Agent & context

agent & retry#

agent {
    max-turns 0          // 0 = unlimited
    worker-max-turns 0  // 0 = unlimited; caps a subagent worker's own turns
}
retry {
    max-retries 10       // auto-retry on timeout/reset/HTTP 408/429/5xx; 0 disables
}

Retries back off on a ladder: 3 s, 5 s, 10 s, 20 s, 30 s, 60 s, …

A retry resumes the turn instead of replaying it. The failed attempt’s partial reply and the tool calls that already returned ride the retried request, so the model carries on with the work it had done and re-issues only what it still needs — the prompt is not sent again as a fresh request, and the reply lands on the same turn rather than under a second, half-finished one. See Sessions.

context#

Context-window management (enabled by default; disabled #true turns it off). One shared input-token budget drives trimming, the overflow guard and compaction:

context {
    reserved 20000                  // tokens kept for the model's reply; input budget = context_length − reserved
    keep-recent-tokens 20000        // tokens kept verbatim as the "tail" when older messages are trimmed
    tool-output-max-chars 16000     // cap on tool output sent to the model (truncated with a read hint); 0 disables
    tool-output-max-bytes 50000     // byte cap applied on top of the char cap; 0 disables
    fallback-context-length 128000  // used when the model catalog has no context length
}

When the input budget is exceeded, Shuvarie compacts: older messages are summarized by the model (the recent tail within keep-recent-tokens stays verbatim) and the summary is spliced in at the cut point, so the ride goes on unbroken. Run it on demand with /compact [instruction]; the optional instruction focuses the summary.