ankra chat
AI-powered chat for troubleshooting and assistance. If a message is provided, sends a one-shot question and prints the conversation id so a follow-up can continue it with--conversation.
If no message is provided, enters interactive chat mode.
Use --cluster to provide cluster context for better answers; without it the
persisted ‘ankra cluster select’ applies, and with neither the chat reads
across the whole organisation.
A one-shot question takes -o json (or -o yaml) for scripts and CI jobs: the
answer is printed once, when the turn ends, as one document carrying the
answer text, the conversation id, the safety mode the turn ran in, the
cluster it was scoped to, the tools the model ran and any write left
awaiting confirmation. A turn that starts and then fails still prints the
document, with its “error” member set, and exits non-zero; a question the
platform refuses to start (a rejected token, an unknown cluster) prints
nothing on stdout and exits with that error’s code. The one exception is
AI paused by the organisation’s billing standing: that refusal prints the
document with “error_code” BILLING_RESTRICTED and the platform’s sentence in
“detail”, whether it came before the turn or on it, so a job can tell it
from a failed answer. Interactive mode has no structured output.
A one-shot question can end before the conversation does, because the agent
paused on a condition and the platform resumes it by itself. That is not a
finished answer: the command exits 9, and the document’s “state” reads
“waiting”. Pass --wait to follow the conversation through its pauses instead;
it exits 5 if a pause outlasts --wait-timeout, with the conversation still
running. “state” is otherwise “completed”, “failed”, or “awaiting_user” for a
turn that needs a confirmation or an answer, which exits 0.
ankra chat actions
Agent-mode chat halts before every write and proposes it as a pending action. Confirm the action to run it, or reject it to discard it. Pending actions expire, so confirm promptly. If the cluster changed since the action was proposed, confirming answers a drift conflict; re-run with--force
to apply it against the changed state anyway.
ankra chat actions confirm
Confirm a pending action proposed by agent-mode chat. The platform re-checks the action against live cluster state before running it. If the state drifted since the proposal, the command reports the drift and exits without running anything; pass--force to run it anyway.
ankra chat actions list
List the actions awaiting confirmation in a conversationankra chat actions reject
Reject a pending AI action so it never runsankra chat delete
Delete a chat conversationankra chat health
Get AI-analyzed cluster healthankra chat history
List chat conversation historyankra chat plans
A chat conversation that needs several writes proposes them as one plan and halts. Approving the plan runs its steps in order; rejecting it discards them. Plans expire, so decide promptly. A plan with high-risk or irreversible steps is only approved when each of those steps is named with--acknowledge.
ankra chat plans approve
Approve a proposed plan and follow its steps as they run. Approving is the decision: there is no further prompt, exactly as with ‘ankra chat actions confirm’. Read the plan first with ‘ankra chat plans get <plan_id>’. A plan with high-risk or irreversible steps is refused, with nothing run, until each of those steps is named with--acknowledge <step-order>; the refusal lists them. It needs the ai.execute
permission.
Exit codes: 0 when every step completed (or the plan was handed to the chat
session that proposed it, which runs the steps), 8 when steps still need
--acknowledge, 3 when the plan is not found, 1 when the plan could not be
approved or a step failed.