Automate browser based workflows with AI
Python aiapiautomationbrowserbrowser-automationcomputergptllmplaywrightpowerautomatepuppeteerpythonrpaseleniumvisionworkflow
feat: dynamic download selectors for file_download blocks in for-loops (#5077)
pedrohsdbf894be1344a5
diff --git a/skyvern/core/script_generations/generate_script.py b/skyvern/core/script_generations/generate_script.pyindex 0a7e4b4d4..3ea075288 100644--- a/skyvern/core/script_generations/generate_script.py+++ b/skyvern/core/script_generations/generate_script.py@@ -1147,6 +1147,7 @@ def _build_block_fn(actions: list[dict[str, Any]],value_to_param: dict[str, str] | None = None,use_semantic_selectors: bool = False,+ is_in_for_loop: bool = False,) -> FunctionDef:# Check if this block is form-filling (4+ input_text/select actions)if use_semantic_selectors and _is_form_filling_block(actions):@@ -1162,28 +1163,37 @@ def _build_block_fn(if block.get("url"):body_stmts.append(cst.parse_statement(f"await page.goto({repr(block['url'])})"))- for act in actions:- if act["action_type"] in [- ActionType.COMPLETE,- ActionType.TERMINATE,- ActionType.NULL_ACTION,- ]:- continue-- # For extraction blocks, assign extract action results to output variable- assign_to_output = act["action_type"] == "extract"+ # For file_download blocks inside for-loops, generate a dynamic click that uses+ # per-iteration context instead of hardcoded xpath/prompt from iteration 0.+ block_type = block.get("block_type")+ if is_in_for_loop and block_type == "file_download":body_stmts.append(- _action_to_stmt(- act,- block,- assign_to_output=assign_to_output,- value_to_param=value_to_param,- use_semantic_selectors=use_semantic_selectors,+ cst.parse_statement(+ 'await page.click(selector=context.download_selector(), prompt=context.prompt, ai="fallback")'))+ else:+ for act in actions:+ if act["action_type"] in [+ ActionType.COMPLETE,+ ActionType.TERMINATE,+ ActionType.NULL_ACTION,+ ]:+ continue++ # For extraction blocks, assign extract action results to output variable+ assign_to_output = act["action_type"] == "extract"+ body_stmts.append(+ _action_to_stmt(+ act,+ block,+ assign_to_output=assign_to_output,+ value_to_param=value_to_param,+ use_semantic_selectors=use_semantic_selectors,+ )+ )# add complete action- block_type = block.get("block_type")if block_type in SCRIPT_TASK_BLOCKS_WITH_COMPLETE_ACTION:complete_action = {"action_type": "complete"}body_stmts.append(_action_to_stmt(complete_action, block))@@ -2818,6 +2828,7 @@ async def generate_workflow_script_python_code(inner_actions,value_to_param=value_to_param,use_semantic_selectors=use_semantic_selectors,+ is_in_for_loop=True,)inner_block_code = cst.Module(body=[inner_fn_def]).codediff --git a/skyvern/core/script_generations/skyvern_page.py b/skyvern/core/script_generations/skyvern_page.pyindex 774c91f89..b74ef2355 100644--- a/skyvern/core/script_generations/skyvern_page.py+++ b/skyvern/core/script_generations/skyvern_page.py@@ -2917,3 +2917,43 @@ class RunContext:self.actions_and_results: list[tuple[Action, list[ActionResult]]] = []# Pre-extracted values from applicant context, keyed by canonical category nameself.extracted_params: dict[str, str | None] = extracted_params or {}++ @property+ def prompt(self) -> str | None:+ """Return the per-iteration prompt from SkyvernContext (set by script_service per loop iteration)."""+ ctx = skyvern_context.current()+ return ctx.prompt if ctx else None++ @property+ def loop_value(self) -> Any | None:+ """Return the current loop iteration value from SkyvernContext.loop_metadata."""+ ctx = skyvern_context.current()+ if ctx and ctx.loop_metadata:+ return ctx.loop_metadata.get("current_value")+ return None++ def download_selector(self) -> str | None:+ """Build a CSS selector targeting a download link from the current loop value.++ Scans the loop_value dict for URL-like strings, extracts the filename,+ and returns a selector like ``a[href*="filename.pdf"]``. Returns None+ if no URL is found.+ """+ value = self.loop_value+ if not value or not isinstance(value, dict):+ return None+ for v in value.values():+ if not isinstance(v, str):+ continue+ # Match URL-like strings (http/https or paths with file extensions)+ if re.match(r"https?://", v) or re.match(r"/.*\.\w+", v):+ # Extract filename from URL path+ filename = v.rstrip("/").rsplit("/", 1)[-1]+ # Strip query params+ filename = filename.split("?")[0]+ if filename and "." in filename:+ # Escape CSS special characters to avoid malformed selectors+ filename = re.sub(r'["\[\]\\]', "", filename)+ if filename:+ return f'a[href*="{filename}"]'+ return Nonediff --git a/skyvern/forge/prompts/skyvern/script-reviewer.j2 b/skyvern/forge/prompts/skyvern/script-reviewer.j2index 0b6f7f54b..f55974bd6 100644--- a/skyvern/forge/prompts/skyvern/script-reviewer.j2+++ b/skyvern/forge/prompts/skyvern/script-reviewer.j2@@ -27,7 +27,9 @@ You are a script reviewer for a browser automation system. Your job is to updateawait page.fill(selector='...', ai='proactive', prompt='Select how you heard about this job')```10. **URL PATTERNS**: When different variants come from different URL hosts or paths, add `url_patterns` to the `page.classify()` call for FREE Tier-0 disambiguation. Use regex on the URL path: `url_patterns={"variant_a": "example\\.com/form-a"}`. URL patterns are checked BEFORE text patterns and cost nothing.-11. **FOR-LOOP BLOCKS**: When code runs inside a for-loop (e.g. by iterating over extracted items), every `page.click()`, `page.goto()`, or `page.download_file()` call MUST use the loop variable (`current_value`) to target the specific item. Never use a static selector or generic prompt that would resolve to the same element on every iteration. Prefer `page.goto(current_value["url"])` for direct URLs, or include the item's title/identifier in the selector: `selector=f'a:has-text("{current_value["title"]}")'`.+11. **FOR-LOOP BLOCKS**: When code runs inside a for-loop (e.g. by iterating over extracted items), every `page.click()`, `page.goto()`, or `page.download_file()` call MUST use per-iteration data to target the specific item. Never use a static selector or generic prompt that would resolve to the same element on every iteration.+ - **File download blocks**: Use `context.download_selector()` which builds a CSS selector (e.g. `a[href*="filename.pdf"]`) from URL-like fields in the current loop value. Combine with `context.prompt` for AI fallback: `await page.click(selector=context.download_selector(), prompt=context.prompt, ai="fallback")`. If `download_selector()` returns `None` (no URL in loop value), fall back to a text-based selector using `context.loop_value`.+ - **Other blocks**: Use `context.loop_value` (dict) to build selectors. Prefer `page.goto(context.loop_value["url"])` for direct URLs, or include the item's title/identifier in the selector: `selector=f'a:has-text("{context.loop_value["title"]}")'`.## Block Navigation Goal```@@ -384,6 +386,12 @@ async def block_fn(page, context):### Properties- `page.url` — Current page URL (str property, free, no LLM call)+### Context Properties (for-loop blocks only)+These properties are available on the `context` object when code runs inside a for-loop:+- `context.loop_value` — Current loop iteration value (dict). Contains the extracted data for this iteration (e.g. `{"title": "...", "download_url": "...", "unique_id": "..."}`).+- `context.prompt` — Per-iteration navigation goal/prompt (str or None). Set automatically from the block's navigation_goal with the current loop value substituted.+- `context.download_selector()` — Builds a CSS selector targeting a download link from `context.loop_value`. Scans the dict values for URL-like strings, extracts the filename, and returns e.g. `'a[href*="filename.pdf"]'`. Returns `None` if no URL is found — in that case, build a selector manually from `context.loop_value`.+### DO NOT USE (these do NOT exist and will crash at runtime)- ~~`page.completed`~~ — Does NOT exist. Use `await page.complete()` to finish.- ~~`page.current_url`~~ — Does NOT exist. Use `page.url` instead.diff --git a/skyvern/services/script_service.py b/skyvern/services/script_service.pyindex 71b845732..e308a6a4f 100644--- a/skyvern/services/script_service.py+++ b/skyvern/services/script_service.py@@ -1820,8 +1820,52 @@ async def download(try:await _prepare_cached_block_inputs(cache_key, prompt)++ # Count downloaded files before running cached function so we can+ # verify that the download actually produced a new file.+ org_id = context.organization_id or ""+ run_id = context.workflow_run_id or ""+ files_before: list = []+ files_before_ok = False+ try:+ async with asyncio.timeout(GET_DOWNLOADED_FILES_TIMEOUT):+ files_before = await app.STORAGE.get_downloaded_files(+ organization_id=org_id,+ run_id=run_id,+ )+ files_before_ok = True+ except asyncio.TimeoutError:+ LOG.warning("Timeout getting downloaded files before cached download")+await _run_cached_function(cached_fn)+ # Verify a new file was actually downloaded.+ # Retry briefly — file may not be visible in storage immediately after the click.+ files_after: list = []+ files_after_ok = False+ for _attempt in range(3):+ try:+ async with asyncio.timeout(GET_DOWNLOADED_FILES_TIMEOUT):+ files_after = await app.STORAGE.get_downloaded_files(+ organization_id=org_id,+ run_id=run_id,+ )+ files_after_ok = True+ except asyncio.TimeoutError:+ LOG.warning("Timeout getting downloaded files after cached download")+ if len(files_after) > len(files_before):+ break+ if _attempt < 2:+ await asyncio.sleep(2)++ # Only raise if both calls succeeded — if either timed out, skip+ # the check to avoid spurious AI fallbacks under degraded storage.+ if files_before_ok and files_after_ok and len(files_after) <= len(files_before):+ raise Exception(+ "Cached download function did not produce a new file. "+ f"Files before: {len(files_before)}, after: {len(files_after)}"+ )+# Update block status to completed if workflow block was createdif workflow_run_block_id:await _update_workflow_block(