Retry a failed job
retry_jobFREE when the original KIE task or a validated stored artifact can be recovered; otherwise PAID only for the sub-steps that still need to run. Re-runs the FAILED step of a job whose status is "error" — the prompt, model and every setting are already on the job, so this is the cheap way out of a failure instead of creating the whole thing again. Work that already succeeded is NOT redone or re-charged: a generation that failed after its image landed reuses the paid result. get_job reports retryMode, retryRecovery and retryCost, and the dryRun quote here is that same number. If get_job reports failureAction:"edit_input", do not use this tool; edit the prompt or references and create a new generation. A non-retryable phase answers 400 JOB_NOT_RETRYABLE — some late phases hold partially-committed state and deliberately have no blanket retry. Eligible overhead Material jobs use stored_artifact recovery to repair the original without new generation or background removal. A confirmed failed repair is not offered again by that processor; refresh get_job and honor server eligibility, even below the general two-attempt limit. No automatic refund is implied. Confirmed background-removal provider failures already fall back automatically on the server, without a new user charge. There is no daily removal retry lockout; honor reported cooldowns and unknown-task safeguards if recovery still fails. For already-authorized generation, automatically execute a zero-credit recovery quote and poll it; no renewed approval or refund. Never duplicate a running recovery. Limit automatic retries to two per failed job per session; stop on recurring failure or any paid quote. DEFAULTS TO A COST PREVIEW — see the dryRun argument. Returns { jobId } — poll get_job.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| jobId | Yes | Id of a job whose status is "error" (from get_job). | |
| dryRun | No | DEFAULTS TO TRUE. While true this returns only a cost quote ({ estimatedCredits, balance, spendCapDaily, spentLast24h, capRemaining }) and executes nothing. Show the user estimatedCredits, then ask a normal confirmation question. Prefer the client's native question UI with Approve / Decline / Discuss choices when available; otherwise accept any unambiguous conversational approval. Never require a fixed phrase or ask the user to type a magic word. Only after approval, re-call with dryRun:false to actually spend. | |
| idempotencyKey | No | Optional Idempotency-Key for the real (dryRun:false) call. Omit and one is minted per call. Reuse the SAME value when retrying a call that failed with ENTITY_BUSY / 402 / 429 so the retry cannot double-dispatch. |