Python: Core: add experimental session-mode harness context provider

[eavanvalkenburg]

Motivation and Context

Part of the experimental Agent Harness feature, mirroring the .NET work shipped in PR #5310 (.NET: Harness Feature branch) and follow-ups #5404, #5365, #5540. The harness is a set of context providers that make it easier to build long-running, plan-and-execute agents with persistent session state.

This PR introduces the Python counterpart of AgentModeProvider from dotnet/src/Microsoft.Agents.AI/Harness/AgentMode/.

Description

Adds the experimental _harness namespace and the first context provider in it: AgentModeProvider, with get_agent_mode / set_agent_mode helpers and a DEFAULT_MODE_SOURCE_ID constant. New HARNESS value added to the ExperimentalFeature enum; all new public symbols decorated with @experimental(ExperimentalFeature.HARNESS).

.NET	Python (this PR)
AgentModeProvider	AgentModeProvider
AgentModeProviderOptions.Modes	available_modes / mode_descriptions kwargs
AgentMode_Get / AgentMode_Set tools	get_mode / set_mode snake_case tools registered via @tool
GetMode / SetMode helpers	get_agent_mode / set_agent_mode module-level helpers

The provider class name now matches .NET exactly as a refinement for cross-language consistency. The remaining differences are deliberate to match each language's conventions:

• Python keeps kwargs-based construction rather than an options object.
• Python tool names use snake_case.
• Default mode list ("plan", "execute") and the default instruction template are aligned with .NET so cross-language behavior matches.

Also folds in a tiny cosmetic refactor of _sessions.py (try/except ImportError: pass → contextlib.suppress(ImportError)) that was made while developing the harness.

This PR is the first of three splitting the _harness package work apart for review:

1. (this PR) agent-mode context provider — mirrors .NET AgentModeProvider
2. Python: Core: add experimental todo-list harness context provider #5612 — todo-list context provider — mirrors .NET TodoProvider
3. Python: Core: add experimental memory harness context provider #5613 — memory context provider — related to but semantically distinct from .NET FileMemoryProvider

The three modules are independent. They share only the HARNESS enum entry and the (empty) _harness/__init__.py. Whichever lands first creates those; the others rebase trivially.

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the Contribution Guidelines
• All unit tests pass, and I have added new tests where possible
• Is this a breaking change? No — gated behind @experimental(ExperimentalFeature.HARNESS).

[Copilot]

Pull request overview

This PR introduces the first experimental Python harness context provider in agent_framework: a session-backed mode provider that exposes get_mode / set_mode tools and helper APIs for reading and updating the current session mode. It extends the existing feature-staging system with a new HARNESS experimental feature and wires the new symbols into the package surface.

Changes:

• Add SessionModeContextProvider plus session-mode helper APIs/constants under the new _harness namespace.
• Export the new harness symbols from agent_framework.__init__ and add ExperimentalFeature.HARNESS.
• Add focused tests for helper behavior, provider configuration, and tool-driven session mode updates; also refactor _sessions.py to use contextlib.suppress(ImportError).

Reviewed changes

Copilot reviewed 5 out of 6 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
python/packages/core/tests/core/test_harness_mode.py	Adds tests for session-mode helpers, experimental metadata, and provider tools.
python/packages/core/agent_framework/_sessions.py	Minor cleanup replacing try/except ImportError with suppress(ImportError).
python/packages/core/agent_framework/_harness/_mode.py	Implements the new session-mode helpers and SessionModeContextProvider.
python/packages/core/agent_framework/_harness/__init__.py	Creates the new harness package namespace.
python/packages/core/agent_framework/_feature_stage.py	Adds the HARNESS experimental feature enum member.
python/packages/core/agent_framework/__init__.py	Re-exports the new harness symbols from the public package API.

[github-actions]

Automated Code Review

Reviewers: 3 | Confidence: 89% | Result: All clear

Reviewed: Correctness, Security Reliability, Design Approach

---

Automated review by eavanvalkenburg's agents

[moonbox3]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework
_sessions.py	387	33	91%	102–104, 106–107, 124–125, 127–129, 206–207, 297, 558–562, 604, 607, 641, 690, 694, 704, 837, 853, 986, 1000–1001, 1024, 1046, 1056, 1098
packages/core/agent_framework/_harness
_mode.py	87	1	98%	61
TOTAL	32200	3711	88%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
6287	30 💤	0 ❌	0 🔥	1m 44s ⏱️

[eavanvalkenburg]

Addressed reviewer feedback about external mode changes confusing the agent.

When the user changes the mode externally (e.g., via a slash-command), the chat history still shows a prior set_mode tool call from the agent. Updating just the system instructions wasn't enough — the model would anchor on the recent tool call and ignore the new mode.

Aligned with the .NET AgentModeProvider implementation:

• set_agent_mode (the public helper used by external callers) now records the previous mode in provider state when it actually changes.
• On the next before_run, AgentModeProvider pops that flag and injects a user-role notification message into the context (Mode changed: ... switched from "X" to "Y". You must now adjust your behavior...), so the most recent context unambiguously reflects the current mode.
• The agent-driven set_mode tool path bypasses this — it writes current_mode directly so it doesn't trigger a redundant self-notification on its own change.

Tests cover: external change injects exactly one notification and clears the flag; a follow-up run without further changes does not re-inject; the agent's own set_mode tool does not queue a notification; and a no-op set_agent_mode(same_mode) does not queue one either.