Plumb EnableMessageInjection through provider AgentConfig so the message-injector loop is reachable

[PratikDhanave (PratikDhanave)]

What

The toolautocall harness already fully implements Config.EnableMessageInjection and its MessageInjector loop (inject.go, wired in autocall.go, tested in autocall_inject_test.go, ported from .NET's MessageInjectingChatClient). However, no provider ever wired the flag: every provider built toolautocall.Config with only Logger/LogSensitiveData, so the message-injector loop was unreachable through the normal agent API.

This PR:

• Adds EnableMessageInjection bool to agent.Config (in agent/agent.go), beside the existing DisableFuncAutoCall.
• Passes EnableMessageInjection: config.EnableMessageInjection into each provider's guarded toolautocall.New(toolautocall.Config{...}) literal: openaiprovider (chat + responses), anthropicprovider, geminiprovider, and aguiprovider.

Because each provider AgentConfig embeds agent.Config and forwards config.Config to agent.New, the new field flows to all providers uniformly. Previously the only way to enable injection was DisableFuncAutoCall=true plus hand-rebuilding the middleware, which loses provider defaults.

Why (parity)

This mirrors the .NET MessageInjectingChatClient / EnqueueMessages surface, which is composable into the standard function-invocation pipeline. Exposing the flag on AgentConfig keeps the Go SDK aligned: tools can enqueue follow-up messages into the auto-call loop without opting out of the provider-managed middleware.

Behavior change

None when the field is false (the default). The value is zero-valued for all existing callers, so the installed middleware behaves exactly as before.

Tests

Added two black-box tests in the canonical provider/openaiprovider/chat_test.go:

• TestChatMessageInjection_ToolInjectsMessage: builds an agent with AgentConfig.EnableMessageInjection=true and a tool that calls MessageInjectorFromContext(ctx).EnqueueMessages(...); asserts the injected message is forwarded to the provider on the following service round and the final answer is produced. This test fails before the wiring change and passes after.
• TestChatMessageInjection_DisabledReturnsNilInjector: negative case with the field false, asserting MessageInjectorFromContext returns nil and nothing is injected.

go build ./..., go vet, and go test (incl. -race on the new tests) pass for the changed packages.

Open design questions

• Scope: This threads the existing harness flag through all five providers for consistency. Should the responses/agui paths be excluded if they are considered out of scope for the initial cut?
• API shape: EnableMessageInjection sits on agent.Config alongside DisableFuncAutoCall. Is that the desired home, or should message injection be surfaced as part of a broader composable-middleware config (cf. Add composable function-invocation middleware to the tool autocall loop #638)?
• Follow-ups: Only the injector enable flag is plumbed; other toolautocall.Config knobs (e.g. AllowConcurrentInvocations, MaximumIterationsPerRequest) remain provider-internal. Happy to plumb those in a follow-up if that is the direction.

[Copilot]

Pull request overview

This PR makes the existing toolautocall message-injection loop reachable through the normal provider agent constructors by plumbing an EnableMessageInjection flag through agent.Config and into each provider’s toolautocall.Config.

Changes:

• Added EnableMessageInjection bool to agent.Config so providers can expose message injection via their embedded agent.Config.
• Threaded EnableMessageInjection into the provider-installed toolautocall.New(toolautocall.Config{...}) middleware across OpenAI (chat + responses), Anthropic, Gemini, and AGUI providers.
• Added OpenAI chat black-box tests validating injection-enabled behavior and the default disabled behavior.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
agent/agent.go	Adds Config.EnableMessageInjection (public surface) and documents intended behavior.
provider/openaiprovider/chat.go	Passes EnableMessageInjection into the OpenAI chat agent’s toolautocall middleware config.
provider/openaiprovider/responses.go	Passes EnableMessageInjection into the OpenAI responses agent’s toolautocall middleware config.
provider/openaiprovider/chat_test.go	Adds end-to-end tests proving injected messages are forwarded on the next provider call (and nil injector when disabled).
provider/anthropicprovider/agent.go	Threads EnableMessageInjection into Anthropic provider toolautocall config.
provider/geminiprovider/agent.go	Threads EnableMessageInjection into Gemini provider toolautocall config.
provider/aguiprovider/agui.go	Threads EnableMessageInjection into AGUI provider toolautocall config.

[github-actions]

API Consistency Review — PR #652: Plumb EnableMessageInjection through provider AgentConfig

Scope

This PR adds one exported field, EnableMessageInjection bool, to agent.Config and wires it through five provider constructors so callers can enable the already-implemented toolautocall.MessageInjector loop without opting out of provider-managed middleware.

Cross-repo parity mapping

SDK	Surface	Default	Notes
.NET	ChatClientAgentOptions.EnableMessageInjection	false (opt-in)	Adds MessageInjectingChatClient to the pipeline when true
Python	MessageInjectionMiddleware	always on — hardcoded in _harness/_agent.py, no opt-out	"It is a no-op when no messages are queued for the session, so there is no opt-out."
Go (this PR)	agent.Config.EnableMessageInjection	false (opt-in)	Matches .NET semantics

Assessment

The Go PR's opt-in default (false) is aligned with the .NET SDK and is a documented, intentional divergence from Python, which always enables the middleware (declared a zero-cost no-op when no messages are queued). The field name EnableMessageInjection matches the .NET option name exactly.

No parity issues are found. The exported API addition (agent.Config.EnableMessageInjection) correctly exposes an existing internal capability through the canonical config surface, preserving semantic parity with the .NET opt-in model.

The public-api-change label is already present and appropriate — this PR adds an exported field to agent.Config.

✅ Parity approved.

Generated by Go API Consistency Review Agent · sonnet46 · 35.1 AIC · ⌖ 5.06 AIC · ⊞ 6K · ◷