What version of the Codex App are you using (From “About Codex” dialog)?
Codex Desktop 26.727.51351 (build 6119)
What subscription do you have?
Paid ChatGPT subscription (exact tier omitted from this public report)
What platform is your computer?
- macOS
26.5.2 (build 25F84)
- Apple Silicon /
arm64
What issue are you seeing?
A long-running Desktop task with three inspection sub-agents entered an unbounded loop:
- The main task repeatedly called the same image-inspection tool on the same four visual inputs.
- It repeatedly emitted near-identical commentary saying that inspection was complete and rendering/export would start next.
- It never transitioned to the requested video editing, rendering, or export work.
- No MP4 was produced and the task's output directory remained empty.
The exported UI transcript contains 190 image-view status events, including 179 identical “viewed 4 images” events. The affected raw rollout grew to 411,653,323 bytes (~392.6 MiB) in only 1,672 JSONL lines, largely because the same 4K images were repeatedly re-read and persisted.
The task ran for at least 2 hours 28 minutes while showing the same inspection/status cycle. The user explicitly asked why it was taking so long. The assistant then apologized and promised five times to stop checking and start exporting, but each promise was followed by another identical image-view action.
A temporary source-path problem occurred when a USB volume became unavailable. However:
- the loop had already started well before that path failure;
- the image was later copied successfully to a stable local workspace path;
- after the copy succeeded, the task still performed another 35 repeated four-image inspections;
- the three video sources remained readable throughout.
Therefore the temporary USB path was not the root cause.
The Desktop logs from the same period repeatedly show event-routing errors across multiple child conversation IDs:
Received item/started for unknown conversation
Received item/completed for unknown conversation
These errors appeared in both the primary renderer and other Desktop renderer contexts. This is consistent with a parent/child conversation registration or rebind race, although that root-cause interpretation is an inference.
What steps can reproduce the bug?
This is the observed reproduction pattern; it is not yet a minimal deterministic test:
- Open Codex Desktop on macOS.
- Start a task with several large visual inputs (in this case three 15-second 4K videos and one 4K image).
- Let the main task spawn three sub-agents to inspect the videos.
- Have the main task inspect four contact-sheet/source images while the child agents run.
- Observe that the three child inspections show as started, but their completion/result state is not reconciled into the parent.
- The parent repeatedly inspects the same four images and rewrites the same progress message instead of advancing.
- Even after a temporary missing source path is restored to a stable local copy, the loop continues.
- Interrupt the task manually.
In the affected rollout:
- 3 sub-agents were spawned;
- 203 agent commentary events were persisted;
- 219 execution records referenced
tools.view_image (221 textual occurrences);
- no execution command invoked
ffmpeg, a renderer, or an exporter.
What is the expected behavior?
- Each inspection subtask should complete once and return a terminal result to the parent.
- The parent should consume those results and proceed to the next requested phase.
- If a child result cannot be reconciled, the task should surface one concrete error or blocker and stop.
- The model/orchestrator should detect repeated identical tool calls with no state change and break the loop.
- Re-reading the same images should not inflate one rollout to hundreds of megabytes.
Additional information
Potentially related:
The distinct symptom here is that the parent does not merely lose a command card or stay visibly idle: it actively re-executes the same image inspection and generates repeated progress commentary indefinitely, consuming time/context and growing the rollout while never performing the requested work.
Privacy: the public report omits usernames, local paths, prompts containing business content, source media, conversation IDs, session IDs, and full logs. A redacted transcript/log excerpt or a privately shared feedback session can be provided if maintainers request it.
What version of the Codex App are you using (From “About Codex” dialog)?
Codex Desktop
26.727.51351(build6119)What subscription do you have?
Paid ChatGPT subscription (exact tier omitted from this public report)
What platform is your computer?
26.5.2(build25F84)arm64What issue are you seeing?
A long-running Desktop task with three inspection sub-agents entered an unbounded loop:
The exported UI transcript contains 190 image-view status events, including 179 identical “viewed 4 images” events. The affected raw rollout grew to 411,653,323 bytes (~392.6 MiB) in only 1,672 JSONL lines, largely because the same 4K images were repeatedly re-read and persisted.
The task ran for at least 2 hours 28 minutes while showing the same inspection/status cycle. The user explicitly asked why it was taking so long. The assistant then apologized and promised five times to stop checking and start exporting, but each promise was followed by another identical image-view action.
A temporary source-path problem occurred when a USB volume became unavailable. However:
Therefore the temporary USB path was not the root cause.
The Desktop logs from the same period repeatedly show event-routing errors across multiple child conversation IDs:
These errors appeared in both the primary renderer and other Desktop renderer contexts. This is consistent with a parent/child conversation registration or rebind race, although that root-cause interpretation is an inference.
What steps can reproduce the bug?
This is the observed reproduction pattern; it is not yet a minimal deterministic test:
In the affected rollout:
tools.view_image(221 textual occurrences);ffmpeg, a renderer, or an exporter.What is the expected behavior?
Additional information
Potentially related:
unknown conversation/ child-thread rebind symptoms, but that report focuses on a missing long-running command card and dropped output deltas.The distinct symptom here is that the parent does not merely lose a command card or stay visibly idle: it actively re-executes the same image inspection and generates repeated progress commentary indefinitely, consuming time/context and growing the rollout while never performing the requested work.
Privacy: the public report omits usernames, local paths, prompts containing business content, source media, conversation IDs, session IDs, and full logs. A redacted transcript/log excerpt or a privately shared feedback session can be provided if maintainers request it.