[BUG] Antigravity Interactions API continuation becomes corrupted after several completed turns

,

Hi,

I have encountered a reproducible issue with stateful conversation continuation in the Gemini Interactions API using the Antigravity managed agent.

Agent: antigravity-preview-05-2026
API revision: 2026-05-20
Service tier: standard
Endpoint: /v1beta/interactions

I maintain a persistent conversation by storing the returned previous_interaction_id and environment_id after each successful interaction.

The conversation initially completed five consecutive interactions successfully. All of those interactions are still retrievable through GET /v1beta/interactions/{id} and report status: completed.

After the fifth turn, however, the continuation chain became unusable.

What I tested

1. Fresh Antigravity interaction, no previous_interaction_id

Result:

HTTP 200

The agent responds normally and a new remote environment is created.

This confirms that the API key, account, Antigravity agent and environment provisioning are working.


2. Continue from interaction #5 using a fresh environment

Request uses:

previous_interaction_id = interaction #5
environment = remote

Result consistently:

HTTP 400

{"error":{"message":"Precondition check failed.","code":"invalid_request"}}

However, interaction #5 itself is retrievable and reports:

status: completed


3. Continue from interaction #4 using a fresh environment

Result consistently:

HTTP 503

The service is currently unavailable.

Repeated retries produce the same result.


4. Continue from interaction #3 using a fresh environment

This produces a particularly unexpected result.

The new request contains only:

Diagnostic only. Reply exactly: CONTEXT_OK

The API returns:

HTTP 200

but instead of responding CONTEXT_OK, it effectively replays the original historical interaction #4.

The response content corresponds to the previous #4 turn and, importantly, the usage statistics are identical to the original interaction #4:

input_tokens:   118585
cached_tokens:  106496
output_tokens:  1455
thought_tokens: 2040
total_tokens:   122080

A new interaction ID and a new environment ID are returned, but the model appears to process/replay the historical next turn rather than the newly supplied input.


Environment verification

I also checked the environment used by the original conversation through:

GET /v1beta/environments/{environment_id}

It returns:

HTTP 200
status: active

Interactions #4 and #5 both reference that same environment and both report status: completed.

I additionally tested continuation using environment: remote to eliminate the original sandbox as the cause. The problem remains.


Current behaviour summary

Fresh interaction              -> HTTP 200, works normally

Continue from interaction #3   -> HTTP 200,
                                  but replays historical #4
                                  instead of processing new input

Continue from interaction #4   -> repeated HTTP 503

Continue from interaction #5   -> repeated HTTP 400
                                  "Precondition check failed"

This appears to be a server-side conversation/continuation state issue rather than an API key, quota or environment problem.

The same API key can create fresh Antigravity interactions successfully at the same time.

Expected behaviour

A completed interaction should remain usable as previous_interaction_id, and the new input should become the next conversation turn.

Using environment: remote should allow the conversation to continue with a new sandbox while preserving the conversation history.

Additional information available

I have retained:

  • exact UTC timestamps,

  • interaction IDs,

  • environment IDs,

  • complete retained interaction responses,

  • per-turn token telemetry,

  • API responses for the 400 / 503 cases,

  • and a minimal reproduction script.

For privacy reasons I am not posting the full interaction IDs publicly, because they belong to a longer persistent conversation.

I can provide the full IDs and raw responses privately to Google engineering if needed.

Is this a known limitation or bug in stateful continuation for antigravity-preview-05-2026?

Thanks.

Thanks for contributing @daareek ! The team is looking into this, hope to have update soon

Hi there, is it possible for you to share your POST /interactions payload, feel free to send it over to alicevik@google.com if you don’t want to share it here.

We think the error is caused by you sending models/antigravity-preview… over the API rather than just agent: “antigravity-preview-05-2026”

You might want to try by dropping models/ prefix and see if it works.