Python: Add a HarnessAgent with available features and sample

[westey-m]

Motivation and Context

#6034

Description

• Add a HarnessAgent class with available features
• Add a basic sample to demonstrate usage (better console to follow)
• Add a compaction strategy based on context window size

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? If yes, add "[BREAKING]" prefix to the title of the PR.

[Copilot]

Pull request overview

Adds a new Python HarnessAgent convenience wrapper (with default context providers and compaction), introduces a context-window–based compaction strategy, and provides a sample demonstrating the bundled agent pipeline.

Changes:

• Introduce HarnessAgent (experimental) that assembles history + compaction + todo/mode/memory/skills providers and optionally telemetry.
• Add ContextWindowCompactionStrategy plus tests covering threshold-triggered tool eviction and truncation.
• Add a HarnessAgent sample (README + interactive research loop) and minor formatting/import tweaks elsewhere.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
python/samples/02-agents/harness/README.md	Adds documentation for the HarnessAgent sample and feature overview.
python/samples/02-agents/harness/harness_research.py	Adds an interactive research assistant sample using HarnessAgent + Foundry web search.
python/packages/foundry_hosting/tests/test_responses.py	Minor whitespace adjustment in tests.
python/packages/foundry_hosting/agent_framework_foundry_hosting/_responses.py	Import cleanup/reordering.
python/packages/core/tests/core/test_harness_agent.py	Adds unit tests validating HarnessAgent assembly and options.
python/packages/core/tests/core/test_compaction.py	Adds unit tests for ContextWindowCompactionStrategy.
python/packages/core/agent_framework/_harness/_agent.py	Implements HarnessAgent plus assembly helpers and defaults.
python/packages/core/agent_framework/_compaction.py	Adds ContextWindowCompactionStrategy and exports it.
python/packages/core/agent_framework/init.py	Exports HarnessAgent, DEFAULT_HARNESS_INSTRUCTIONS, and ContextWindowCompactionStrategy.
.devcontainer/devcontainer.json	Updates the Python devcontainer image/features and adds VS Code extensions.

[github-actions]

Automated Code Review

Reviewers: 4 | Confidence: 88%

✓ Correctness

The core logic of ContextWindowCompactionStrategy and HarnessAgent is correct. The TokenBudgetComposedStrategy properly re-annotates tokens between phases, and the two-phase compaction pipeline works as designed. The main new issue is a duplicate from pathlib import Path import introduced in _responses.py, which is clearly a merge/editing artifact. The previously-flagged issues (_assemble_instructions empty-string fallback, sample docstring placement, test coverage for disable_memory, and test docstring/assertion mismatch) remain valid and unresolved.

✓ Security Reliability

No significant security or reliability issues found. The PR introduces proper input validation in both ContextWindowCompactionStrategy and HarnessAgent constructors, correctly bounds threshold parameters, and delegates to well-tested existing compaction infrastructure. The only concrete code issue is a duplicate import introduced in _responses.py, which is harmless at runtime but indicates an incomplete merge.

✓ Test Coverage

The PR adds a HarnessAgent class and ContextWindowCompactionStrategy with good assembly/construction test coverage, but the primary runtime API method HarnessAgent.run() has zero test coverage. The existing tests verify assembly and configuration but never execute the agent's main behavior. Additionally, get_session() is untested and there is no test verifying that both compaction phases (tool eviction + truncation) fire together in a single call.

✓ Design Approach

The main design issue I found is in the new HarnessAgent sample: it instructs the model to save its final report to file memory, but the sample never configures a memory store, so the HarnessAgent assembly path never adds MemoryContextProvider. That leaves the sample advertising a capability it does not actually expose.

---

Automated review by westey-m's agents

[moonbox3]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework
_compaction.py	664	55	91%	104, 158–159, 177–178, 193–196, 203–205, 229, 243, 250, 265–266, 320, 327, 343, 393, 417, 447, 494, 500, 502, 521, 565–570, 582, 654, 656, 671, 715, 777, 899, 960, 962, 978, 985, 990, 1005, 1013, 1107, 1110, 1128, 1144, 1240, 1267, 1272, 1351
packages/core/agent_framework/_harness
_agent.py	70	4	94%	129, 323–324, 326
packages/foundry_hosting/agent_framework_foundry_hosting
_responses.py	684	105	84%	183–186, 251, 328–329, 339, 376, 431, 445, 497, 500–504, 523, 526, 532, 534, 555–557, 586–588, 593, 595, 602, 604–605, 607, 609, 615–618, 620–622, 626, 629, 634–640, 643–644, 646–647, 655–660, 939, 952, 1418–1420, 1422, 1469–1470, 1472–1473, 1475–1476, 1478–1479, 1484, 1493, 1496–1498, 1500, 1514, 1527, 1572–1573, 1575, 1579–1583, 1585, 1592–1593, 1595–1596, 1602, 1604–1608, 1615, 1621, 1643, 1649
TOTAL	36397	4312	88%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
7273	34 💤	0 ❌	0 🔥	1m 53s ⏱️