Gemini Deep Research Interactions stuck in in_progress after May 26 schema migration

Gemini API
1

Subject: Gemini Deep Research Interactions stuck in in_progress after May 26 schema migration

Hello Google Gemini API team,

We are seeing a regression with Gemini Deep Research via the Interactions API after the May 26, 2026 migration to the new steps schema.

We run a production Telegram bot that takes a Polymarket event, builds a prompt, starts Gemini Deep Research, then delivers the final report as PDF to the user.

Since May 26, some Deep Research interactions never complete. They remain in:

status: in_progress
output_text: empty
error: None
steps: only one user_input step

They stay like this for hours. There is no completed, failed, requires_action, cancelled, or expired terminal status.

Environment:

  • Python 3.12.3 (production), Python 3.10 (test instance)

  • google-genai==2.7.0 (production), google-genai==1.75.0 (test instance)

  • httpx==0.28.1

  • Model: deep-research-pro-preview-12-2025 (production), also tested deep-research-preview-04-2026 and deep-research-max-preview-04-2026

  • The issue reproduces on both SDK versions

API method in production bot:

client.interactions.create(
    input=full_prompt,
    agent="deep-research-pro-preview-12-2025",
    background=True,
    store=True,
)

Then we poll with:

client.interactions.get(interaction_id)

What we observe:

  • Before May 26, we successfully ran 300+ Deep Research tasks across different prompts and Polymarket events.

  • After May 26, some tasks get stuck indefinitely.

  • For stuck tasks, interactions.get(id) repeatedly returns in_progress, with no output and no error.

  • The only step is user_input; there are no model/tool/search/url steps, so it looks like the agent worker never starts after the interaction is accepted.

  • Network tests from the same server are stable. Raw HTTP requests to generativelanguage.googleapis.com work consistently.

  • The issue is not the SDK client lifecycle bug: our real code keeps the genai.Client instance alive.

  • Deep Research works normally through the Gemini web interface (AI Studio) during the same time period when API calls are stuck.

  • A separate CLI implementation using the same server, same model, same universal prompt, and same events has completed 8/10 runs successfully when using:

client.interactions.create(
    input=prompt_text,
    agent="deep-research-pro-preview-12-2025",
    background=True,
    stream=True,
)

This suggests the background=True, store=True polling path may be using a different execution path or queue than the streaming path.

  • One interaction completed successfully (2.3 min) using a brand-new API key from a fresh Google Cloud project with store=True. Subsequent requests with the same key and identical parameters returned to the stuck behavior. This suggests possible per-key or per-project rate limiting or queue starvation, rather than a pure code-path bug.

We also tested / considered:

  • The issue does not look like prompt parsing failure, because the stuck interaction never proceeds past the initial user_input step.

  • It does not look like collaborative planning waiting for approval, because there is no model_output plan step and no requires_action status.

  • The new steps schema parsing works for completed interactions.

  • The problem reproduces across all three available Deep Research models (deep-research-pro-preview-12-2025, deep-research-preview-04-2026, deep-research-max-preview-04-2026).

  • The problem reproduces with trivial prompts (e.g., “What is 2+2?”), not just large production prompts.

  • The problem is that some interactions never reach any terminal state.

Stuck interaction IDs:

# Production bot (deep-research-pro-preview-12-2025, without store=True):
v1_ChdURWtaYW83S0Z0aUY0LUVQNnFTQ21BWRIXVEVrWmFvN0tGdGlGNC1FUDZxU0NtQVk
  stuck 31min, killed by bot restart

v1_ChdjMUVaYXJpTUVyX2hnOFVQc2ZYMWdBTRIXYzFFWmFyaU1Fcl9oZzhVUHNmWDFnQU0
  stuck 8min, killed by VPS reboot

# Test bot (deep-research-pro-preview-12-2025, with store=True):
v1_ChdRVmtaYXR6N1BLX0FnOFVQMHFxSTRBMBIXUVZrWmF0ejdQS19BZzhVUDBxcUk0QTA
  stuck 16+min

# Test script (deep-research-max-preview-04-2026, store=True, trivial "2+2" prompt):
v1_ChdUWGNaYXJiZk04X2JqdU1QNzZTRHNBZxIXVFhjWmFyYmZNOF9ianVNUDc2U0RzQWc
  stuck 10+min

# Test script (deep-research-preview-04-2026, store=True) — SUCCEEDED:
v1_Chc5SDhaYXFMMEx2bkJqdU1QOVlTd3lRaxIXOUg4WmFxTDBMdm5CanVNUDlZU3d5UWs
  completed in 2.3min ✓

# Test bot (deep-research-preview-04-2026, store=True, new API key from new project):
v1_ChdDb1VaYXVQSEs2bmJqdU1QZ2J1QmdBZxIXQ29VWmF1UEhLNm5ianVNUGdidUJnQWc
  stuck 13+min

# Test script (deep-research-preview-04-2026, store=True, longer prompt):
v1_ChdhWTRaYXByZ09kNklxZmtQeDVfdjBRNBIXYVk0WmFwcmdPZDZJcWZrUHg1X3YwUTQ
  stuck 5.5+min

Questions:

  1. Is there a known issue with Deep Research Interactions created with background=True, store=True after the May 26 migration?

  2. Are background=True, store=True and background=True, stream=True routed through different backend execution paths or queues?

  3. Should production integrations use stream=True first and then resume/poll with interactions.get(), instead of pure polling?

  4. What is the expected terminal behavior for a Deep Research interaction that fails to start? Should it eventually become failed or expired?

  5. Can you confirm the current default for agent_config.collaborative_planning for Deep Research?

  6. Should we explicitly set:

agent_config={
    "type": "deep-research",
    "thinking_summaries": "auto",
    "collaborative_planning": False,
}
  1. Is service_tier relevant for Deep Research reliability? Could flex cause accepted interactions to remain in in_progress indefinitely?

Expected behavior:

  • If the task is accepted, it should either start producing agent steps and eventually complete, or fail/expire with a terminal status.

  • It should not stay in in_progress indefinitely with only user_input.

This regression is affecting paid user-facing workflows, so any guidance from the Gemini API / Deep Research team would be very helpful.

Thank you.

Is there sensitive data here? Or can I post it directly on the forum?

Interaction IDs are public task identifiers; they do not contain API keys or data. Everything else — SDK versions, call code, problem description — is also non-sensitive.

You can post it on the forum as is. The only thing is — do not include your API key in the message, but it’s not there anyway.

Subject: Gemini Deep Research Interactions stuck in in_progress after May 26 schema migration

Hello Google Gemini API team,

We are seeing a regression with Gemini Deep Research via the Interactions API after the May 26, 2026 migration to the new steps schema.

We run a production Telegram bot that takes a Polymarket event, builds a prompt, starts Gemini Deep Research, then delivers the final report as PDF to the user.

Since May 26, some Deep Research interactions never complete. They remain in:

`status: in_progress
output_text: empty
error: None
steps: only one user_input step`

They stay like this for hours. There is no completed, failed, requires_action, cancelled, or expired terminal status.

**Environment:**

  • Python 3.12.3 (production), Python 3.10 (test instance)
  • google-genai==2.7.0 (production), google-genai==1.75.0 (test instance)
  • httpx==0.28.1
  • Model: deep-research-pro-preview-12-2025 (production), also tested deep-research-preview-04-2026 and deep-research-max-preview-04-2026
  • The issue reproduces on both SDK versions

**API method in production bot:**

`client.interactions.create(
input=full_prompt,
agent=“deep-research-pro-preview-12-2025”,
background=True,
store=True,
)`

Then we poll with:

`client.interactions.get(interaction_id)`

**What we observe:**

  • Before May 26, we successfully ran 300+ Deep Research tasks across different prompts and Polymarket events.
  • After May 26, some tasks get stuck indefinitely.
  • For stuck tasks, `interactions.get(id)` repeatedly returns `in_progress`, with no output and no error.
  • The only step is `user_input`; there are no model/tool/search/url steps, so it looks like the agent worker never starts after the interaction is accepted.
  • Network tests from the same server are stable. Raw HTTP requests to `generativelanguage.googleapis.com` work consistently.
  • The issue is not the SDK client lifecycle bug: our real code keeps the `genai.Client` instance alive.
  • Deep Research works normally through the Gemini web interface (AI Studio) during the same time period when API calls are stuck.
  • A separate CLI implementation using the same server, same model, same universal prompt, and same events has completed 8/10 runs successfully when using:

`client.interactions.create(
input=prompt_text,
agent=“deep-research-pro-preview-12-2025”,
background=True,
stream=True,
)`

This suggests the `background=True, store=True` polling path may be using a different execution path or queue than the streaming path.

  • One interaction completed successfully (2.3 min) using a brand-new API key from a fresh Google Cloud project with `store=True`. Subsequent requests with the same key and identical parameters returned to the stuck behavior. This suggests possible per-key or per-project rate limiting or queue starvation, rather than a pure code-path bug.

**We also tested / considered:**

  • The issue does not look like prompt parsing failure, because the stuck interaction never proceeds past the initial `user_input` step.
  • It does not look like collaborative planning waiting for approval, because there is no `model_output` plan step and no `requires_action` status.
  • The new steps schema parsing works for completed interactions.
  • The problem reproduces across all three available Deep Research models (`deep-research-pro-preview-12-2025`, `deep-research-preview-04-2026`, `deep-research-max-preview-04-2026`).
  • The problem reproduces with trivial prompts (e.g., “What is 2+2?”), not just large production prompts.
  • The problem is that some interactions never reach any terminal state.

**Stuck interaction IDs:**

`# Production bot (deep-research-pro-preview-12-2025, without store=True):
v1_ChdURWtaYW83S0Z0aUY0LUVQNnFTQ21BWRIXVEVrWmFvN0tGdGlGNC1FUDZxU0NtQVk
stuck 31min, killed by bot restart

Test bot (deep-research-pro-preview-12-2025, with store=True):

v1_ChdRVmtaYXR6N1BLX0FnOFVQMHFxSTRBMBIXUVZrWmF0ejdQS19BZzhVUDBxcUk0QTA
stuck 16+min

Test script (deep-research-max-preview-04-2026, store=True, trivial “2+2” prompt):

v1_ChdUWGNaYXJiZk04X2JqdU1QNzZTRHNBZxIXVFhjWmFyYmZNOF9ianVNUDc2U0RzQWc
stuck 10+min

Test script (deep-research-preview-04-2026, store=True) — SUCCEEDED:

v1_Chc5SDhaYXFMMEx2bkJqdU1QOVlTd3lRaxIXOUg4WmFxTDBMdm5CanVNUDlZU3d5UWs
completed in 2.3min ✓

Test bot (deep-research-preview-04-2026, store=True, new API key from new project):

v1_ChdDb1VaYXVQSEs2bmJqdU1QZ2J1QmdBZxIXQ29VWmF1UEhLNm5ianVNUGdidUJnQWc
stuck 13+min`

**Questions:**

  1. Is there a known issue with Deep Research Interactions created with `background=True, store=True` after the May 26 migration?
  2. Are `background=True, store=True` and `background=True, stream=True` routed through different backend execution paths or queues?
  3. Should production integrations use `stream=True` first and then resume/poll with `interactions.get()`, instead of pure polling?
  4. What is the expected terminal behavior for a Deep Research interaction that fails to start? Should it eventually become `failed` or `expired`?
  5. Can you confirm the current default for `agent_config.collaborative_planning` for Deep Research?
  6. Should we explicitly set:

`agent_config={
“type”: “deep-research”,
“thinking_summaries”: “auto”,
“collaborative_planning”: False,
}`

  1. Is `service_tier` relevant for Deep Research reliability? Could `flex` cause accepted interactions to remain in `in_progress` indefinitely?

**Expected behavior:**

  • If the task is accepted, it should either start producing agent steps and eventually complete, or fail/expire with a terminal status.
  • It should not stay in `in_progress` indefinitely with only `user_input`.

This regression is affecting paid user-facing workflows, so any guidance from the Gemini API / Deep Research team would be very helpful.

Thank you.

Prepare this data so that I can copy it easily, like you do with code, for example.

The last question was numbered as 1 instead of 7. Here is the corrected version — copy it in its entirety:

Subject: Gemini Deep Research Interactions stuck in in_progress after May 26 schema migration

Hello Google Gemini API team,

We are seeing a regression with Gemini Deep Research via the Interactions API after the May 26, 2026 migration to the new steps schema.

We run a production Telegram bot that takes a Polymarket event, builds a prompt, starts Gemini Deep Research, then delivers the final report as PDF to the user.

Since May 26, some Deep Research interactions never complete. They remain in:

status: in_progress
output_text: empty
error: None
steps: only one user_input step


They stay like this for hours. There is no completed, failed, requires_action, cancelled, or expired terminal status.

**Environment:**

- Python 3.12.3 (production), Python 3.10 (test instance)
- google-genai==2.7.0 (production), google-genai==1.75.0 (test instance)
- httpx==0.28.1
- Model: deep-research-pro-preview-12-2025 (production), also tested deep-research-preview-04-2026 and deep-research-max-preview-04-2026
- The issue reproduces on both SDK versions

**API method in production bot:**

```python
client.interactions.create(
    input=full_prompt,
    agent="deep-research-pro-preview-12-2025",
    background=True,
    store=True,
)

Then we poll with:

client.interactions.get(interaction_id)

What we observe:

  • Before May 26, we successfully ran 300+ Deep Research tasks across different prompts and Polymarket events.

  • After May 26, some tasks get stuck indefinitely.

  • For stuck tasks, interactions.get(id) repeatedly returns in_progress, with no output and no error.

  • The only step is user_input; there are no model/tool/search/url steps, so it looks like the agent worker never starts after the interaction is accepted.

  • Network tests from the same server are stable. Raw HTTP requests to generativelanguage.googleapis.com work consistently.

  • The issue is not the SDK client lifecycle bug: our real code keeps the genai.Client instance alive.

  • Deep Research works normally through the Gemini web interface (AI Studio) during the same time period when API calls are stuck.

  • A separate CLI implementation using the same server, same model, same universal prompt, and same events has completed 8/10 runs successfully when using:

client.interactions.create(
    input=prompt_text,
    agent="deep-research-pro-preview-12-2025",
    background=True,
    stream=True,
)

This suggests the background=True, store=True polling path may be using a different execution path or queue than the streaming path.

  • One interaction completed successfully (2.3 min) using a brand-new API key from a fresh Google Cloud project with store=True. Subsequent requests with the same key and identical parameters returned to the stuck behavior. This suggests possible per-key or per-project rate limiting or queue starvation, rather than a pure code-path bug.

We also tested / considered:

  • The issue does not look like prompt parsing failure, because the stuck interaction never proceeds past the initial user_input step.

  • It does not look like collaborative planning waiting for approval, because there is no model_output plan step and no requires_action status.

  • The new steps schema parsing works for completed interactions.

  • The problem reproduces across all three available Deep Research models (deep-research-pro-preview-12-2025, deep-research-preview-04-2026, deep-research-max-preview-04-2026).

  • The problem reproduces with trivial prompts (e.g., “What is 2+2?”), not just large production prompts.

  • The problem is that some interactions never reach any terminal state.

Stuck interaction IDs:

# Production bot (deep-research-pro-preview-12-2025, without store=True):
v1_ChdURWtaYW83S0Z0aUY0LUVQNnFTQ21BWRIXVEVrWmFvN0tGdGlGNC1FUDZxU0Nt...
  stuck 31min, killed by bot restart

# Test bot (deep-research-pro-preview-12-2025, with store=True):
v1_ChdRVmtaYXR6N1BLX0FnOFVQMHFxSTRBMBIXUVZrWmF0ejdQS19BZzhVUDBxcUk0...
  stuck 16+min

# Test script (deep-research-max-preview-04-2026, store=True, trivial "2+2" prompt):
v1_ChdUWGNaYXJiZk04X2JqdU1QNzZTRHNBZxIXVFhjWmFyYmZNOF9ianVNUDc2U0Rz...
  stuck 10+min

# Test script (deep-research-preview-04-2026, store=True) — SUCCEEDED:

v1_Chc5SDhaYXFMMEx2bkJqdU1QOVlTd3lRaxIXOUg4WmFxTDBMdm5CanVNUDlZU3d5...
  completed in 2.3min ✓

# Test bot (deep-research-preview-04-2026, store=True, new API key from new project):
v1_ChdDb1VaYXVQSEs2bmJqdU1QZ2J1QmdBZxIXQ29VWmF1UEhLNm5ianVNUGdidUJn...
  stuck 13+min

Questions:

  1. Is there a known issue with Deep Research Interactions created with background=True, store=True after the May 26 migration?

  2. Are background=True, store=True and background=True, stream=True routed through different backend execution paths or queues?

  3. Should production integrations use stream=True first and then resume/poll with interactions.get(), instead of pure polling?

  4. What is the expected terminal behavior for a Deep Research interaction that fails to start? Should it eventually become failed or expired?

  5. Can you confirm the current default for agent_config.collaborative_planning for Deep Research?

  6. Should we explicitly set:

agent_config={
    "type": "deep-research",
    "thinking_summaries": "auto",
    "collaborative_planning": False,
}
  1. Is service_tier relevant for Deep Research reliability? Could flex cause accepted interactions to remain in in_progress indefinitely?

Expected behavior:

  • If the task is accepted, it should either start producing agent steps and eventually complete, or fail/expire with a