[BUG] Task can stall after a completed tool result with no next API request #1883

Problem (one or two sentences)

After a tool such as read_file completes, Zoo Code can remain active with the Stop button visible but never start the next API request. Cancelling and resuming can continue the task, but an interrupted batch read leaves the model without the rest of the files and can change the work that follows.

Context (who is affected and when)

This is intermittent during longer agentic tasks with auto-approved native tool calls. It has occurred across multiple builds and providers, most often during file-reading sequences, and is especially damaging when a batch read_file turn stops partway through.

Reproduction steps

1. Run Zoo Code 3.84.0 with read-only tool auto-approval enabled.
2. Start a task that causes several native read_file calls, including a multi-file or batch turn.
3. Continue until a read card completes and is marked answered.
4. In an affected turn, observe that no new API request or rate-limit wait appears. The tool card has no spinner, but the Stop button still indicates an active task.
5. Cancel the task and resume it. The task can continue, but it does not reliably resume the original incomplete batch read.

The failure is intermittent. A deterministic regression test can reproduce the terminal state by completing the tool result while deliberately losing only the readiness-latch update.

Expected result

Once the stream is closed and every completed tool call has a matching result, Zoo starts the next API request and preserves the full batch-read sequence.

Actual result

The task can remain active indefinitely between the completed tool turn and the next API request. It may eventually wake up, or require cancel/resume and leave the model with incomplete batch context.

Variations tried (optional)

The completed-card variant has appeared with multiple providers and model profiles. It is distinct from a partial tool call waiting for stream termination: the tool card is complete and answered, and the real tool result already exists.

App Version

3.84.0 and current main at c38db3ed761318776e8d47630487dac76b37686a

API Provider (optional)

openai, anthropic

Model Used (optional)

Occurs with GPT and Claude models.