Skip to content
Docs menu

Requests and request traces

Search every request your organisation sent through Venom Cloud, open one request's trace, read its stored prompt and response, and set how long content is kept.

View .md

The Requests page (/requests) is the console's log explorer. Every request that reached Venom Cloud is a row: which key sent it, which model it asked for and which one served it, the daemon or provider it went to, status, tokens, cost and latency. You type one line of the log query language to filter it, group it or chart it.

Who can open it: any role with logs:read (owner, admin, developer, viewer). The billing role cannot.

A first search

Open Requests and type, then press Enter:

status:5xx since:1h

That lists the server errors of the last hour, newest first. Some more to try:

key:ci-* is:error | count by error
model:gpt-* latency:>2s
| timeseries 5m count by status

The URL carries the whole view (query, time zone, columns, comparison, chart type, open row), so you can paste it to a teammate. Without a time term the page reads the last 24 hours and says "last 24h (default)".

The page

Part What it does
Query bar Parses as you type: colours, red squiggles with a one-click fix, completion of fields and of values seen in your own recent rows. Runs on Enter, or after 700 ms of quiet when the query has no error.
Window picker Quick ranges, ‹ and › to shift the window, local zone or UTC. Dragging across the histogram zooms in.
Volume histogram Requests over the window. With Compare on, a dashed overlay of the previous period, yesterday or last week, and unusual buckets marked.
Fields sidebar Every field of the matched rows: the top 10 values with counts, or p50, p95, p99, min, max and average for numbers. Click a value to filter, Alt-click or − to exclude, add the field as a column, or run a one-click aggregation such as "Cost by key". It reads the newest 20,000 matches.
Rows / Patterns tabs Rows: the requests, with columns you choose, reorder, resize and pin. Patterns: the rows grouped into request shapes (status, error, fallback, path, how it was served) with a count and a sparkline each.
Row panel Opens a row in place: every field, and "Surrounding": what the same session, key or daemon did within ±1, 5 or 15 minutes.
Live tail New rows every 3 s; pauses while your pointer is over the rows or a row is open.
Saved queries Save the current query under a name, privately or shared with the org. Manage: rename, share or unshare, copy a link, delete. At most 200 per person. Not available to the operator until an org is picked.
Share menu Copy link, Copy as curl (the same query against the admin API), CSV or JSON of the loaded rows, Download every match, Create alert.
Disclosure line Says whether request content is logged here and for how long, with a link to Data & logging.

Keyboard shortcuts

Outside a text field; ? shows the list.

Key Does
/ focus the query
Enter run (in the query bar); open the highlighted row
s save the query
j / k next / previous row
h / l previous / next value in the row
o open the row
f / x filter to / exclude the highlighted value
c copy the row as JSON
t open the time picker
[ / ] shift the window back / forward
Shift + L live tail on or off
b fields sidebar on or off
p rows or patterns
Esc close the row or the sheet

Export

Choice What you get Who
CSV The loaded rows, shown columns only, up to 10,000 rows, built in the browser. Cells that would run as spreadsheet formulas are defused. anyone who can open the page
JSON The loaded rows, every field, up to 10,000 rows. anyone who can open the page
Download every match (audited) Every row the query matches, from the server, NDJSON or CSV, at most 50,000 rows. Only for a plain search (no | stages, no content terms). "Include prompts and responses" adds the stored bodies. Writes a logs.exported audit row, and content.exported with bodies. org:manage (owner, admin); bodies also need content:read
Create alert Opens Alerts with the current search, its time range and pipes dropped. org:manage

One request: /requests/{id}

Click a row's time, or open /requests/{id} with a request id, to see its trace:

  • Header: the model, the status, the error code (client_gone (partial) when the client left mid-stream and was billed for what was streamed), a background badge, the id with a copy button, and buttons for "Same session" and "This key's requests".
  • Figures: latency and time to first token, tokens in and out (and cached), cost, attempts (marked when hedged or when it fell back).
  • Who and what: key, model (and the served model when it differs), endpoint, session. Each value links back into the explorer filtered to it.
  • Where it went: served by the cloud's own provider key or by a daemon (linked to the daemon page), the mode (direct or relay), the route reason, the fallback reason, routing time.
  • Timeline: each relayed job as a bar: the first attempt, failovers to the next candidate, hedges. A direct request has no job (the client went straight to the daemon with a ticket); jobs are kept 7 days, the request's metadata longer.
  • Route decision: per job, the eligible daemons in cost order when it was queued, with the one that took it marked.
  • Prompt and response: the stored bodies, behind a "Show prompt and response" button. Opening them writes a content.viewed audit row.

The content card says why nothing is shown when the body is not stored (direct mode, content logging off, no vault key on the deployment) or when your role can't read it.

Data and logging: /settings/logging

Settings, Data & logging, says what Venom Cloud keeps about your requests.

Kept For how long Notes
Metadata: key, key group, model, route, daemon or provider, tokens, cost, latency, status up to 90 days always
Prompt and response bodies the org's retention: default 30 days, 1 to 90 Up to 1 MiB each, for every request the cloud relays or serves. Sealed with the org's key, deleted at expiry. Needs VENOM_VAULT_KEYS on the deployment; without it only metadata is kept.
Searchable shapes: tool names called, image sent, message count, prompt and response lengths with the bodies names and sizes only, never text

Requests sent directly to a daemon (307 with a ticket) never pass through the cloud, so their bodies are not seen or stored there.

Control Who Effect
Retention (days) content:settings (owner, admin) Applies to new bodies. Bodies already stored keep the expiry they were stored with. Audited as content.settings; the change can be undone from the audit log.
Content logging on or off the deployment's operator only Off: only metadata is kept for that org.

Reading a stored body needs content:read (owner, admin, developer). Viewers and billing members see the metadata only.

Troubleshooting

Symptom Cause and fix
"No such request here" on /requests/{id} The id belongs to another organisation, is mistyped, or the row passed the retention period.
"Stored, but you can't open it" Your role lacks content:read, or the deployment has no VENOM_VAULT_KEYS.
"Not stored" The request went direct to a daemon, content logging is off for the org, or the body expired.
The query doesn't run and shows a red squiggle A query error: hover it for the message and the fix. See diagnostics.
403 content_forbidden You searched prompt:, response: or content: without content:read.
504 query_timeout The query ran past 5 s. Narrow the window or add a filter.
No "Download every match" Only owners and admins get it, and only for a plain search without pipes or content terms.