Error codes
Checks at the gateway
Each code of this family: what it means, its most frequent cause, what to do and who can do it.
policy_blocked
- What it means
- This call was blocked by a rule of your workspace.
- Most frequent cause
- The rule “‹rule›” does not allow this kind of call.
- What to do
- Use one of the compliant options below if there is one, or ask the rule’s owner for a time-limited waiver, with a short justification.
- Who can do it
- An owner or admin of the workspace · Open the screen
sovereignty_blocked
- What it means
- This call was blocked by the data-residency policy.
- Most frequent cause
- Your workspace does not allow this data to be sent to ‹provider›, whose servers are outside the permitted region.
- What to do
- Use a model hosted in the EU if one is offered below, or ask an administrator to review the residency policy.
- Who can do it
- An owner or admin of the workspace · Open the screen
jurisdiction_blocked
- What it means
- This call was blocked by the law that applies to it.
- Most frequent cause
- The strictest rule of the laws covering this call — where its people are, where your entities are, where ‹provider› runs the model — does not allow it as sent.
- What to do
- Open the decision to see the rule and its reasoning, or ask an administrator to declare a compliant provider in Rules › Jurisdictions.
- Who can do it
- An owner or admin of the workspace · Open the screen
kill_switch_active
- What it means
- This call was blocked: the emergency stop is on.
- Most frequent cause
- An owner or administrator turned on the emergency stop for this workspace, mandate, agent or provider. No call is served while it is on.
- What to do
- Ask a workspace owner or administrator why, and when it can be lifted.
- Who can do it
- An owner or admin of the workspace · Open the screen
data_exfiltration_blocked
- What it means
- This call was blocked: the message contains secret content.
- Most frequent cause
- Part of the message looks like a password, a key or a token, and secrets are never sent to an AI model.
- What to do
- Remove the secret from the message and send it again. No waiver is possible for secrets.
- Who can do it
- Any member of the workspace
prompt_injection_blocked
- What it means
- This call was blocked: the message reads as an attempt to override the model’s instructions.
- Most frequent cause
- This call is governed in enforce, and part of the message asks the model to ignore or reveal its instructions, or to play a persona without rules (a prompt injection). In enforce, such a message is never sent to the model.
- What to do
- Rephrase the request without addressing the model’s instructions. If the message is legitimate, an owner or administrator finds the refusal, and the signal that matched, among the sealed decisions.
- Who can do it
- Any member of the workspace
redaction_failed
- What it means
- This call was blocked: its personal data could not be masked.
- Most frequent cause
- The message contains sensitive personal data and masking it failed, so nothing was sent.
- What to do
- Remove the personal data from the message, or try again in a moment.
- Who can do it
- Any member of the workspace
governance_error
- What it means
- This call was blocked: the governance checks could not run.
- Most frequent cause
- Azoth blocks a call rather than serve it unchecked.
- What to do
- Try again in a moment. If it keeps happening, tell your administrator.
- Who can do it
- Azoth, or the operator of this deployment
loop_detected
- What it means
- This call was blocked: the same call repeated in a loop.
- Most frequent cause
- An agent sent the same request again and again, so Azoth paused it to stop the spend.
- What to do
- Check the agent’s logic, then try again.
- Who can do it
- The developer of the calling code
tenant_concurrency_limit
- What it means
- This call was blocked: too many calls are running at once.
- Most frequent cause
- Your workspace has a limit on simultaneous AI calls, and it was reached.
- What to do
- Wait a moment and try again.
- Who can do it
- The developer of the calling code
human_review_required
- What it means
- This call is waiting for a person to approve it.
- Most frequent cause
- A rule of your workspace requires a human approval before this kind of call.
- What to do
- It runs once another person approves it. Follow it with the tracking link.
- Who can do it
- An owner or admin of the workspace · Open the screen
human_review_not_resumable
- What it means
- This call was blocked: it needs a human approval, and it cannot be held.
- Most frequent cause
- A rule requires a human approval, but this kind of call cannot wait for one.
- What to do
- Send it through the chat API, which can wait for an approval, or ask the rule’s owner to review the rule.
- Who can do it
- The developer of the calling code
approval_expired
- What it means
- This call was not run: its approval expired.
- Most frequent cause
- Nobody approved the call before its approval request expired.
- What to do
- Send the call again to create a new approval request.
- Who can do it
- Any member of the workspace
ungrounded_answer
- What it means
- This answer was not returned: its sources do not support enough of it.
- Most frequent cause
- The mandate is high-stakes, and too few of the answer’s statements were found in the passages sent with the call (an estimate).
- What to do
- Send the passages the answer should rest on, rephrase the question, or ask an administrator to review the mandate’s threshold.
- Who can do it
- The developer of the calling code
grounding_review_required
- What it means
- This answer is waiting for a person to check it.
- Most frequent cause
- The mandate is high-stakes, and the passages sent with the call support too little of the answer (an estimate).
- What to do
- It is returned once another person approves it. Fetch it with the release link.
- Who can do it
- An owner or admin of the workspace
grounding_requires_complete_answer
- What it means
- This call was blocked: its answer must be checked before it is sent.
- Most frequent cause
- The mandate is high-stakes, and a streamed answer cannot be checked against its sources before it reaches you.
- What to do
- Send the call again without streaming.
- Who can do it
- The developer of the calling code
pending
- What it means
- The approval this request waits for has not been decided yet: the held call has not run, the held answer is not handed over.
- Most frequent cause
- The call was resumed, or the answer fetched, before a reviewer approved it in Rules › Approvals.
- What to do
- This approval is still waiting for a decision. Try again once a reviewer has approved it in Rules › Approvals.
- Who can do it
- An owner or admin of the workspace · Open the screen
rejected
- What it means
- A reviewer rejected the approval this request waits for: the held call will never run, the held answer will never be handed over.
- Most frequent cause
- The person who reviewed it in Rules › Approvals judged that the call, or the answer, should not go out.
- What to do
- This approval was rejected: the held call will not run. To try again, change the call — or ask the reviewer why — then send it again.
- Who can do it
- The developer of the calling code
already_executed
- What it means
- The held call of this approval has run already: an approval runs its call once.
- Most frequent cause
- A second resume of the same approval — a retry after a timeout, or two processes resuming it at once.
- What to do
- This approved call has run already: an approval runs its call once, and its answer went to the first resume. To run it again, send the call again — it is governed anew.
- Who can do it
- The developer of the calling code
already_released
- What it means
- The held answer of this approval was handed over already: it is handed over once, then erased.
- Most frequent cause
- A second fetch of the same answer — a retry after a timeout, or two processes fetching it at once.
- What to do
- This held answer was handed over already: it is handed over once, then erased. Use the answer of the first fetch, or send the call again.
- Who can do it
- The developer of the calling code
not_requester
- What it means
- Only the API key that sent a held call can run it once it is approved.
- Most frequent cause
- Another key of the workspace — a key rotated since, or another service — resumed the approval.
- What to do
- This approval belongs to the API key that sent the call: resume it with that key.
- Who can do it
- The developer of the calling code
idempotency_key_invalid
- What it means
- The
Idempotency-Keyheader is empty, too long, or holds a character it may not hold. - Most frequent cause
- A key longer than 255 characters, with a space, or with characters outside visible ASCII.
- What to do
- The Idempotency-Key header must be 1 to 255 visible ASCII characters, without spaces. Send one unique value per call, such as a UUID.
- Who can do it
- The developer of the calling code
idempotency_key_reused
- What it means
- A call already used this
Idempotency-Keywithin the hour with another body: one key names one call. - Most frequent cause
- A client that sends a fixed key, or a retry whose body changed — a new timestamp, another message.
- What to do
- This Idempotency-Key was already used with another request body. Use a new key for a new call; reuse a key only to retry the very same call.
- Who can do it
- The developer of the calling code
idempotency_key_in_progress
- What it means
- The first call with this
Idempotency-Keyhas not answered yet: it is never run a second time meanwhile. - Most frequent cause
- A retry sent before the first call answered — a client timeout shorter than the model’s answer.
- What to do
- A call with this Idempotency-Key is still running. Wait a moment, then retry with the same key to receive its answer.
- Who can do it
- The developer of the calling code
approval_pending
- What it means
- The step this agent execution paused on still waits for a reviewer’s decision: nothing was run.
- Most frequent cause
- The execution was resumed before a reviewer approved the step in Rules › Approvals.
- What to do
- This agent step is still waiting for a decision. Resume the execution once a reviewer has approved it in Rules › Approvals.
- Who can do it
- An owner or admin of the workspace · Open the screen
approval_rejected
- What it means
- A reviewer rejected the step this agent execution paused on: the step will not run, and the execution cannot be resumed.
- Most frequent cause
- The person who reviewed the step in Rules › Approvals judged that the agent should not take this action.
- What to do
- A reviewer rejected this agent step: the execution will not resume. To try again, change the plan — or ask the reviewer why — then run the agent again.
- Who can do it
- The developer of the calling code
approval_not_found
- What it means
- This agent execution has no approval request left to resume it with.
- Most frequent cause
- The request was never filed, its step was resumed already, or the execution belongs to another workspace.
- What to do
- No approval request waits for this agent execution. Check the execution id; if it was resumed already, its result is the first resume’s.
- Who can do it
- The developer of the calling code
already_resumed
- What it means
- The approved step of this execution has run already: an approval runs its step once.
- Most frequent cause
- A second resume of the same execution — a retry after a timeout, or two processes resuming it at once.
- What to do
- This execution was resumed already: an approval runs its step once. Read the execution’s result instead of resuming it again.
- Who can do it
- The developer of the calling code
execution_not_found
- What it means
- No agent execution has this id in this workspace.
- Most frequent cause
- An id typed by hand, taken from another workspace, or sent with a key of another workspace.
- What to do
- No agent execution with this id was found in this workspace. Check the id, and the API key you resume it with.
- Who can do it
- The developer of the calling code
not_awaiting_approval
- What it means
- This agent execution is not paused for an approval, so there is nothing to resume.
- Most frequent cause
- It finished, failed or is still running — or it was resumed already.
- What to do
- This execution is not waiting for an approval: there is nothing to resume. Read its status before resuming it.
- Who can do it
- The developer of the calling code
agent_plan_invalid
- What it means
- The plan the agent’s planner made did not pass its checks, so nothing ran. The answer’s
detailslists what failed. - Most frequent cause
- A goal the planner could not turn into steps the agent may take — a tool it does not have, or more steps than allowed.
- What to do
- The agent could not plan this goal into steps it may take. Rephrase the goal, or give the agent the tools it needs, then try again.
- Who can do it
- The developer of the calling code