deepseek-harness/packages/llm/llm-retry

@deepseek-ai/dsh-llm-retry

Function plugin that retries selected transient model-request failures on the agent loop's closed-step recovery seam. It does not wrap ctx.llm.stream(): every adapter call remains one provider attempt, and every retry opens a fresh numbered step.

The default policy permits two retries for RATE_LIMIT, SERVER, TIMEOUT, and TRANSPORT, using bounded exponential backoff from 500 ms to 10 seconds with 10 percent jitter. Delay bounds must fit Node's supported timer range. A valid providerRetryAfterMs replaces local backoff when it is within the configured cap; an over-cap instruction delegates to the next recovery policy instead.

Before waiting, the plugin appends a non-surface llm/retry event with the failure and scheduled delay. Cancellation and plugin disposal abort the wait; disposal drains the plugin's active backoffs, and a callback captured before disposal fails closed if invoked afterward.

- name: '@deepseek-ai/dsh-llm-retry'
  config:
    maxTransientRetries: 2
    initialDelayMs: 500
    maxDelayMs: 10000
    jitterRatio: 0.1
    retryableCodes: [RATE_LIMIT, SERVER, TIMEOUT, TRANSPORT]

Model Experience

Transient request recovery

What the model sees

No retry event, delay, or failure prose is model-visible. After a retry, the next numbered step reconstructs the same explicit provider/model request from durable session history; failed chunks never enter derived messages.

Token effect

Each retry is a new provider request and may repeat input-token billing. The finite budget caps attempts; llm/retry itself contributes no tokens.

KV Cache effect

The reconstructed request preserves the prior prefix and is eligible for provider cache reuse under that provider's rules. The non-surface status event does not change cache identity.

Known Limitations and Deferred Work

• Agent steps are the only retry boundary — direct ctx.llm.stream() consumers remain single-attempt because a raw stream cannot separate already-emitted chunks durably.
• Finite plugin budgets add — this policy counts only configured transient codes; context-overflow compaction counts only its own code. A future policy with overlapping codes must document and test registration-order behavior.
• llm/retry records scheduling, not completion — later step and turn events establish success, exhaustion, or cancellation.