Outage explanation and one-click repair
HYPHA tells you what is happening, the probable cause and its evidence, the fix. When the fix is a known action, it proposes it: a human approves, the agent runs it.
What the video shows
- The Problems view: four current alerts, sorted by priority.
- The Explain button of the "systemd unit failed" alert: shop-worker.service on web-01.
- The explanation panel: the service killed by memory exhaustion, the log lines that prove it, the fix. The wait for the analysis is cut.
- The proposed action, "Restart shop-worker.service on web-01", to approve in one click.
Explain
The Explain button sits on an alert (Problems view, overview, monitors), on a host page, on a log line's details, on a trace and on an application error.
HYPHA then gathers, within the workspace and for the object concerned:
- the host: system, cores, memory, tags, last contact;
- the monitor: query, thresholds, state, value and its ten latest changes;
- the host's non-OK checks;
- its logs from the last 30 minutes, errors first, repeated messages grouped;
- its events from the last 24 hours;
- its metrics from the last 30 minutes, summarised (CPU, memory, load, disks, swap, network);
- its most demanding processes;
- for a trace, the error spans and the critical path.
The model answers in four parts: What is happening, Probable cause (with its evidence: a log line, a measured value), Fix (commands for the host's system), Preventing it. It must not make anything up; if the evidence is not enough, it says so and states what to check. The panel also shows the model used, the duration and the context read.
Where the model runs
- Hosted service: on S33D's GPU servers, in France. Data sent to the model goes to no third party.
- Self-hosting: on your own Ollama-style model server (
HYPHA_LLM_URL). Without a model server, the button does not appear.
Limit: 60 explanations per hour per workspace; the monthly number depends on the plan (see pricing).
Repair in one click
HYPHA reads the Fix part again. If it finds an action it knows (systemctl restart|start|stop <unit>, Restart-Service <service>, service <name> restart), it proposes it as a button, for example "Restart shop-worker.service on web-01", on two conditions: the host's agent accepts that action, and the unit or service exists on the host.
- A click opens a confirmation: action, host, reason.
- If you are an administrator, the action is created and approved in the same step. Otherwise, it waits for an administrator's approval in Settings › Actions.
- The agent receives it through its control channel, checks that it is allowed, runs it, then returns the exit code and the output.
- Follow-up is live: pending, sent, running, succeeded or failed. The agent sends the unit's state right away.
An unapproved action expires after one hour; an approved action that the agent has not picked up expires after 10 minutes. Everything goes to the audit log.
What the agent accepts
Actions are off by default. On each host, you decide:
actions:
enabled: true
allow: [restart_service, start_service] # allowed actions
services: ["nginx*", "shop-*"] # targeted units or services
deny_services: ["sshd*", "ssh*", "hypha-agent*", "systemd-*"]
scripts:
clear-cache: /etc/hypha/scripts/clear-cache.sh
timeout: 120s
- No free command: only named actions (
restart_service,start_service,stop_service,run_script,update_agent). - An action outside the list, a forbidden service or an expired action are refused by the agent, with the reason.
- The agent's own service is never touched.
- A script is designated by its name; its path comes from the agent's configuration, never from the server. It runs without arguments, its output is capped at 64 KiB.
- In a container, actions on the host's services are not available.
An action can also be proposed by hand, through the proposal form (action, unit or script, reason) or the API. History and pending actions are in Settings › Actions.
Updated: 19 November 2026