Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@code_puppy/agents/event_stream_handler.py`:
- Around line 140-150: The merge function _merge_tool_name can leak partial
streaming names (e.g. "agent_share_your_reasoning") before the final value
arrives; change _merge_tool_name to detect the special reasoning tool string
(e.g. "agent_share_your_reasoning") and avoid returning any intermediate/partial
concatenation that would be a prefix of that special name — instead keep
returning current_name until the merged result equals the full special name, and
only then return the full name; apply the same suppression logic in the other
merging code path noted (lines ~385-404) so partial prefixes of the reasoning
tool are never emitted.

In `@code_puppy/cli_runner.py`:
- Around line 572-575: The code registers a runtime with
register_active_interactive_runtime(runtime) but only clears it on the normal
exit path, which can leave a stale global runtime after exceptions; to fix,
ensure you always call clear_active_interactive_runtime(runtime) in a finally
block paired with the try that follows the register so the runtime is removed
regardless of errors (referencing PromptRuntimeState,
register_active_interactive_runtime, clear_active_interactive_runtime, and
runtime.mark_idle) — move or add a finally that invokes
clear_active_interactive_runtime(runtime) and performs any necessary cleanup
after runtime.mark_idle().
- Around line 1194-1202: The code currently ignores the return value of
BackgroundInteractiveCommand.run(), so even when it returns False (indicating
failure) the code still emits a "completed" success lifecycle; change the await
asyncio.to_thread(command_result.run, command_result.cancel_event) call to
capture the return value (e.g., result = await
asyncio.to_thread(command_result.run, command_result.cancel_event)) and only
call emit_interject_queue_lifecycle(runtime, "completed", item=source_item,
level="success") when the captured result is truthy and
command_result.cancel_event.is_set() is False (i.e., if result and not
command_result.cancel_event.is_set()) so failed runs that return False do not
emit a success completion.
- Around line 1289-1299: The code cancels the active run via
cancel_active_run("interject") before checking whether runtime.request_interject
succeeded, which can drop the current task if the queue is full; to fix, call
runtime.request_interject(stripped_task,
allow_command_dispatch=allow_command_dispatch) first and inspect its ok result
(and position/item), and only call cancel_active_run("interject") if ok is True;
if ok is False, call emit_warning with the queue limit message (using
get_queue_limit()) and do not cancel the active run. Ensure this change is
applied where requested_action == "interject" and references the same symbols
(requested_action, runtime.request_interject, cancel_active_run, emit_warning,
get_queue_limit).
- Around line 1859-1867: The code treats the return of
run_prompt_with_attachments as a single object but that function returns a tuple
(result, task); change the call to unpack it (e.g., response, task = await
run_prompt_with_attachments(...)), then preserve the existing None check on
response (if response is None: return) before using response.output to set
agent_response; ensure prompt-only mode which returns (None, None) is handled by
the unpack and early return to avoid accessing response.output when response is
None.

In `@code_puppy/tools/command_runner.py`:
- Around line 1218-1225: The call to
get_active_interactive_runtime().notify_shell_started() is performed before
entering the try/finally, so if notify_shell_started or later
notify_shell_finished raises, it can short-circuit spinner/keyboard cleanup;
move notify_shell_started() into the main try block after acquiring keyboard
context (or after deciding release_keyboard_context) and ensure
notify_shell_finished() is invoked from the finally block but wrapped in its own
try/except to swallow/log any exceptions so they don't prevent cleanup;
reference get_active_interactive_runtime, notify_shell_started,
notify_shell_finished, _acquire_keyboard_context, and the
release_keyboard_context flag when applying the change.
- Around line 288-295: The _normalize_shell_cwd helper currently strips
leading/trailing whitespace which alters valid paths; change it to only collapse
whitespace-only inputs to None by returning None when cwd is None or cwd.strip()
is empty, otherwise return the original cwd unchanged (preserving any
intentional leading/trailing spaces) so callers receive the original non-empty
value.

In `@tests/test_cli_runner_full_coverage.py`:
- Around line 130-142: The prompt_side_effect helper currently wraps every
non-PromptSubmission value into PromptSubmission(action="submit", text=value),
which converts a test-provided None (meant to represent cancellation) into a
submission; modify prompt_side_effect so that after resolving value from
input_fn it returns None unchanged if value is None, returns the value unchanged
if isinstance(value, PromptSubmission), and only otherwise wraps into
PromptSubmission(action="submit", text=value); reference prompt_side_effect,
input_fn and PromptSubmission when making the change.

---

Outside diff comments:
In `@code_puppy/plugins/antigravity_oauth/register_callbacks.py`:
- Around line 258-266: The auth flow currently treats authentication as
successful even if add_models_to_config(...) fails; update
_perform_authentication to only return True when both the auth step and
add_models_to_config succeed — if add_models_to_config returns False, emit the
warning (emit_warning) and return False so the session does not switch to an
unregistered model; locate the block that calls add_models_to_config and adjust
the control flow so success requires both the access token/project_id handling
and the model registration call (refer to add_models_to_config,
_perform_authentication, emit_success, emit_warning, and emit_info to find the
correct spot).

---

Nitpick comments:
In `@code_puppy/command_line/core_commands.py`:
- Around line 176-218: The function handle_tutorial_command currently returns
either a bool or a BackgroundInteractiveCommand but lacks a type annotation;
update its signature to include an explicit return type (e.g., Union[bool,
BackgroundInteractiveCommand]) and add the necessary typing import (from typing
import Union) so IDEs and linters understand the union return type; keep the
existing behavior and symbols (handle_tutorial_command and
BackgroundInteractiveCommand) unchanged elsewhere.

In `@tests/tools/test_file_modifications_extended.py`:
- Around line 225-244: Add companion tests to cover the other validation paths
for register_replace_in_file: create two new test functions named
test_register_replace_in_file_rejects_missing_new_str and
test_register_replace_in_file_rejects_non_string_values that mirror the existing
test_register_replace_in_file_rejects_missing_old_str structure, register the
tool via register_replace_in_file(Agent()), call the registered
"replace_in_file" function with replacements missing "new_str" in the first test
and with a non-string "old_str" (e.g., integer) in the second, and assert
result["success"] is False, result["changed"] is False and that the error
message contains "new_str" for the first test (and appropriate validation
indication for the second).

Verify each finding against the current code and only fix it if needed.

Nitpick comments:
In `@tests/test_cli_runner_full_coverage.py`:
- Around line 1521-1528: The test uses a long fixed sleep inside the fake_run
coroutine (function fake_run) when handling the "first task" prompt, which
should be replaced by waiting on an event to avoid timing flakiness: create an
asyncio.Event or Future (e.g., first_task_continue) and in fake_run, replace
await asyncio.sleep(10) with await first_task_continue.wait() (or awaiting the
Future); then signal or cancel that event/future from the test flow when you
want the first task to continue or be cancelled (use the existing
first_task_started Event to know when to trigger it). Update references to
first_task_started and any assertions around render_order/started_prompts so the
test coordinates via the new first_task_continue event instead of a fixed sleep.
- Around line 190-192: The test currently only checks the emitted payload text
but not that the emitter was invoked exactly once; update the test(s) that
inspect message_bus.emit (e.g., the block capturing emitted =
message_bus.emit.call_args[0][0]) to also assert a single emission using
message_bus.emit.assert_called_once() (or assert message_bus.emit.call_count ==
1) before examining emitted.text, and apply the same single-call assertion to
the similar cases at the other locations (the queue and command-completion
lifecycle tests).