web-retriever: ask before persisting extracted data to a file (#792)

Previously web-retriever's core workflow had an unconditional step 7:
always write extracted data to a file. Benchmarking against qa-kitten
(same model, same tasks, only the agent differs) showed this firing
even on tasks with no requested output format, silently dropping a
JSON file the user never asked for.

New behavior:
- If the user explicitly requested a file (any format), always create
  it -- this takes precedence over everything else.
- If the task/user already specified an inline output format (a single
  answer, a specific field, yes/no), answer inline -- skip file logic.
- If the result is small enough to read comfortably in chat, answer
  inline.
- Otherwise (genuinely open-ended, large/structured extraction with no
  explicit file request), ask via ask_user_question whether to save to
  a file before writing anything.

ask_user_question hard-errors instead of prompting when interactive
tools are unavailable (sub-agent invocation via invoke_agent,
autonomous-loop mode, non-interactive/CI) -- web-retriever's primary
advertised invocation path is exactly that sub-agent case, so the
prompt explicitly tells the model not to retry on that error and to
fall back to inline output instead.

Re-verified against the exact repro case: same task/model that
previously produced an unrequested file now answers inline with
file_tool_calls=0.

Jira: PUP-637

Co-authored-by: TJ Webb <thomas.webb@walmart.com>

Author: thomwebb

code_puppy/agents/agent_web_retriever.py

@@ -188,10 +188,35 @@ def get_system_prompt(self) -> str:
    from snapshot/locator results.
 6. **Paginate/crawl**: repeat navigate -> snapshot -> extract across a
    URL list or "next page" control as needed.
-7. **Persist**: write extracted data to a file the user can actually use
-   - `create_file`/`replace_in_file` for JSON, CSV, or markdown tables,
-   matching whatever format was requested (default to JSON if unspecified).
-   Tell the user exactly where the file landed.
+7. **Confirm before persisting -- don't silently drop files on disk.**
+   Writing an unrequested file is a surprising side effect, not a free
+   convenience. Before calling `create_file`/`replace_in_file`, check in
+   this order:
+   - Did the user explicitly ask for a file (any format)? Always create
+     it -- this wins over everything below. You may still add a brief
+     inline summary alongside it.
+   - Otherwise, did the user (or the task itself) already specify an
+     inline output format -- an inline answer, a specific field to
+     report, a yes/no, a single value? If so, just answer inline. Skip
+     the rest of this step; don't ask, don't write a file.
+   - Otherwise, is the result small enough to read comfortably in chat (a
+     handful of fields, one item, a short list)? Answer inline.
+   - Otherwise (genuinely open-ended extraction that produced a
+     multi-row/structured result with no inline format specified and no
+     explicit file request): try `ask_user_question` to check whether
+     they want it saved to a file (and in what format) before writing
+     anything. `ask_user_question` returns an error instead of prompting
+     when interactive tools are unavailable (sub-agent invocation,
+     autonomous-loop mode, non-interactive/CI environment) -- if you get
+     that error, do NOT retry the call. Fall back to answering inline
+     (report the full extracted data as text/markdown) and note briefly
+     that you defaulted to inline output because file confirmation
+     wasn't available.
+   - If the user confirmed a file via `ask_user_question`, or explicitly
+     asked for one up front, use `create_file`/`replace_in_file` for
+     JSON, CSV, or markdown tables, matching whatever format was
+     requested (default to JSON if unspecified). Tell the user exactly
+     where the file landed.
 8. **Save the recipe**: `browser_save_workflow` after a non-trivial
    multi-step scrape/automation succeeds, so the next run doesn't start
    from zero. Name workflows descriptively (include the domain and goal).