dsh-camel

Running long tasks against free models often means a single RATE_LIMIT error kills the whole run, and the built-in retry path simply gives up. dsh-camel plugs that gap: instead of letting a throttled free model terminate your task, it keeps the task alive within a configurable wait window and resumes automatically once the quota recovers.

The design philosophy is prevent first, retry second. Camel maintains an independent adaptive quota state per exact provider and model pair, gradually ramping up during clean windows and backing off on rate-limit hits to reduce how often the limit is triggered in the first place. When a limit is still hit, it honors Retry-After headers and retries in either bounded or unbounded mode, emitting standard llm/retry events so downstream handlers get first refusal before Camel takes over. Every policy is disabled after installation; task-level overrides, pause and resume controls, /camel status, and /camel explain keep the full lifecycle transparent, and no LLM is ever called to classify failures or tune the policy.

It is built for DeepSeek Harness users who rely heavily on free models, hit rate limits frequently, and refuse to manually re-send every throttled turn. Camel owns RATE_LIMIT codes exclusively. It will not switch models, rotate accounts, or make model-based judgments. Pair it with dsh-continue for non-rate network recovery, and the two divide labor naturally by error type out of the box.