Python: Reject candidate_count>1 in the Gemini chat client instead of silently dropping candidates by Wilson0511 · Pull Request #9929 · microsoft/agent-framework

Wilson_Jiang (Wilson0511) · 2026-11-01T09:06:09Z

Description

candidate_count flows to the Gemini API through the generic option passthrough, but both the streaming and non-streaming parsers read candidates[0] only — so candidate_count > 1 is billed for every candidate while all but the first are silently discarded, and finish reasons and usage are attributed to candidate 0 alone. Reject the option up front in _prepare_config with ChatClientInvalidRequestException, mirroring the Ollama client's explicit rejection of unsupported tool_choice modes. candidate_count=1 and absent flow through unchanged.

RegressionCheck

test_candidate_count_above_one_is_rejected: {"candidate_count": 2} raises before any request is built; {"candidate_count": 1} still forwards (config.candidate_count == 1); absent stays None.
Full tests/ for the gemini package on this branch: failure set identical to main (pre-existing environment failures only; zero new).

🤖 Generated with ZCode
Co-Authored-By: ZCode noreply@zcode.dev

… silently dropping candidates Both the streaming and non-streaming parsers read candidates[0] only. candidate_count flows to the API through the generic option passthrough, so a multi-candidate request is billed for every candidate while all but the first are silently discarded, and finish reasons and usage are taken from candidate 0 alone. Reject the option up front, mirroring the Ollama client's explicit rejection of unsupported tool_choice modes. 🤖 Generated with ZCode Co-Authored-By: ZCode <noreply@zcode.dev>

Wilson_Jiang (Wilson0511) · 2026-11-01T09:13:08Z

The initial push of this branch was mis-based (it accidentally contained an unrelated worktree snapshot) and was force-pushed to a single clean commit (c284568, gemini files only). The Copilot review above was submitted against that stale commit — its summary ("shell API change", "Redis migration test") references files that are not part of this PR. The current diff is exactly two gemini files (+32 lines).

Copilot

🔵 Needs a closer look

The shell API change breaks existing callers, and the Redis migration test leaves a test-owned key behind during cleanup.

No open issues 🎉

What changed in this PR

This PR addresses #9928 by rejecting Gemini requests for multiple candidates rather than silently discarding responses. It also includes broader changes to Python Agent Framework context management, resource limits, and DevUI file handling.

Changes:

Reject unsupported Gemini candidate counts and preserve both MCP approval lists in OpenAI and Foundry clients.
Add experimental agent-triggered compaction, skill unloading, and identity-preserving Redis scope filters.
Bound shell output and Monty attachment capture, and restrict DevUI file URLs, with supporting tests and documentation.

File	Description
`python/packages/tools/tests/test_shell_truncate_and_quote.py`	Test bounded streaming capture and decoding.
`python/packages/tools/tests/test_shell_parse_rc.py`	Test incremental completion-marker parsing.
`python/packages/tools/tests/test_local_shell_tool.py`	Cover output limits, cancellation, and pipe cleanup.
`python/packages/tools/tests/test_docker_shell_tool.py`	Cover bounded Docker output and termination cleanup.
`python/packages/tools/README.md`	Document independent output limits.
`python/packages/tools/agent_framework_tools/shell/_types.py`	Update truncation field documentation.
`python/packages/tools/agent_framework_tools/shell/_truncate.py`	Add bounded byte and text buffers.
`python/packages/tools/agent_framework_tools/shell/_tool.py`	Expose separate stdout and stderr limits.
`python/packages/tools/agent_framework_tools/shell/_session.py`	Bound persistent capture and parse markers incrementally.
`python/packages/tools/agent_framework_tools/shell/_executor.py`	Use concurrent bounded stateless capture.
`python/packages/tools/agent_framework_tools/shell/_docker.py`	Apply per-stream limits and revise cleanup.
`python/packages/tools/agent_framework_tools/shell/_capture.py`	Add shared subprocess capture and pipe cleanup.
`python/packages/redis/tests/test_providers.py`	Update scope-schema expectations and mocks.
`python/packages/redis/tests/test_context_provider_sdk.py`	Verify scope tokens and client ownership.
`python/packages/redis/tests/test_context_provider_integration.py`	Add live isolation and migration tests.
`python/packages/redis/tests/test_context_provider_edges.py`	Cover scope encoding and schema validation.
`python/packages/redis/README.md`	Document incompatible-schema migration.
`python/packages/redis/agent_framework_redis/_context_provider.py`	Add hashed scope filters and compatibility checks.
`python/packages/openai/tests/openai/test_openai_chat_client.py`	Test continuation rejection and combined approval lists.
`python/packages/openai/agent_framework_openai/_chat_client.py`	Declare continuation aliases and merge approval lists.
`python/packages/monty/tests/monty/test_monty_codeact.py`	Test attachment budgets and bounded reads.
`python/packages/monty/README.md`	Document per-execution capture limits.
`python/packages/monty/agent_framework_monty/_instructions.py`	Describe capture limits in agent instructions.
`python/packages/monty/agent_framework_monty/_execute_code_tool.py`	Enforce attachment count and byte budgets.
`python/packages/gemini/tests/test_gemini_client.py`	Test candidate-count rejection and defaults.
`python/packages/gemini/agent_framework_gemini/_chat_client.py`	Reject requests for multiple candidates.
`python/packages/foundry/tests/foundry/test_foundry_chat_client.py`	Verify both MCP approval lists survive.
`python/packages/foundry/agent_framework_foundry/_chat_client.py`	Merge MCP approval lists without overwriting.
`python/packages/devui/frontend/yarn.lock`	Lock added browser-test dependencies.
`python/packages/devui/frontend/test/file-urls.test.mjs`	Test URL restrictions and embedded-file lifecycle.
`python/packages/devui/frontend/src/lib/file-urls.ts`	Validate file URLs and embedded formats.
`python/packages/devui/frontend/src/components/features/agent/message-renderers/OpenAIContentRenderer.tsx`	Use validated links and managed blob URLs.
`python/packages/devui/frontend/package.json`	Add file-URL tests and Happy DOM.
`python/packages/devui/frontend/package-lock.json`	Lock updated frontend dependencies.
`python/packages/core/tests/core/test_function_invocation_logic.py`	Verify rejection-result metadata.
`python/packages/core/tests/core/test_compaction.py`	Cover trigger semantics and skill unloading.
`python/packages/core/tests/core/test_clients.py`	Test client-managed-history requirements.
`python/packages/core/tests/core/test_agents.py`	Test compaction within agent execution.
`python/packages/core/AGENTS.md`	Document compaction architecture and constraints.
`python/packages/core/agent_framework/_types.py`	Define rejection and skill-provenance keys.
`python/packages/core/agent_framework/_tools.py`	Mark rejected approval results.
`python/packages/core/agent_framework/_skills.py`	Record resolved skill identity on results.
`python/packages/core/agent_framework/_feature_stage.py`	Register the experimental compaction feature.
`python/packages/core/agent_framework/_compaction.py`	Add trigger-driven compaction and skill unloading.
`python/packages/core/agent_framework/_clients.py`	Reject incompatible service-managed history options.
`python/packages/core/agent_framework/__init__.pyi`	Export new strategy types.
`python/packages/core/agent_framework/__init__.py`	Expose new strategies through lazy exports.

_{Review effort: Balanced}

Eduard van Valkenburg (eavanvalkenburg) · 2026-11-02T13:29:52Z

+                # unsupported-tool_choice rejection).
+                raise ChatClientInvalidRequestException(
+                    "Gemini chat client supports a single candidate. "
+                    f"candidate_count={value} would silently drop every candidate but the first; "


the default should be that folks remove it altogether, but if passed, it needs to be 1

Eduard van Valkenburg (eavanvalkenburg) · 2026-11-02T13:30:12Z

            if key == "stop" and isinstance(value, str):
                # ChatOptions allows a single stop string; stop_sequences takes a list.
                value = [value]
+            if key == "candidate_count" and value > 1:


0 as value would also be wrong, so let's make it:

Suggested change

if key == "candidate_count" and value > 1:

if key == "candidate_count" and value != "1":

Copilot AI balanced review requested due to automatic review settings November 1, 2026 09:06

Wilson_Jiang (Wilson0511) requested review from SergeyMenshykh and Eduard van Valkenburg (eavanvalkenburg) as code owners November 1, 2026 09:06

Wilson_Jiang (Wilson0511) temporarily deployed to github-app-auth November 1, 2026 09:06 — with GitHub Actions Inactive

Wilson_Jiang (Wilson0511) requested review from Darren Cohen (dargilco), Jose Alvarez (jpalvarezl), Roger Barreto (rogerbarreto), sophia-ramsey and westey (westey-m) as code owners November 1, 2026 09:06

Wilson_Jiang (Wilson0511) temporarily deployed to github-app-auth November 1, 2026 09:06 — with GitHub Actions Inactive

agent-framework-automation Bot added documentation Usage: [Issues, PRs], Target: documentation in the code base and learn docs python Usage: [Issues, PRs], Target: Python labels Nov 1, 2026

Copilot started reviewing on behalf of Wilson_Jiang (Wilson0511) November 1, 2026 09:06

View session

Wilson_Jiang (Wilson0511) closed this Nov 1, 2026

Wilson_Jiang (Wilson0511) force-pushed the fix/gemini-multi-candidate-silent-drop branch from db38dfa to b7884ad Compare November 1, 2026 09:08

Wilson_Jiang (Wilson0511) had a problem deploying to github-app-auth November 1, 2026 09:08 — with GitHub Actions Failure

Wilson_Jiang (Wilson0511) reopened this Nov 1, 2026

Wilson_Jiang (Wilson0511) temporarily deployed to github-app-auth November 1, 2026 09:11 — with GitHub Actions Inactive

Copilot AI reviewed Nov 1, 2026

View reviewed changes

Eduard van Valkenburg (eavanvalkenburg) mentioned this pull request Nov 2, 2026

Python: Gemini chat client silently drops all but the first candidate when candidate_count > 1 #9928

Open

Eduard van Valkenburg (eavanvalkenburg) reviewed Nov 2, 2026

View reviewed changes

Navigation Menu

Repository navigation

Python: Reject candidate_count>1 in the Gemini chat client instead of silently dropping candidates - #9929

Python: Reject candidate_count>1 in the Gemini chat client instead of silently dropping candidates#9929
Wilson_Jiang (Wilson0511) wants to merge 1 commit into
microsoft:mainfrom
Wilson0511:fix/gemini-multi-candidate-silent-drop

Wilson_Jiang (Wilson0511) commented Nov 1, 2026

Uh oh!

Wilson_Jiang (Wilson0511) commented Nov 1, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Eduard van Valkenburg (eavanvalkenburg) Nov 2, 2026

Uh oh!

Eduard van Valkenburg (eavanvalkenburg) Nov 2, 2026

Uh oh!

This branch was previously deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	if key == "candidate_count" and value > 1:
	if key == "candidate_count" and value != "1":

Conversation

Wilson_Jiang (Wilson0511) commented Nov 1, 2026

Description

RegressionCheck

Uh oh!

Wilson_Jiang (Wilson0511) commented Nov 1, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🔵 Needs a closer look

Uh oh!

Eduard van Valkenburg (eavanvalkenburg) Nov 2, 2026

Choose a reason for hiding this comment

Uh oh!

Eduard van Valkenburg (eavanvalkenburg) Nov 2, 2026

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