Python: default dict function_choice_behavior to auto-invoking for auto and required by DZH166 · Pull Request #14752 · microsoft/semantic-kernel

DZH166 · 2026-11-08T15:45:55Z

Motivation and Context

FunctionChoiceBehavior.from_dict is the parser used whenever a function_choice_behavior is supplied as a dictionary, e.g. PromptExecutionSettings(function_choice_behavior={"type": "auto"}) (see PromptExecutionSettings.parse_function_choice_behavior). It currently pops auto_invoke with a hard default of False:

auto_invoke = data.pop("auto_invoke", False)

Consequences on the current main:

FunctionChoiceBehavior.from_dict({"type": "auto"}) -> maximum_auto_invoke_attempts == 0, so auto_invoke_kernel_functions is False: tools are advertised to the model but are never invoked.
FunctionChoiceBehavior.from_dict({"type": "required"}) -> also 0 attempts, so the model is forced to emit a tool call that the SDK then never executes.
This contradicts every other way of requesting the same behavior:
- FunctionChoiceBehavior.from_string("auto") -> 5 attempts,
- FunctionChoiceBehavior.Auto() -> 5 attempts,
- FunctionChoiceBehavior.Required() -> 1 attempt.

So the dict notation silently downgrades "auto"/"required" to tool-call-advertisement-only mode unless the caller knows about the undocumented auto_invoke key. An agent configured from a serialized settings dict (or any flow that round-trips settings through dictionaries) stops calling tools with no error and no warning.

Description

In FunctionChoiceBehavior.from_dict, default auto_invoke to behavior_type != "none" instead of False, so the dict form defaults exactly like from_string and the Auto/Required classmethods. NoneInvoke keeps auto-invocation off, and an explicit auto_invoke key in the dict still wins (both directions).
Extended the docstring of from_dict to state the default.
Added regression tests:
- test_from_dict_defaults_auto_invoke_consistently (parametrized over auto/required/none/explicit overrides),
- test_from_dict_matches_from_string_for_auto,
- two PromptExecutionSettings tests covering the dict parse path used by parse_function_choice_behavior.

No existing behavior changes for callers that pass auto_invoke explicitly or pass maximum_auto_invoke_attempts in the dict (the existing parametrized from_dict tests keep passing unchanged).

Validation

Red -> green evidence (Python 3.13, Windows):

Red (before the fix), pytest tests/unit/connectors/ai/test_function_choice_behavior.py tests/unit/connectors/ai/test_prompt_execution_settings.py: 4 failed, 38 passed — the four failures are exactly the new implicit-default cases ({"type": "auto"} -> expected 5 got 0, {"type": "required"} -> expected 1 got 0, from_dict vs from_string mismatch, and the settings-level dict parse).
Green (after the fix): 42 passed.
Broader regression run after the fix: 439 passed (kernel, agents, function-choice, settings, chat-completion agent suites) and 218 passed (open_ai connector suite), plus the previously passing suite around the change stayed green.

Contribution Checklist

The code builds clean without any errors or warnings
The PR follows the SK Contribution Guidelines and the pre-submission formatting script raises no violations
All unit tests pass, and I have added new tests where possible
I didn't break anyone 🙂

AI assistance

This change was prepared with the help of an AI coding agent (code analysis, patch drafting, and test writing); the final code, tests, and validation were reviewed by a human before submission.

…to and required

Copilot

🟡 Changes recommended

The documentation incorrectly promises that auto_invoke=False always disables invocation, omitting the explicit attempt-count override and NoneInvoke exception.

1 open issue

Document exceptions to the auto-invoke disable guarantee · New

What changed in this PR

This PR aligns Python’s dictionary-based function-choice configuration with the existing string and factory defaults, enabling automatic tool invocation for auto and required.

Changes:

Defaults auto_invoke based on the behavior type while preserving explicit settings.
Documents the new default.
Adds regression tests for dictionary parsing, overrides, and PromptExecutionSettings integration.

File	Description
`python/tests/unit/connectors/ai/test_prompt_execution_settings.py`	Tests default invocation and explicit opt-out through settings parsing.
`python/tests/unit/connectors/ai/test_function_choice_behavior.py`	Tests behavior defaults, overrides, and consistency with string parsing.
`python/semantic_kernel/connectors/ai/function_choice_behavior.py`	Updates the dictionary default and its documentation.

_{Review effort: Balanced}

DZH166 · 2026-11-08T16:24:34Z

+
+        The default for `auto_invoke` matches `from_string` and the classmethods:
+        auto-invocation is enabled for the "auto" and "required" types and disabled
+        for "none". An explicit `auto_invoke` key in the dict always wins.


Correct — both exceptions are real, and the docstring is now precise (2b12e2b):

An explicit `auto_invoke` key overrides that default, with two exceptions: * an explicit `maximum_auto_invoke_attempts` in the same dict wins, because the classmethods only fill the count in when it is absent -- so `{"type": "auto", "auto_invoke": False, "maximum_auto_invoke_attempts": 3}` still auto-invokes. Leave the count out when `auto_invoke=False` is meant to disable invocation for "auto" and "required". * the "none" behavior ignores `auto_invoke` entirely: `NoneInvoke()` always sets the attempt count to 0, whatever the dict says.

While there I also made from_dict() copy the input (data = dict(data)): it pop()s type, auto_invoke, functions and filters, so the caller's config dict came back emptied. Behavior is otherwise unchanged.

…sable invocation from_dict() lets an explicit maximum_auto_invoke_attempts win over auto_invoke, and the none behavior ignores auto_invoke entirely. Spell both out, and copy the input dict so from_dict() does not strip the caller's keys.

Muhammad Abdullah Butt (1AbdullahButt)

Looks correct. Here's what I checked.

The default. auto_invoke = data.pop("auto_invoke", behavior_type != "none") (function_choice_behavior.py:209) passes True to Auto()/Required(). Those then setdefault the attempt count to DEFAULT_MAX_AUTO_INVOKE_ATTEMPTS and 1. That matches from_string("auto") and from_string("required"). NoneInvoke() ignores the flag and sets 0, as the docstring says.
Existing callers. The only caller in the package is PromptExecutionSettings.parse_function_choice_behavior (prompt_execution_settings.py:49). The existing from_dict tests and the function_choice_yaml/defined_function.yaml sample all pass maximum_auto_invoke_attempts explicitly. So on main they already auto-invoked through setdefault, and this change doesn't move them. Only dicts without a count change behaviour, which is the bug.
data = dict(data). It's a shallow copy, so filters and functions are still the caller's objects. Nothing here mutates them (_combine_filter_dicts builds a new dict), so that's fine.
git merge-tree against main is clean.

I didn't run the tests.

Two notes:

This turns on function execution for any existing {"type": "auto"} or {"type": "required"} settings dict without a count, including YAML prompt templates written that way. Before, those only advertised the tools. That's the intended fix, but it is a behaviour change that runs code. A line in the release notes would help, pointing to auto_invoke: false as the way to keep advertise-only.
type_map[behavior_type] is case-sensitive, while from_string lowercases (data.lower()). So {"type": "Auto"} raises KeyError while "Auto" as a string works. That's not from this PR, but behavior_type = data.pop("type", "auto").lower() would line the two up, since this PR is about making them agree.

Reviewed with help from Claude Code.

Python: default dict function_choice_behavior to auto-invoking for au…

1d02a01

…to and required

DZH166 requested a review from a team as a code owner November 8, 2026 15:45

Copilot AI balanced review requested due to automatic review settings November 8, 2026 15:45

DZH166 temporarily deployed to github-app-auth November 8, 2026 15:45 — with GitHub Actions Inactive

semantic-kernel-automation Bot added the python Pull requests for the Python Semantic Kernel label Nov 8, 2026

Copilot started reviewing on behalf of DZH166 November 8, 2026 15:46

View session

Copilot AI reviewed Nov 8, 2026

View reviewed changes

DZH166 temporarily deployed to github-app-auth November 8, 2026 16:24 — with GitHub Actions Inactive

Muhammad Abdullah Butt (1AbdullahButt) reviewed Nov 12, 2026

View reviewed changes

Navigation Menu

Repository navigation

Python: default dict function_choice_behavior to auto-invoking for auto and required - #14752

Python: default dict function_choice_behavior to auto-invoking for auto and required#14752
DZH166 wants to merge 2 commits into
microsoft:mainfrom
DZH166:codex/fcb-from-dict-auto-invoke-20261108

DZH166 commented Nov 8, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

DZH166 Nov 8, 2026

Uh oh!

Muhammad Abdullah Butt (1AbdullahButt) left a comment

Uh oh!

This branch was previously deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

DZH166 commented Nov 8, 2026

Motivation and Context

Description

Validation

Contribution Checklist

AI assistance

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

DZH166 Nov 8, 2026

Choose a reason for hiding this comment

Uh oh!

Muhammad Abdullah Butt (1AbdullahButt) left a comment

Choose a reason for hiding this comment

Uh oh!

This branch was previously deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants