[GH-ISSUE #157] [FEATURE]: Survival mode for Anthropic Subscription's 5-hour rolling usage windows #54

New issue

Closed

opened 2026-02-27 07:20:11 +03:00 by kerem · 0 comments

kerem commented

2026-02-27 07:20:11 +03:00

Owner

Originally created by @UmbraDeorum on GitHub (Feb 22, 2026).
Original GitHub issue: https://github.com/KeygraphHQ/shannon/issues/157

Originally assigned to: @ajmallesh on GitHub.

When an agent hits Anthropic's usage limit mid-run, the current PRODUCTION_RETRY policy caps backoff at 30 minutes:

tsconst PRODUCTION_RETRY = {
  initialInterval: '5 minutes',
  maximumInterval: '30 minutes',  // ← too short
  backoffCoefficient: 2,
  maximumAttempts: 50,
};

Anthropic's usage limits reset on a rolling 5-hour window from when the session started for subscription models. With a 30-minute ceiling, the retry schedule plateaus well short of that window, causing the workflow to terminate with no completed agents and an un-resumable workspace.

The parallel phase launches all 5 vuln/exploit pipeline pairs simultaneously via Promise.allSettled, which spikes LLM usage and makes hitting the limit more likely.

pipeline:
  max_concurrent_pipelines: 5

Describe the solution you'd like

Either provide a preset 5+hour safe timeout for subscription users as a config option, or expose the values in PRODUCTION_RETRY, so that operators are able to configure them before launching (or both).

Exposing a max_concurrent_pipelines config option would let operators tune the number of simultaneous pipelines launched, allowing for better control in low-tier subscriptions.

This requires adding the field to Config in src/types/config.ts and passing it through to buildPipelineConfigs() in src/temporal/workflows.ts.

Include a section in README for subscriptions users, so that they may set this configuration properly beforehand, saving them from failed runs.

Describe alternatives you've considered

At this time, the only alternative is to directly modify the values in the respective sections in the code, which most likely means that a failed run has already taken place before someone goes through the effort of looking for them.

Additional context

No response

Originally created by @UmbraDeorum on GitHub (Feb 22, 2026). Original GitHub issue: https://github.com/KeygraphHQ/shannon/issues/157 Originally assigned to: @ajmallesh on GitHub. ### Is your feature request related to a problem? Please describe. **A.** When an agent hits Anthropic's usage limit mid-run, the current `PRODUCTION_RETRY` policy caps backoff at 30 minutes: ``` tsconst PRODUCTION_RETRY = { initialInterval: '5 minutes', maximumInterval: '30 minutes', // ← too short backoffCoefficient: 2, maximumAttempts: 50, }; ``` Anthropic's usage limits reset on a rolling 5-hour window from when the session started for subscription models. With a 30-minute ceiling, the retry schedule plateaus well short of that window, causing the workflow to terminate with no completed agents and an un-resumable workspace. **B.** The parallel phase launches all 5 vuln/exploit pipeline pairs simultaneously via Promise.allSettled, which spikes LLM usage and makes hitting the limit more likely. ``` pipeline: max_concurrent_pipelines: 5 ``` ### Describe the solution you'd like **A.** Either provide a preset 5+hour safe timeout for subscription users as a config option, or expose the values in PRODUCTION_RETRY, so that operators are able to configure them before launching (or both). **B.** Exposing a `max_concurrent_pipelines` config option would let operators tune the number of simultaneous pipelines launched, allowing for better control in low-tier subscriptions. This requires adding the field to Config in `src/types/config.ts` and passing it through to `buildPipelineConfigs()` in `src/temporal/workflows.ts`. Include a section in README for subscriptions users, so that they may set this configuration properly beforehand, saving them from failed runs. ### Describe alternatives you've considered At this time, the only alternative is to directly modify the values in the respective sections in the code, which most likely means that a failed run has already taken place before someone goes through the effort of looking for them. ### Additional context _No response_

kerem closed this issue

2026-02-27 07:20:11 +03:00

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

starred/shannon-KeygraphHQ#54

No description provided.

Rows
Columns