Foundry memory items emit assistant content as input_text instead of output_text

[PratikDhanave (PratikDhanave)]

What

toResponseItem in provider/foundryprovider/memory.go selected the message
role among assistant/system/user but always emitted the content part with type
input_text. As a result, an assistant memory item was serialized with
input_text rather than output_text.

This selects the content-part type by role: assistant items now emit
output_text, while user and system items keep input_text.

Why

This aligns with the repo's Responses convention and .NET/Python parity. In
provider/openaiprovider/responses.go, assistant turns are output messages
built with OfOutputText (output_text), while user/system turns use
OfInputText (input_text). Foundry memory items should follow the same
input/output distinction so the assistant turn is represented as model output,
not user input.

Tests

Extended TestMemoryProviderInvokedUpdatesMemories to decode the serialized
items and assert the content-part type: the user item is input_text and the
assistant item is output_text. The test fails before the fix (assistant part
was input_text) and passes after.

[Copilot]

Pull request overview

This PR fixes Foundry memory-item serialization so that assistant messages are represented as model output (output_text) rather than user input (input_text), aligning Foundry’s behavior with the repo’s Responses conventions and other SDK parity.

Changes:

• Update toResponseItem to emit output_text when the role is assistant, and input_text otherwise.
• Extend TestMemoryProviderInvokedUpdatesMemories to assert the serialized content-part type for both user and assistant memory items.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
provider/foundryprovider/memory.go	Selects content-part "type" based on message role, emitting output_text for assistant items.
provider/foundryprovider/memory_test.go	Adds assertions verifying user items serialize as input_text and assistant items as output_text.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

Parity Review — ✅ Approved

This PR fixes a bug in the unexported toResponseItem helper in provider/foundryprovider/memory.go. No exported (public) Go APIs are added, removed, or changed.

Findings

Scope: Internal implementation fix — toResponseItem is an unexported function; no exported identifiers or user-visible API contracts are affected.

Parity: The fix explicitly improves alignment with upstream behavior. In both the Go provider/openaiprovider/responses.go and the upstream .NET/Python implementations, assistant turns are serialized as output_text while user/system turns use input_text. The previous Go behavior (always emitting input_text) was a divergence from upstream semantics; this PR corrects it.

Labels: The public-api-change label is not warranted — only unexported implementation details changed. The existing parity-approved label is appropriate and will be retained.

No parity issues identified.

Warning

Firewall blocked 1 domain

The following domain was blocked by the firewall during workflow execution:

• awmgmcpg

To allow these domains, add them to the network.allowed list in your workflow frontmatter:

network:
  allowed:
    - defaults
    - "awmgmcpg"

See Network Configuration for more information.

Generated by Go API Consistency Review Agent · 21.6 AIC · ⌖ 5.55 AIC · ⊞ 5.9K · ◷