CLI Issue - Losing Connection Mid-Prompt

Summary

CommandCode registers a failure.

Expected Behavior

Completed prompt with an implementation report.

Actual Behavior

Error: Connection error.

Type "continue" to try again. If the issue persists, contact support:
https://commandcode.ai/discord
Trace ID: 2ad8ca50de4a82e887c4ac68dbd0215c

Steps to reproduce the issue

1. Run any prompt requiring reasoning

Command Code Version

1.54.0

Operating System

macOS

Terminal/IDE

Terminal

Shell

bash

Session file (optional)

Attatched

Fix prompt (optional)

Please investigate and repair a recurring CommandCode platform/runtime connectivity failure affecting autonomous coding turns.

This issue has now occurred at least THREE TIMES across DIFFERENT, UNRELATED PROMPTS/WORKFLOWS. The repeated failures are not specific to the repository task shown below.

Environment:

• Command Code version: v1.54.0
• Models in the most recent affected session: mimo-v2.5-pro · taste-1
• Host: macOS
• Most recent failure: September 14, 2026

SEVERITY / SCOPE

This should be investigated as a recurring CommandCode infrastructure, runtime, model-routing, or session-management defect.

The same general failure has occurred at least three times while executing unrelated prompts. This substantially reduces the likelihood that the underlying cause is the specific codebase, prompt, repository operation, or task being executed.

Please do NOT focus the investigation on repairing the user’s repository or application code. The autonomous QA/IT investigation should focus on CommandCode itself and its managed dependencies.

MOST RECENT INCIDENT

CommandCode was executing:

Review the prompt, ncaa_repair_sep_14.txt, and execute.

The agent operated normally for approximately 35 seconds and successfully performed numerous READ, LIST, GLOB, SEARCH, and SHELL operations.

The final successful shell operation completed normally and returned:

Games in Sept 8-16 window: 303

Immediately afterward, CommandCode failed with:

⚠ Error: Connection error.

CommandCode instructed:

Type "continue" to try again. If the issue persists, contact support.

Trace ID:

f162a2399c6ab11ab3279a106bb1caaa

The user entered:

continue

Instead of recovering the turn, CommandCode reported:

Cannot run /feedback while Command Code is working. Press Esc or Ctrl+C to interrupt the current turn first.

It then produced another:

⚠ Error: Connection error.

Second Trace ID:

2ad8ca50de4a82e887c4ac68dbd0215c

CommandCode then opened the GitHub feedback workflow and exported:

command-code-session-a836e855.html

RECURRING FAILURE PATTERN

This is not an isolated occurrence.

At least two additional connection failures have occurred during other, unrelated CommandCode prompts.

Those incidents involved different work/tasks, making the common factor CommandCode rather than the specific NCAAF workflow shown in the latest session.

Please correlate available account/client/backend telemetry across recent sessions rather than examining only the two trace IDs above.

INVESTIGATION REQUIRED

Determine whether these recurring failures originate from:

• CommandCode backend/gateway connectivity
• model-provider connectivity
• model routing or failover
• streaming/session transport
• long-lived request handling
• context/session persistence
• tool-result serialization or ingestion
• tool execution → model handoff
• backend worker termination/recycling
• timeout handling
• rate limiting
• authentication/session token handling
• client/backend state synchronization
• network retry logic
• another CommandCode-managed dependency

Please specifically determine whether there is a relationship between the failures and autonomous turns involving multiple sequential tool calls.

RECOVERY PATH DEFECT

There also appears to be a separate or related defect in failure recovery.

CommandCode explicitly instructed the user to type:

continue

After doing exactly that, the client reported:

Cannot run /feedback while Command Code is working.

This suggests a possible client/backend state synchronization defect.

Please investigate whether:

1. The frontend considered the turn failed while the backend still considered it active.
2. A backend worker remained marked as running after its connection was lost.
3. continue was incorrectly routed or interpreted.
4. Recovery attempted to interact with /feedback unexpectedly.
5. The session contained stale turn state.
6. The original model/worker could no longer be reached.
7. The retry mechanism failed to reconstruct or resume the previous agent state.

TRACE / TELEMETRY CORRELATION

Please investigate:

f162a2399c6ab11ab3279a106bb1caaa

2ad8ca50de4a82e887c4ac68dbd0215c

Also search the user’s recent CommandCode telemetry for the other connection failures occurring around the same period.

The purpose is to determine whether all three incidents share:

• the same backend exception;
• the same model/provider;
• the same gateway;
• the same worker lifecycle event;
• the same timeout;
• the same streaming failure;
• the same client state transition;
• or another common infrastructure signature.

If the previous incidents have their own trace IDs available in telemetry, please include them in the correlation analysis.

REPRODUCTION

Attempt to reproduce the failure using autonomous turns that:

• run for an extended period;
• perform many sequential READ/SEARCH/SHELL operations;
• return moderately large tool outputs;
• transition repeatedly between tool execution and model reasoning;
• use mimo-v2.5-pro;
• use other supported models as a control;
• experience an intentionally simulated transient connection interruption;
• attempt continue afterward.

The objective is to separate:

MODEL-SPECIFIC failure

from

COMMANDCODE RUNTIME failure

from

CLIENT/TRANSPORT failure

from

GENERAL INFRASTRUCTURE failure.

REQUIRED VALIDATION

Do not classify the issue as resolved solely because a single prompt completes successfully.

Validate the fix across multiple unrelated autonomous tasks and multiple consecutive tool/model cycles.

At minimum validate:

1. Multiple unrelated prompts complete without the recurring connection failure.
2. Long autonomous turns remain stable.
3. Repeated tool calls do not destabilize the session.
4. Tool-result → model transitions remain stable.
5. A transient connection interruption can recover correctly.
6. continue actually resumes the interrupted turn.
7. Agent context and completed work survive recovery.
8. Frontend and backend agree on whether a turn is RUNNING, FAILED, or RECOVERABLE.
9. Recovery does not incorrectly invoke or conflict with /feedback.
10. No repository-specific modifications are required to solve the platform failure.

REQUIRED QA REPORT

Please return:

• Root cause, if established.
• If root cause is not yet established, the highest-confidence failure domain and supporting evidence.
• Whether all three incidents appear related.
• Component/service responsible.
• Whether the defect is:
  • CommandCode client,
  • CommandCode backend,
  • model provider,
  • networking/transport,
  • model routing,
  • session management,
  • or another component.
• Explanation of the failed continue recovery behavior.
• Reproduction results.
• Fix implemented.
• Tests/validation performed.
• Remaining uncertainty.
• Temporary CommandCode-side workaround, if applicable.
• Release/version containing the permanent fix, if known.

Do not close this as a repository-specific issue unless telemetry provides affirmative evidence that the user’s local repository caused the CONNECTION failures.

The exported session:

command-code-session-a836e855.html

should be treated as evidence for the most recent occurrence, while the broader investigation should correlate this incident against the other recent connection failures.

Additional context

OS: macOS Tahoe 26.6.2

Bắt đầu với command-code-session-a836e855.html và đối chiếu các trace ID f162a2399c6ab11ab3279a106bb1caaa và 2ad8ca50de4a82e887c4ac68dbd0215c với telemetry gần đây. Tái hiện các lượt tự động kéo dài với các lần gọi công cụ lặp lại và một lần gián đoạn được mô phỏng, sau đó xác minh các session ổn định, khả năng khôi phục bằng continue hoạt động đúng, context được bảo toàn và không xung đột với /feedback.