@syncended/dsh-retry

Configuration

The defaults apply immediately after installation. To customize them, edit the existing model-error-retry row in $DSH_HOME/profiles/ /cordis.patch.yml; do not add a duplicate row with the same id.

- id: model-error-retry
  config:
    maxRetries: 3
    overloadMaxRetries: 5
    retryableCodes:
      - PI_AI_ERROR
      - UNKNOWN
      - STREAM_CLOSED
      - EMPTY_RESPONSE
      - RATE_LIMIT
      - SERVER
      - TIMEOUT
      - TRANSPORT
    initialDelayMs: 750
    overloadInitialDelayMs: 2000
    maxDelayMs: 15000
    jitterRatio: 0.15
    respectRetryAfter: true
    resumeInterrupted: true
    resumeDisposed: false
    resumeMaxAgeMs: 86400000
    resumePrompt: >-
      DeepSeek Harness stopped before the previous task produced a complete final
      assistant response, or before a scheduled retry started. Continue the unfinished
      work from the durable session history. Re-check the current workspace and
      external state before acting. Do not blindly repeat tool calls that may have side
      effects; verify their outcome first.
    providers:
      - openai
      - openrouter
    excludeProviders: []

Restart the Host after changing this file.

Option	Default	Description
maxRetries	2	Retries after an ordinary original failure. 0 disables request-error retries but not resumeInterrupted.
overloadMaxRetries	5	Retries for PI_AI_ERROR with an explicit overload message.
retryableCodes	see below	Exact provider-neutral codes. * matches any request error.
initialDelayMs	500	Initial ordinary exponential-backoff delay.
overloadInitialDelayMs	2000	Initial explicit-overload delay.
maxDelayMs	10000	Maximum local delay. A larger Retry-After leaves the failure terminal.
jitterRatio	0.1	Symmetric jitter ratio from 0 to 1.
providers	[]	Exact provider-route allowlist; empty allows all routes.
excludeProviders	[]	Exact provider-route denylist; takes precedence over the allowlist.
respectRetryAfter	true	Honor provider Retry-After when it does not exceed maxDelayMs.
resumeInterrupted	true	On cold resume, continue an unmatched plugin retry, a model step with no assistant message, or a durably closed tool boundary before its next model step.
resumeDisposed	false	Also accept aborted/disposed as an unfinished retry. Disabled because HMR and intentional teardown can produce this reason.
resumeMaxAgeMs	86400000	Oldest pending retry eligible for automatic continuation; 24 hours by default.
resumePrompt	built in	Model-visible safety instruction for a confirmed interrupted continuation.

Default fallback retry codes:

STREAM_CLOSED, MALFORMED_RESPONSE, PI_AI_ERROR, API_ERROR, MODEL_ERROR,
INTERNAL_ERROR, PROVIDER_ERROR, UNKNOWN

Permanent failures such as AUTH, INVALID_REQUEST, MISSING_CREDENTIAL, UNKNOWN_MODEL, and context overflow are not retried by default. Add an exact code or * only after considering cost and repeated side effects.

Validation rules

• Retry counts must be non-negative safe integers.
• initialDelayMs and overloadInitialDelayMs must not exceed maxDelayMs.
• jitterRatio must be between 0 and 1.
• retryableCodes cannot be empty while request-error retries are enabled.
• resumePrompt cannot be empty when interrupted-session recovery is enabled.
• Provider filters compare exact route strings; they are not regular expressions or prefixes.