Give Architect its project context (#2557)

* feat(backend): tell architect which project it is in

The Architect asked "which project?" on every session because it was never
told. session.project_id reaches the DB tenant GUC, the X-Project-Id header on
every tool call, and the Celery headers — but never the prompt.

It cannot resolve this itself either: the project table is not covered by the
project_isolation RLS policy, so list_projects returns every project in the
organization with nothing marking the active one.

Resolve the project in build_agent, pass it to ArchitectAgent, and render a
block in the iteration prompt. A project that no longer resolves degrades to
no block rather than failing the turn.

The block is explicit that reads cover this project plus organization-level
entities, and that other projects' contents are not readable — RLS matches
project_id = current OR project_id IS NULL, so claiming otherwise would have
the agent give wrong reasons for what it sees.

* fix(backend): stop telling architect to resolve a project

create_endpoint's tool doc said project_id was "REQUIRED — the endpoint cannot
be created without it", and its parameter doc said to resolve it via
list_projects and "ask if ambiguous". Both are wrong: EndpointCreate.project_id
is Optional and auto_stamp fills it from the request scope. The agent followed
the docs, so it asked.

Correct create_endpoint and get_project in mcp_tools.yaml and mirror the fix
into the public tool-catalog.md. Rework the phase guidance that pushed the same
way: discovery no longer resolves project_id before create_endpoint, planning no
longer plans a project, creation only creates one when the user asked for a new
project, and save_plan's project field is documented as new-project-only.

* test(backend): cover architect project context

Pin the two facts that make the lookup necessary — the project table has no
project_id column, so list_projects cannot self-filter — and the degradation
paths, so a deleted project or a malformed id omits the block instead of
failing the turn.

Also assert no tool doc tells the agent to ask for a project, and that the
prompt block stays honest about organization-level rows and unreadable
projects.

* fix(sdk): flatten and clip project fields before the prompt

Project names and descriptions are user-controlled and the context block is
line-structured, so a newline in either forged a field: a description of
"harmless\nProject: Evil" rendered as a second Project: line and the agent read
a project the user never set.

Collapse whitespace and clip to config caps. The description is rendered every
turn, so an unbounded one was paid for on each.

Also name the exception in the resolver's warning, matching the sibling
degrade-gracefully logs in attachments.py and chat.py.

Author: nicolai-rhesis

apps/backend/src/rhesis/backend/app/mcp_server/mcp_tools.yaml

@@ -15,10 +15,10 @@ tools:
     path: /projects/{project_id}
     description: >
       Get one project by ID. ENTITY: Project. PREREQUISITES: project_id
-      from list_projects. CHAIN: use before create_endpoint (project_id
-      is required) or when confirming project context for an endpoint.
-      NEVER: ask the user for a project ID when they gave a name — use
-      list_projects with $filter first.
+      from list_projects. CHAIN: use when the user asks about a specific
+      project by name. NOT needed before create_endpoint — that takes its
+      project from the request scope. NEVER: ask the user for a project ID
+      when they gave a name — use list_projects with $filter first.
     parameters:
       project_id:
         description: "REQUIRED. UUID of the project."
@@ -999,18 +999,18 @@ tools:
     requires_confirmation: true
     description: >
       Register a new REST API endpoint (an AI system to be tested — a chat
-      API, completion API, or chatbot). Resolve the target project FIRST with
-      list_projects and pass its id as project_id (REQUIRED — the endpoint
-      cannot be created without it). Do NOT send server-managed fields
+      API, completion API, or chatbot). The endpoint is created in the
+      current project automatically — do NOT resolve a project or ask the
+      user which one to use. Do NOT send server-managed fields
       (id, user_id, organization_id, status_id, created_at, updated_at).
       After creating, you can verify reachability with check_endpoint.
     parameters:
       name:
         description: "REQUIRED. Endpoint name, unique within the project (string)."
       project_id:
         description: >
-          REQUIRED. UUID of the project this endpoint belongs to. Resolve it
-          via list_projects (use the user's project, or ask if ambiguous).
+          Omit. The project scope is taken from the request, not from this
+          field.
       connection_type:
         description: 'Connection protocol. Use "REST" for HTTP APIs (default for chatbots).'
       url:

apps/backend/src/rhesis/backend/app/services/architect/runner.py

@@ -117,6 +117,39 @@ async def prepare_and_load_session(
     return resolved_session_id, session_data
 
 
+def _resolve_project_context(db, project_id: Optional[str], organization_id: str) -> Optional[dict]:
+    """Look up the session's project so the agent can name it.
+
+    The agent cannot resolve this itself: ``project`` is not covered by the
+    ``project_isolation`` RLS policy, so ``list_projects`` returns every
+    project in the organization with nothing marking the active one.
+
+    A project that no longer resolves is not fatal — the prompt block is
+    simply omitted and the agent behaves as it did before.
+    """
+    if not project_id:
+        return None
+
+    from uuid import UUID
+
+    from rhesis.backend.app.crud import project as project_crud
+
+    try:
+        project = project_crud.get_project(db, UUID(project_id), organization_id=organization_id)
+    except Exception as exc:
+        logger.warning("Could not resolve project %s for Architect context: %s", project_id, exc)
+        return None
+
+    if not project:
+        return None
+
+    return {
+        "project_id": str(project.id),
+        "name": project.name or "",
+        "description": project.description or "",
+    }
+
+
 @observe()
 async def build_agent(
     session_data: Dict[str, Any],
@@ -144,6 +177,7 @@ async def build_agent(
             raise ValueError(f"User {user_id} is inactive")
         delegation_token = create_service_delegation_token(user, "backend")
         model = get_user_generation_model(db, user)
+        project_context = _resolve_project_context(db, project_id, organization_id)
 
     agent_state = session_data["agent_state"]
     snapshot = ArchitectAgentStateSnapshot(
@@ -174,6 +208,7 @@ async def build_agent(
         event_handlers=[ws_handler, *tracing_handlers],
         max_iterations=snapshot.max_iterations,
         verbose=False,
+        project_context=project_context,
     )
     agent.restore_state(snapshot)
 

sdk/src/rhesis/sdk/agents/architect/agent.py

@@ -139,6 +139,7 @@ def __init__(
         history_window: Optional[int] = None,
         verbose: bool = False,
         event_handlers: Optional[List[AgentEventHandler]] = None,
+        project_context: Optional[Dict[str, Any]] = None,
     ):
         self._cfg = config or ArchitectConfig()
         templates_dir = Path(__file__).parent / "prompt_templates"
@@ -155,6 +156,10 @@ def __init__(
             jinja_env=build_architect_jinja_env(templates_dir),
         )
         self._conversation_history: List[Dict[str, Any]] = []
+        # The project this session is scoped to. Tool calls carry it as
+        # X-Project-Id, so the agent must not ask the user which project
+        # they mean — see _format_project_context.
+        self._project_context: Dict[str, Any] = project_context or {}
         self._plan: Optional[ArchitectPlan] = None
         self._mode: AgentMode = AgentMode.DISCOVERY
         self._workflow_path: WorkflowPath = WorkflowPath.UNSET
@@ -1516,8 +1521,52 @@ def _build_prompt(
             plan_progress_text=plan_progress_text,
             discovery_state_text=discovery_text,
             attachments_text=attachments_text,
+            project_context_text=self._format_project_context(),
         )
 
+    def _format_project_context(self) -> str:
+        """Describe the project this session is scoped to.
+
+        Without this the agent has no way to know: the ``project`` table is
+        not covered by the ``project_isolation`` RLS policy, so
+        ``list_projects`` returns every project in the organization with
+        nothing marking the active one. It would ask the user instead.
+        """
+        name = self._clean_project_field(
+            self._project_context.get("name"), self._cfg.project_name_max_chars
+        )
+        project_id = (self._project_context.get("project_id") or "").strip()
+        if not name and not project_id:
+            return ""
+
+        label = name or "(unnamed)"
+        lines = [f"Project: {label}"]
+        if project_id:
+            lines.append(f"Project ID: {project_id}")
+        description = self._clean_project_field(
+            self._project_context.get("description"),
+            self._cfg.project_description_max_chars,
+        )
+        if description:
+            lines.append(f"About: {description}")
+        return "\n".join(lines)
+
+    @staticmethod
+    def _clean_project_field(value: Any, max_chars: int) -> str:
+        """Flatten and clip a project field before it enters the prompt.
+
+        The block is line-structured, so a newline inside a name or
+        description would forge a field the user never set (``About: x\nProject:
+        Other``). Collapsing whitespace removes that, and the clip keeps a long
+        description from being paid for on every turn.
+        """
+        if not isinstance(value, str):
+            return ""
+        flattened = " ".join(value.split())
+        if len(flattened) > max_chars:
+            return flattened[:max_chars].rstrip() + "…"
+        return flattened
+
     def _format_discovery_state(self) -> str:
         """Format the discovery state for the iteration prompt."""
         ds = self._discovery_state

sdk/src/rhesis/sdk/agents/architect/config.py

@@ -40,6 +40,12 @@ class ArchitectConfig:
     recent_msg_max_chars: int = 2_000
     older_msg_max_chars: int = 500
 
+    # ── project context ───────────────────────────────────────────
+    # Rendered every turn, so an unbounded description would be paid for
+    # on each one.
+    project_name_max_chars: int = 200
+    project_description_max_chars: int = 500
+
     # ── streaming / tool-result preview ───────────────────────────
     tool_result_preview_chars: int = 4_000
     reasoning_preview_chars: int = 200

sdk/src/rhesis/sdk/agents/architect/prompt_templates/iteration_prompt.j2

@@ -3,6 +3,23 @@ Current Mode: {{ mode }}
 
 User Message: {{ user_query }}
 
+{% if project_context_text %}
+Current Project (already resolved — do not ask):
+{{ project_context_text }}
+
+Every tool call you make is already scoped to this project. Do not ask the user
+which project they mean, and do not call `list_projects` to pick one. Entities
+you create land here automatically — omit `project_id` and let the request scope
+fill it in.
+
+Reads return this project's entities plus organization-level ones (those with no
+project). Other projects' contents are not visible: you can name them via
+`list_projects` and talk about them, but you cannot read or write their
+requirements, metrics, test sets or runs from this session. If the user asks for
+another project's data, say that plainly rather than reporting an empty result as
+though the project were empty.
+{% endif %}
+
 {% if phase_knowledge_text %}
 Phase Guidance:
 {{ phase_knowledge_text }}

sdk/src/rhesis/sdk/agents/architect/prompt_templates/telemachus-save-plan.j2

@@ -4,7 +4,7 @@ When you have formulated the plan, you MUST call **save_plan** before presenting
 **save_plan is strictly validated** — the call fails if any field is missing, has the wrong type, or uses a value outside the allowed set. Send arguments exactly as specified below.
 
 Top-level structure (all three array fields are required, even if empty):
-- `project`: `{name, description}` — optional. Omit entirely when no project is needed (ad-hoc tests, single test set for an existing endpoint). Do NOT pass `null` or an empty object.
+- `project`: `{name, description}` — only for creating a **new** project, which is rare. The session is already scoped to a project and entities land there automatically, so never use this field to name the project you are already working in. Omit it entirely in the normal case. Do NOT pass `null` or an empty object.
 - `requirements`: array (required, may be empty)
 - `test_sets`: array (required, may be empty)
 - `metrics`: array (required, may be empty)
@@ -53,7 +53,7 @@ Common validation pitfalls that will be rejected:
 - `test_type: "single-turn"` or `"singleturn"` → must be `"Single-Turn"` or `"Multi-Turn"` (exact case, hyphen)
 - `num_tests: "15"` → must be the integer `15`
 - Top-level keys missing because the section was empty → send `[]`, not omit
-- `project: null` or `project: {}` when there is no project → omit the key entirely
+- `project: null`, `project: {}`, or `project` naming the current project → omit the key entirely
 - Extra fields not listed above are ignored, but typos in field names mean the data is dropped
 
 When save_plan fails the error response lists each invalid field with its reason. Read it, fix the offending fields, and re-call save_plan with the corrected arguments. Do NOT proceed to other tools while the plan is unsaved.

skills/rhesis/references/phases/creation.md

@@ -5,7 +5,7 @@ Execute the approved plan exactly — no extra entities.
 ## Order (mandatory)
 
 1. Reuse lookup — resolve IDs via `list_*` + `$filter` if needed
-2. `create_project` — only if planned
+2. `create_project` — only when the user asked for a **new** project. The session already has one; never create a project to "hold" this work
 3. `create_requirement` — **(new)** only; name + description
 4. Resolve all requirement IDs
 5. Metrics — **(reuse)** skip; **(improve)** `improve_metric`; **(new)** `create_metric` with plan name, **`metric_scope`**, `score_type`, `evaluation_prompt`, plus `description`, `evaluation_steps`, `reasoning`, `explanation` written to the depth in `metric-authoring.md`. Do NOT use `generate_metric`

skills/rhesis/references/phases/discovery.md

@@ -2,7 +2,7 @@
 
 When the user names an endpoint or wants to test an AI application:
 
-1. Resolve endpoint: `list_endpoints` with `$select=name,id,url,description`. If missing, `create_endpoint` (resolve `project_id` via `list_projects` first).
+1. Resolve endpoint: `list_endpoints` with `$select=name,id,url,description`. If missing, `create_endpoint` — it lands in the current project automatically, so omit `project_id`.
 2. `check_endpoint` — report failures before proceeding.
 
 If the project has **no** endpoint at all, lead your reply with that before any test design: nothing can run against the project until one is connected, so there are no responses to score. Show both routes — register the application's HTTP API, or wrap a Python function with the SDK's `@endpoint` decorator ([connecting an application](https://docs.rhesis.ai/docs/getting-started/connecting-application)) — and offer to create it now. Say it once; if the user would rather design tests first, continue.

skills/rhesis/references/phases/planning.md

@@ -48,7 +48,7 @@ State the type for each test set when you present the plan, and check that the m
 
 Propose existing entities when they match. Say explicitly: "I'll reuse 'Refuses Harmful Requests'."
 
-Skip project for ad-hoc work.
+Never plan a project. The session is already scoped to one, and entities land there automatically — plan a project only when the user explicitly asks for a new one.
 
 ## Approval gate
 

skills/rhesis/references/tool-catalog.md

@@ -21,11 +21,11 @@ List configured endpoints (the AI systems under test).
 ### `create_endpoint`
 Register a new REST API endpoint (the AI system to be tested — a chat API, completion API, or chatbot).
 
-Resolve the target project **first** with `list_projects` and pass its id — the endpoint cannot be created without a `project_id`. After creating, verify reachability with `check_endpoint`.
+The endpoint is created in the current project automatically — do not resolve a project or ask which one to use. After creating, verify reachability with `check_endpoint`.
 
 **Key parameters:**
 - `name` (required) — endpoint name, unique within the project
-- `project_id` (required) — UUID of the owning project; resolve via `list_projects`
+- `project_id` — omit; the project scope is taken from the request
 - `connection_type` — `"REST"` for HTTP APIs (default for chatbots)
 - `url` (required for REST) — full endpoint URL, e.g. `https://api.example.com/chat`
 - `method` — HTTP method, e.g. `"POST"` or `"GET"`
@@ -42,7 +42,7 @@ Resolve the target project **first** with `list_projects` and pass its id — th
 - `auth_token` — bearer token / API key (write-only, stored encrypted)
 - `client_id`, `client_secret`, `token_url`, `scopes`, `audience` — OAuth client-credentials fields (`client_secret` is write-only, stored encrypted)
 
-**Common mistakes:** Omitting `project_id` — the create fails because it is a required relationship. Always resolve the project with `list_projects` first. Do NOT send server-managed fields (`id`, `user_id`, `organization_id`, `status_id`, `created_at`, `updated_at`) — `status_id` is auto-assigned to "Active".
+**Common mistakes:** Sending `project_id` — it is filled from the request scope, and passing one is how endpoints end up in the wrong project. Do NOT send server-managed fields (`id`, `user_id`, `organization_id`, `status_id`, `created_at`, `updated_at`) — `status_id` is auto-assigned to "Active".
 
 ---
 
@@ -534,7 +534,7 @@ Create a knowledge source for grounding Single-Turn `generate_test_set`.
 ### `get_project`
 Get one project by ID.
 
-**CHAIN:** after `list_projects` when you need full project detail before `create_endpoint`.
+**CHAIN:** after `list_projects` when the user asks about a specific project by name. Not needed before `create_endpoint` — that takes its project from the request scope.
 
 **Key parameters:** `project_id` (required)