Mind · Strategy
Retry
Classified re-attempt with backoff conditioned on failure type
Definition
Mechanism
Intelligent re-attempt of failed coordination with failure-informed strategy. After BREAK + COMPENSATE, agent evaluates: (1) CLASSIFY failure—transient (timeout, rate-limit, network blip) vs persistent (capability gap, protocol mismatch, explicit rejection). (2) CHECK retry_hint from BREAK (partner may say 'don't retry' or 'wait 30s'). (3) CONSULT failure_history—same error repeating? CircuitBreaker threshold reached? (4) COMPUTE Backoff—adaptive based on failure type: transient uses ExponentialBackoff, persistent uses longer fixed delay or triggers abort. (5) VERIFY changed_conditions—has something changed that makes retry worthwhile? (6) EXECUTE retry if within budget and conditions favor success, else ABORT with retry_exhausted status. Retry CARRIES FORWARD: failure context, partner state observations, environmental data. Retry RESETS: coordination state (fresh start, don't resume mid-stream). It handles transient failures by re-queuing the task, distinguishing them from terminal failures that trigger Break and Compensate.
Preconditions
- A failure has occurred and its details are available to classify
Postconditions
- Success or definitive failure
Invariants
- Classifies the failure before deciding, and records the classification.
- Backoff applied between attempts.
- Attempts do not exceed `max_attempts`.
Risks
Failure modes
Misclassifying persistent failure as transient (wastes retry budget).
Misclassifying transient as persistent (gives up too soon).
Backoff too aggressive (slow recovery from transient issues).
Backoff too timid (hammers already-failing system).
CircuitBreaker too sensitive (abandons recoverable situations).
Retry succeeds but same failure recurs (didn't address root cause).
Changed_conditions check misses relevant changes.
Relationships
Builds on
Composes with
Infrastructure
Geometrically increasing capped retry delay
References
Infrastructure
Delay subsequent attempts after failure or contention
Infrastructure
Failure-announcement signal {severity, reason, recoverable}, ACK-required
Infrastructure
Fail-fast protection against cascading failures
Infrastructure
Structured rollback via logged inverses in LIFO order
Used by
References
Pre-computed responses to critical failure
Typed failure signaling that lateral reframing, not retry, is required
When improvement stalls under one method, change the method instead of repeating it
Infrastructure
Substrate-level re-attempt: same call, after a delay
Reference
Parameters
max_attemptsInteger, [1, 10]- Attempts before giving up. Referenced by the third invariant, which previously named a Max_Retries nothing defined.