[pull] develop from baserow:develop

backend/pyproject.toml

@@ -22,7 +22,7 @@ dependencies = [
     "psycopg2-binary==2.9.11",
     "Faker==40.1.0",
     "gunicorn==23.0.0",
-    "uvicorn[standard]==0.40.0",
+    "uvicorn[standard]==0.52.4",
     "websockets==16.0.0",
     "requests==2.34.2",
     "itsdangerous==2.2.0",

e2e-tests/fixtures/user.ts

@@ -44,7 +44,7 @@ export async function createUser(
 ): Promise<User> {
   const password = faker.internet.password();
   const response: any = await getClient().post("user/", {
-    name: faker.name.fullName(),
+    name: faker.person.fullName(),
     email: faker.internet.email(),
     password,
     language: "en",

e2e-tests/package.json

@@ -17,7 +17,7 @@
     "codegen": "playwright codegen"
   },
   "dependencies": {
-    "@faker-js/faker": "7.6.0",
+    "@faker-js/faker": "10.5.0",
     "@nuxt/test-utils": "^3.21.0",
     "@playwright/test": "^1.48.0",
     "axios": "1.18.0",

e2e-tests/yarn.lock

@@ -44,10 +44,10 @@
     picocolors "^1.0.0"
     sisteransi "^1.0.5"
 
-"@faker-js/faker@7.6.0":
-  version "7.6.0"
-  resolved "https://registry.yarnpkg.com/@faker-js/faker/-/faker-7.6.0.tgz#9ea331766084288634a9247fcd8b84f16ff4ba07"
-  integrity sha512-XK6BTq1NDMo9Xqw/YkYyGjSsg44fbNwYRx7QK2CuoQgyy+f1rrTDHoExVM5PsyXCtfl2vs2vVJ0MN0yN6LppRw==
+"@faker-js/faker@10.5.0":
+  version "10.5.0"
+  resolved "https://registry.yarnpkg.com/@faker-js/faker/-/faker-10.5.0.tgz#d2f6a8c7f08d087ac5f077d6babd0821edf24c03"
+  integrity sha512-bsxD8WLS5lIj7aaoCx1YJkktqYj5vlBUE6HWzu2Q51ksrGJ0H737ECCKlFU7Yf8Br45z9t99frBp/J7kzbMPAg==
 
 "@jridgewell/gen-mapping@^0.3.5":
   version "0.3.13"

enterprise/backend/src/baserow_enterprise/assistant/assistant.py

@@ -363,7 +363,10 @@ async def _save_ai_response(
         await AssistantChatPrediction.objects.acreate(
             human_message=human_msg,
             ai_response=ai_msg,
-            prediction={"answer": answer},
+            prediction={
+                "answer": answer,
+                "posthog_trace_id": self._telemetry.trace_id,
+            },
         )
         return AiMessage(
             id=ai_msg.id,
@@ -469,7 +472,11 @@ async def _run_agent(
         """
 
         try:
-            with self._telemetry.trace(self._chat, user_prompt) as tracer:
+            with self._telemetry.trace(
+                self._chat,
+                user_prompt,
+                cancelled_by_user=lambda: self._tool_helpers.is_cancelled,
+            ) as tracer:
                 answer, run_result = await self._run_agent_with_retries(
                     user_prompt, message_history, queue
                 )
@@ -635,10 +642,16 @@ def _looks_like_json_tool_call(text: str) -> bool:
         """Return True if *text* looks like a tool call dumped as JSON.
 
         Checks for ``{"name": ..., "arguments": ...}`` pattern in the first
-        200 chars. Does not require valid JSON (the output may be truncated).
+        200 chars after an optional code fence. Does not require valid JSON
+        because the output may be truncated.
+
+        :param text: The final text returned by the agent.
+        :return: Whether the text appears to contain an unexecuted tool call.
         """
 
         stripped = text.strip()
+        if stripped.startswith("```"):
+            stripped = stripped.split("\n", 1)[-1].strip()
         return (
             bool(stripped)
             and stripped[0] == "{"

enterprise/backend/src/baserow_enterprise/assistant/prompts.py

@@ -8,17 +8,24 @@
 """
 
 RULES = """\
+<contracts>
+Three invariants. A tool call that breaks one is invalid — do not send it.
+A. IDs. Every `*_id` argument must carry a real ID you have in hand: returned by a tool call in this conversation, present in `<ui_context>`, or given to you by the user. Never invent, guess, or carry over an ID from a different resource. Baserow IDs start at 1, so 0 is never an ID. If you do not have the ID yet, call the list_*/create_* tool that returns it, then pass back the exact value it returned.
+B. Modes. Have tools → call them. Each tool is owned by exactly one mode, and `<available_tools>` is the authority: it names what the current `<mode>` can call and what each other mode owns. To use a tool owned by another mode, call switch_mode first. If a tool call comes back rejected as an unknown name, that means wrong mode, not missing feature — re-read `<available_tools>`, switch to the owning mode, and retry it once. Only describe manual UI steps once you have confirmed no mode owns a tool for the action; `<limitations>` lists what genuinely cannot be done in any mode.
+C. Payloads. Send every required argument on the first attempt, not only the ones you are confident about. For create_*, update_* and setup_* tools the payload is the point of the call: one carrying just IDs and a `thought` is always incomplete.
+</contracts>
 <rules>
 1. Use the `thought` parameter on EVERY tool call. It is shown to the user, so write it as a brief user-facing status (e.g. "Checking existing pages" not "Calling list_pages to get page IDs"). Never use tool names or internal references.
-2. Have tools → call them. No tools in current mode → check other modes before saying something is not possible. If another mode has the tool, switch_mode and use it. Only explain manual UI steps if no mode covers the action.
-3. One tool per turn. Wait for the result. Never reply and call a tool in same turn.
-4. Request priority: action > follow-up (reuse prior IDs, never search docs) > question. When a tool result contains next_steps, act on them immediately — do not ask for permission to continue.
-5. You start in the mode matching your UI context (database/application/automation). If the user asks a how-to or feature question, call switch_mode("explain"), then search_user_docs.
-6. After finishing the tool calls in a different mode (not just after switching — after the actual work is done and results received), switch back to the original domain mode (check <mode> and <ui_context>).
-7. Reply in concise Markdown. Never expose raw JSON or internal IDs unless asked.
-8. Before starting work, use list_* to understand what exists and avoid duplicates. But don't list resources you just created — create_* tools already return IDs and refs. When a request references resources by name/ID, verify they exist before building on them. If not found, ask — don't guess. But when the task *requires* creating resources in another domain (e.g. building an app that needs new tables), switch_mode and create them yourself — don't ask the user to do it manually.
-9. Before responding to the user, verify ALL parts of `<current_task>` are addressed. If anything is missing, continue working.
-10. At the start, verify the request fits the current UI context (e.g. don't add "Inquiries" table to a "Project Management" DB). If it doesn't match and not explicitly requested, ask the user which target to use.
+2. One tool per turn. Wait for the result. Never reply and call a tool in same turn.
+3. Request priority: action > follow-up (reuse prior IDs, never search docs) > question. When a tool result contains next_steps, act on them immediately — do not ask for permission to continue.
+4. You start in the mode matching your UI context (database/application/automation). If the user asks a how-to or feature question, call switch_mode("explain"), then search_user_docs.
+5. After finishing the tool calls in a different mode (not just after switching — after the actual work is done and results received), switch back to the original domain mode (check <mode> and <ui_context>).
+6. Reply in concise Markdown. Never expose raw JSON or internal IDs unless asked.
+7. Before starting work, use list_* to understand what exists and avoid duplicates. But don't list resources you just created — create_* tools already return IDs and refs. When a request references resources by name/ID, verify they exist before building on them. If not found, ask — don't guess. But when the task *requires* creating resources in another domain (e.g. building an app that needs new tables), switch_mode and create them yourself — don't ask the user to do it manually.
+8. Before responding to the user, verify ALL parts of `<current_task>` are addressed. If anything is missing, continue working.
+9. At the start, verify the request fits the current UI context (e.g. don't add "Inquiries" table to a "Project Management" DB). If it doesn't match and not explicitly requested, ask the user which target to use.
+10. When a task needs a database, application, or automation that does not exist yet, call create_builders first and build on the ID it returns (contract A).
+11. For database formula creation or repair, call generate_formula so the result is validated. Never return or save a handwritten formula. Use save_to_field=true when the user asks to create, fix, save, or apply it; use false only when they explicitly want formula text without changing the table.
 </rules>
 """
 

enterprise/backend/src/baserow_enterprise/assistant/retrying_model.py

@@ -67,6 +67,8 @@
 def _is_transient_provider_error(exc: Exception) -> bool:
     """Return True for provider errors that are transient and safe to retry."""
 
+    if isinstance(exc, ModelHTTPError) and exc.status_code == 429:
+        return True
     msg = str(exc)
     return any(needle in msg for needle in _RETRYABLE_MESSAGES)
 
@@ -413,6 +415,8 @@ async def request(
                     ):
                         raise
                     delay = self._delay_for(attempt)
+                    if isinstance(exc, ModelHTTPError) and exc.retry_after is not None:
+                        delay = min(exc.retry_after, self.max_delay)
                     logger.warning(
                         "[assistant] Model request failed (attempt {}/{}), "
                         "retrying in {:.1f}s: {}",