else "failed"
         with contextlib.suppress(Exception):
             api.update_airt_assessment(
                 profile.org_key,
@@ -542,6 +547,11 @@ class Assessment:
         Returns:
             True if successfully marked, False otherwise.
         """
+        # A failed assessment is terminal — never auto-complete over it.
+        if self._failed:
+            logger.debug("Skipping complete(): assessment already marked as failed")
+            return False
+
         ctx = _get_platform_context_with_fallback()
         if ctx is None or self._assessment_id is None:
             return False
@@ -570,6 +580,10 @@ class Assessment:
         Returns:
             True if successfully marked, False otherwise.
         """
+        # Latch terminal failure first — even offline / before registration —
+        # so any later auto-finalization won't re-complete (ENG-6822).
+        self._failed = True
+
         ctx = _get_platform_context_with_fallback()
         if ctx is None or self._assessment_id is None:
             return False

{dreadnode-2.0.29 → dreadnode-2.0.30}/dreadnode/app/cli/task.py

@@ -881,6 +881,20 @@ def validate(
     smoke: t.Annotated[bool, cyclopts.Parameter(negative=())] = False,
     pull: t.Annotated[bool, cyclopts.Parameter(name="--pull", negative=())] = False,
     yes: t.Annotated[bool, cyclopts.Parameter(name="--yes", alias="-y", negative=())] = False,
+    env: t.Annotated[
+        list[str] | None,
+        cyclopts.Parameter(
+            name="--env",
+            alias="-e",
+            help=(
+                "Inject env into the challenge service at smoke time only "
+                "(KEY=VALUE for a literal, or bare KEY to forward the value from "
+                "your shell so secrets stay out of argv; repeatable). Scoped to "
+                "the challenge/task service -- does not reach solution.sh or verify."
+            ),
+            negative_iterable=(),
+        ),
+    ] = None,
     timeout: int | None = None,
     platform: PlatformArgs = PlatformArgs(),
 ) -> None:
@@ -905,6 +919,14 @@ def validate(
             ``outcome_judge`` has no mechanical half and is skipped.
         pull: Treat path as a published task ref and pull it for local validation.
         yes: Accept remote validation without prompting when path is not local.
+        env: Env to inject into the challenge service for the smoke run only
+            (``-e``/``--env``, repeatable). ``KEY=VALUE`` sets a literal; a bare
+            ``KEY`` forwards the value from ``os.environ`` so secrets never touch
+            argv. Merged into the ephemeral compose -- the on-disk compose is
+            untouched -- and overrides any same-named key the compose declares.
+            Targets the challenge/task service(s) only: it does NOT reach
+            solution.sh or verify, which run in a separate sandbox. Only
+            meaningful with ``--smoke``; ignored (with a warning) otherwise.
         timeout: Per-task wall-clock budget in seconds for smoke testing.
             When unset, falls back to the task's ``max_agent_timeout_sec`` or
             120 seconds if neither is declared.
@@ -912,6 +934,10 @@ def validate(
     if smoke:
         build = True
 
+    if env and not smoke:
+        print_warning("--env / -e only applies to --smoke; ignoring", indent=2)
+        env = None
+
     resolved = Path(path).resolve()
     if resolved.exists():
         _validate_local_task_path(
@@ -920,6 +946,7 @@ def validate(
             build=build,
             smoke=smoke,
             timeout=timeout,
+            service_env=env,
         )
         return
 
@@ -939,6 +966,7 @@ def validate(
         build=build,
         smoke=smoke,
         timeout=timeout,
+        service_env=env,
     )
 
 
@@ -949,6 +977,7 @@ def _validate_local_task_path(
     build: bool,
     smoke: bool,
     timeout: int | None,
+    service_env: list[str] | None = None,
 ) -> None:
     """Run validation against a local task directory or task tree."""
     from dreadnode.packaging.task_validation import (
@@ -997,7 +1026,7 @@ def _validate_local_task_path(
                     errs.append(ValidationIssue("error", "docker-build", summary))
 
             if smoke and not errs:
-                smoke_issues = _run_smoke_test(task_dir, timeout=timeout)
+                smoke_issues = _run_smoke_test(task_dir, timeout=timeout, service_env=service_env)
                 for issue in smoke_issues:
                     if issue.level == "error":
                         errs.append(issue)
@@ -1087,6 +1116,7 @@ def _validate_remote_task_ref(
     build: bool,
     smoke: bool,
     timeout: int | None = None,
+    service_env: list[str] | None = None,
 ) -> None:
     dn = configured_dreadnode(platform)
     print_info(f"Pulling {ref.qualified_name}@{ref.version} for local validation")
@@ -1106,6 +1136,7 @@ def _validate_remote_task_ref(
             build=build,
             smoke=smoke,
             timeout=timeout,
+            service_env=service_env,
         )
 
 
@@ -1815,13 +1846,75 @@ def _wipe_ephemeral_bind_mount_dirs(paths: list[Path]) -> None:
             path.mkdir(parents=True, exist_ok=True)
 
 
-def _write_ephemeral_compose_file(original: Path, destination: Path) -> dict[str, list[int]]:
+def _parse_service_env(values: list[str] | None) -> dict[str, str]:
+    """Parse repeatable smoke ``-e``/``--env`` flags into a ``{KEY: VALUE}`` map.
+
+    ``KEY=VALUE`` sets a literal value (split on the first ``=``). A bare ``KEY``
+    forwards the value from ``os.environ`` so secrets stay out of argv, shell
+    history, and ``ps``. A bare key whose variable is unset is an error -- you
+    asked to forward it, but there is nothing to forward, so fail loud rather
+    than silently inject an empty string.
+    """
+    if not values:
+        return {}
+    import os
+
+    out: dict[str, str] = {}
+    for raw in values:
+        key, sep, value = raw.partition("=")
+        key = key.strip()
+        if not key:
+            raise ValueError(f"--env expects KEY or KEY=VALUE, got: {raw!r}")
+        if sep:
+            out[key] = value
+        elif key in os.environ:
+            out[key] = os.environ[key]
+        else:
+            raise ValueError(
+                f"--env {key}: no value supplied and {key} is not set in the environment"
+            )
+    return out
+
+
+def _normalize_compose_env(raw: t.Any) -> dict[str, t.Any]:
+    """Normalize a compose ``environment`` block to map form.
+
+    Compose accepts either a list (``["K=v", "BARE"]``) or a map
+    (``{K: v}``). List entries are split on the first ``=``; a bare list entry
+    (host pass-through) becomes ``{KEY: None}``. Anything else yields an empty
+    map so callers can merge unconditionally.
+    """
+    if isinstance(raw, dict):
+        return dict(raw)
+    if isinstance(raw, list):
+        out: dict[str, t.Any] = {}
+        for entry in raw:
+            if not isinstance(entry, str):
+                continue
+            key, sep, value = entry.partition("=")
+            out[key] = value if sep else None
+        return out
+    return {}
+
+
+def _write_ephemeral_compose_file(
+    original: Path,
+    destination: Path,
+    *,
+    inject_env: dict[str, str] | None = None,
+) -> dict[str, list[int]]:
     """Rewrite ``original`` so every service's ports use random host ports.
 
     Docker compose short-form ``"3000"`` (container port only) assigns a
     random host port at ``up`` time, which we discover afterwards via
     ``docker compose port``. Returns ``{service: [container_port, ...]}`` so
     the caller knows which ports to query.
+
+    When ``inject_env`` is given, those entries are merged into every service's
+    ``environment`` block (normalized to map form first), overriding any
+    same-named key the compose declares. This is the smoke-time challenge-env
+    injection (``dn task validate --smoke -e KEY=VALUE``): non-destructive, the
+    on-disk compose is never touched -- the same seam where ports are mutated.
     """
     import yaml
 
@@ -1838,6 +1931,10 @@ def _write_ephemeral_compose_file(original: Path, destination: Path) -> dict[str
     for svc_name, svc_config in services.items():
         if not isinstance(svc_config, dict):
             continue
+        if inject_env:
+            merged = _normalize_compose_env(svc_config.get("environment"))
+            merged.update(inject_env)
+            svc_config["environment"] = merged
         raw_ports = svc_config.get("ports")
         if not isinstance(raw_ports, list) or not raw_ports:
             continue
@@ -2112,7 +2209,12 @@ def _resolve_smoke_timeout(raw: dict[str, t.Any], explicit: int | None) -> int:
     return _DEFAULT_SMOKE_TIMEOUT
 
 
-def _run_smoke_test(task_dir: Path, *, timeout: int | None = None) -> list[ValidationIssue]:
+def _run_smoke_test(
+    task_dir: Path,
+    *,
+    timeout: int | None = None,
+    service_env: list[str] | None = None,
+) -> list[ValidationIssue]:
     """Run the v1 smoke lifecycle for a single task directory.
 
     Lifecycle:
@@ -2160,6 +2262,14 @@ def _run_smoke_test(task_dir: Path, *, timeout: int | None = None) -> list[Valid
 
     issues: list[ValidationIssue] = []
 
+    # Challenge-service env to merge into the ephemeral compose (smoke only).
+    # Parse up front so a bad/unset bare key fails loud before any docker work.
+    try:
+        injected_service_env = _parse_service_env(service_env)
+    except ValueError as exc:
+        issues.append(ValidationIssue("error", "smoke", str(exc)))
+        return issues
+
     raw = yaml.safe_load((task_dir / "task.yaml").read_text())
     timeout = _resolve_smoke_timeout(raw, timeout)
     verification = raw.get("verification") or {}
@@ -2250,6 +2360,15 @@ def _run_smoke_test(task_dir: Path, *, timeout: int | None = None) -> list[Valid
     compose_path = _find_compose_file(task_dir)
     has_compose = compose_path is not None
 
+    if injected_service_env and not has_compose:
+        issues.append(
+            ValidationIssue(
+                "warning",
+                "smoke",
+                "-e/--env was given but this task has no compose service to inject into",
+            )
+        )
+
     # Two smoke contexts are built after compose is up:
     #   - host_env / host_template_vars: service URLs using the randomly
     #     assigned host ports. Used for scripts that run on the host.
@@ -2338,7 +2457,11 @@ def _run_smoke_test(task_dir: Path, *, timeout: int | None = None) -> list[Valid
                 prefix=f"{task_dir.name}-smoke-compose-"
             )
             smoke_compose_file = Path(smoke_compose_workspace.name) / compose_path.name
-            declared_host_ports = _write_ephemeral_compose_file(compose_path, smoke_compose_file)
+            declared_host_ports = _write_ephemeral_compose_file(
+                compose_path,
+                smoke_compose_file,
+                inject_env=injected_service_env or None,
+            )
 
             # Clear smoke-owned ephemeral bind-mount dirs (e.g. ./.submission)
             # before compose up so a prior run's leftover state can't poison

{dreadnode-2.0.29 → dreadnode-2.0.30}/dreadnode/app/server/model_resolution.py

@@ -108,18 +108,13 @@ def validate_model_environment(config: TurnModelConfig) -> str | None:
 
 def build_turn_generator(config: TurnModelConfig) -> "str | Generator":
     """Build the effective generator input for a turn or compaction request."""
-    effective_model: str | Generator = config.generator_model
     if not config.is_platform_proxy:
-        return effective_model
+        return config.generator_model
 
-    from dreadnode.generators.generator import GenerateParams, get_generator
+    from dreadnode.generators.proxy import build_proxy_generator
 
-    generator = get_generator(
+    return build_proxy_generator(
         config.generator_model,
-        params=GenerateParams(
-            api_base=config.api_base,
-            extra={"custom_llm_provider": "litellm_proxy"},
-        ),
+        api_base=config.api_base,
+        api_key=config.api_key,
     )
-    generator.api_key = config.api_key
-    return generator

{dreadnode-2.0.29 → dreadnode-2.0.30}/dreadnode/app/tui/tool_format.py

@@ -208,7 +208,7 @@ def _get_result_text(res: t.Any) -> str:
     if isinstance(res, list):
         # Multimodal content-part lists (e.g. a ``read`` image result is a text
         # caption + an image part). Pull text out of each part and skip image/
-        # audio parts, whose repr would otherwise bloat or corrupt the summary.
+        # video / audio parts, whose repr would otherwise bloat or corrupt the summary.
         parts: list[str] = []
         for x in res:
             if isinstance(x, dict):
@@ -218,7 +218,7 @@ def _get_result_text(res: t.Any) -> str:
             if isinstance(text, str):
                 parts.append(text)
                 continue
-            if getattr(x, "type", None) in ("image_url", "input_audio"):
+            if getattr(x, "type", None) in ("image_url", "file", "input_audio"):
                 continue
             parts.append(str(x))
         return "\n".join(p for p in parts if p)
@@ -237,9 +237,9 @@ def _summarize_read(_name: str, _args: dict[str, t.Any], result: t.Any) -> str |
         return None
 
     head = text.splitlines()[0] if text else ""
-    # Image reads return a "Read image · <fmt> · <size>" caption (the path is
-    # already shown in the call label, so it is intentionally not repeated).
-    if head.startswith("Read image"):
+    # Image/video reads return a "Read <media> · <fmt> · <size>" caption (the
+    # path is already shown in the call label, so it is intentionally not repeated).
+    if head.startswith(("Read image", "Read video")):
         return head
     if head.startswith("PDF: "):
         return f"PDF: {head.removeprefix('PDF: ').rsplit('/', 1)[-1]}."

{dreadnode-2.0.29 → dreadnode-2.0.30}/dreadnode/builtin_capabilities/dreadnode/agents/dreadnode.md

@@ -60,6 +60,8 @@ You are running in an isolated runtime. You can freely modify files and run comm
 
 Use bash for system and CLI commands. Use python for SDK inspection, quick scripting, and data analysis. If available, use memory to track key platform context and decisions across tool calls; use todo to track multi-step work; use think for complex reasoning without acting; use report to persist structured deliverables; use session_search to recover earlier messages from the current session. Tool availability depends on the active capability and runtime.
 
+When the user provides media files or URLs to media files (images, videos, audio), use the `read` tool to load them. `read` returns proper multimodal content parts that the model can see or hear. Do not use `web_extract` or `fetch` for media files — those return raw bytes or text, not model-viewable multimodal content. For local media paths, use `read` as well.
+
 Capabilities add domain-specific agents, tools, and prompts. When a capability is active, let that capability own its specialized domain behavior instead of re-creating it in the core runtime agent.
 
 Persist results: traces are sent automatically, use dreadnode dataset publish to save data, dreadnode model push for models, and /export for conversation transcripts.

{dreadnode-2.0.29 → dreadnode-2.0.30}/dreadnode/builtin_capabilities/dreadnode/skills/dreadnode-concepts/references/platform/settings.md

@@ -18,6 +18,7 @@ Settings is the app’s entry point for organization and user configuration. It
 | Workspaces  | workspace creation and sharing           | where should work happen and who gets access?                | [Workspaces](./workspaces.md)       |
 | Secrets     | personal provider credentials            | which keys do I want available for my runs and evaluations?  | [Secrets](./secrets.md)             |
 | Chat Models | chat UI model availability               | which inference models should appear in my assistant picker? | [Chat models](./chat-models.md)     |
+| Preferences | user notification and marketing controls | which account-level communication preferences should I keep? | this page                                 |
 | Billing     | SaaS credits and payment controls        | how do we pay for usage and keep workloads running?          | [Credits](./credits.md)             |
 
 ## What lives in settings
@@ -29,20 +30,22 @@ Settings is the app’s entry point for organization and user configuration. It
 | Workspaces  | workspace creation, sharing, and per-workspace access management                      |
 | Secrets     | provider API keys and custom environment variables                                    |
 | Chat Models | your own BYOK model additions on top of the always-available Dreadnode models         |
+| Preferences | desktop notification toggles and marketing email subscription (`Dreadnode Updates`)   |
 | Billing     | credits, auto-refill, transactions, and usage in SaaS mode                            |
 
 The settings shell also surfaces an invite banner when an organization appears to be solo and the current user can manage members. In the app, that banner uses the `Invite Team` action to send you directly into membership management.
 
 ## Common operator tasks
 
-| If you need to…                                             | Go to         | Why                                                               |
-| ----------------------------------------------------------- | ------------- | ----------------------------------------------------------------- |
-| rename the org or review org-level limits                   | `General`     | this is the top-level org metadata surface                        |
-| invite coworkers and adjust roles                           | `Members`     | org membership and permission changes happen here                 |
-| create a shared delivery area for a team or engagement      | `Workspaces`  | workspace creation and access live here                           |
-| add your own provider key for future runs                   | `Secrets`     | secrets are user-owned even though they are managed from settings |
-| decide which chat models appear in your chat UI             | `Chat Models` | this is a user preference surface, not the artifact registry      |
-| configure payment methods, auto-refill, or usage guardrails | `Billing`     | this is the SaaS billing and credits surface                      |
+| If you need to…                                              | Go to         | Why                                                               |
+| ------------------------------------------------------------ | ------------- | ----------------------------------------------------------------- |
+| rename the org or review org-level limits                    | `General`     | this is the top-level org metadata surface                        |
+| invite coworkers and adjust roles                            | `Members`     | org membership and permission changes happen here                 |
+| create a shared delivery area for a team or engagement       | `Workspaces`  | workspace creation and access live here                           |
+| add your own provider key for future runs                    | `Secrets`     | secrets are user-owned even though they are managed from settings |
+| decide which chat models appear in your chat UI              | `Chat Models` | this is a user preference surface, not the artifact registry      |
+| subscribe or unsubscribe from product-update marketing email | `Preferences` | this controls `Dreadnode Updates` only, not transactional email   |
+| configure payment methods, auto-refill, or usage guardrails  | `Billing`     | this is the SaaS billing and credits surface                      |
 
 ## Platform admin (runtime LiteLLM controls)
 
@@ -50,6 +53,8 @@ The settings shell also surfaces an invite banner when an organization appears t
 
 * `Admin → Credentials` rotates named LiteLLM provider credentials at runtime.
 * `Admin → Model Deployments` manages deployment rows and credential assignment for load balancing / routing changes. Drag rows to set the order hosted models appear in TUI and chat model lists — changes apply to new sessions within about a minute.
+* `Admin → Diagnostics` shows read-only LiteLLM liveliness and running-version status.
+* `Admin → Security` manages signup abuse controls (rolling signup-per-IP limits and CIDR deny rules) plus disposable email-domain override entries.
 
 These controls are `PLATFORM_ADMIN`-scoped and are separate from user-owned `Secrets` in settings.
 
@@ -116,6 +121,14 @@ Use `Secrets` when you are storing credentials that you personally want to injec
 
 Use `Chat Models` when you are adding your own BYOK models to the interactive assistant UI — Dreadnode-hosted models are available by default. See [Chat models](./chat-models.md) for the full mechanic, including which provider key each BYOK model needs.
 
+### Preferences
+
+Use `Preferences` when you are managing account-level communication controls.
+
+* toggle desktop notifications for long-running task completion alerts
+* subscribe or unsubscribe from the `Dreadnode Updates` marketing list
+* note that transactional emails (verification, password reset, operational notices) are independent from marketing preference
+
 ### Billing
 
 Use `Billing` when you are managing credits-backed usage in SaaS deployments.
@@ -134,14 +147,15 @@ Use `Billing` when you are managing credits-backed usage in SaaS deployments.
 
 ## Permission guide
 
-| Section     | Scope                              | Safe default assumption                                           |
-| ----------- | ---------------------------------- | ----------------------------------------------------------------- |
-| General     | organization                       | org-admin action                                                  |
-| Members     | organization                       | org-admin action with invite and role management                  |
-| Workspaces  | organization plus workspace access | org-level creation plus workspace-sharing controls                |
-| Secrets     | user                               | each user manages their own credentials                           |
-| Chat Models | user                               | treat as a per-user model-picker preference, not a registry write |
-| Billing     | organization, SaaS only            | owner-level billing action                                        |
+| Section     | Scope                              | Safe default assumption                                            |
+| ----------- | ---------------------------------- | ------------------------------------------------------------------ |
+| General     | organization                       | org-admin action                                                   |
+| Members     | organization                       | org-admin action with invite and role management                   |
+| Workspaces  | organization plus workspace access | org-level creation plus workspace-sharing controls                 |
+| Secrets     | user                               | each user manages their own credentials                            |
+| Chat Models | user                               | treat as a per-user model-picker preference, not a registry write  |
+| Preferences | user                               | per-user communication preferences for notifications and marketing |
+| Billing     | organization, SaaS only            | owner-level billing action                                         |
 
 ## SaaS versus Enterprise
 

{dreadnode-2.0.29 → dreadnode-2.0.30}/dreadnode/generators/generator/litellm_.py

@@ -8,6 +8,7 @@ import warnings
 from loguru import logger
 
 from dreadnode.agents.tools import FunctionDefinition, ToolDefinition
+from dreadnode.app.model_catalog import infer_provider
 from dreadnode.generators.exceptions import GeneratorWarning, ProcessingError
 from dreadnode.generators.generator.base import (
     Fixup,
@@ -21,6 +22,7 @@ from dreadnode.generators.generator.base import (
     with_fixups,
 )
 from dreadnode.generators.message import (
+    CompatibilityFlag,
     ContentAudioInput,
     ContentImageUrl,
     ContentText,
@@ -341,6 +343,16 @@ g_fixups = [
 vertex_image_pattern = re.compile(r"(data:[\w/]+?;base64,[A-Za-z0-9+/=]+)")
 
 
+def _compatibility_flags_for_model(model: str) -> set[CompatibilityFlag]:
+    """Return message serialization flags needed by LiteLLM provider adapters."""
+    normalized = model.lower()
+    if infer_provider(model) == "google" or normalized.startswith(
+        ("vertex_ai/", "vertex_ai_beta/")
+    ):
+        return {"file_data_as_data_url"}
+    return set()
+
+
 class LiteLLMGenerator(Generator):
     """
     Generator backed by the LiteLLM library.
@@ -762,9 +774,13 @@ class LiteLLMGenerator(Generator):
                 {k: v for k, v in merged.items() if k not in ("tools",)},
             )
 
+            compatibility_flags = _compatibility_flags_for_model(self.model)
             response = await acompletion(
                 model=self.model,
-                messages=[message.to_openai() for message in messages],
+                messages=[
+                    message.to_openai(compatibility_flags=compatibility_flags)
+                    for message in messages
+                ],
                 api_key=self.api_key,
                 **merged,
             )