Diagnostics tab¶
Diagnostics is where you find out what happened or why something isn't working.
It has four subtabs, each with its own address: Audit (/diagnostics/audit, the
default), Logs, Performance and Doctor. A blue or red badge on the tab
means Doctor has something for you.
Old bookmarks to /audit, /logs, /debug, /doctor and /activity/... still
work. The command palette has an entry for each subtab. On a peer's page, Audit needs
that feature, and the others need unrestricted trust.
Doctor¶
What this host and install still need for tclaude to work properly. Start here when something doesn't work. The default Needs attention view shows what matters, and All checks shows everything. The sidebar groups checks into Decisions, Sandbox, Tools, Harnesses and Hooks & setup.
- 🔵 decision: a question only you can answer, such as the default sandbox for new agents (see The install default). Answer it in place.
- 🔴 blocking: something stops tclaude or spawns, such as missing
tmux, no harness, or tclaude's sandbox being the default on a host where it can't run. - 🟡 warning: degraded behavior, such as missing harness hooks.
- ⚪ info and ✅ ok: optional extras and passing checks.
A fix is a command to copy (sudo apt install bubblewrap, tclaude setup) or a
button. Some fixes open a batch preview: it lists what the fix would change, with
every item selected. Deselect any you want left alone, then apply. One example is
restoring agents that lost their --harness-config
on a relaunch. Another is the cron jobs, standing orders, automations and runs that
still refer to retired or deleted agents: copy
tclaude agent cleanup-dependents --orphaned.
The view never changes the host by itself. Results are cached for a minute;
Re-check refreshes them.
Audit¶
The trail of changes made through the daemon: spawns, messages, lifecycle actions, group and permission changes, sudo grants, proxy calls and power actions. Read-only polling isn't recorded.
Columns are When, Actor (operator, agent, tclaude or unknown), Action (⊞ means it came
from the dashboard), Target, Detail and Outcome (ok, denied, rejected or err). Denied
attempts are included. Spawn rows have a ? with the request and resolved
parameters. Search and filter by source and outcome.
Records are kept for 30 days (audit.retention_days; negative keeps them forever).
See Permissions and audit for what's recorded.
Logs¶
A paged view of ~/.tclaude/data/output.log, with level pills, structured fields,
search, a minimum level, a since filter, and a 2-second live stream. Turn on
rotated to include rotated files.
⤓ export downloads every line matching your filter as JSON lines, oldest first. It ignores paging and the view's 64 MiB limit, so it can cover more than the tab shows.
Reading a 🚫 refusal¶
The 🚫 badge means the daemon refused a callback
from a session. Each refusal logs a broker: refused a brokered callback warning, with
the caller's process, the row it resolved to, the ID it claimed, and which proof was
missing. Repeats within ten seconds are counted, not logged again.
Two related lines are about timing, not refusal:
broker: caller exited before the daemon handled its callbackmeans the caller left first. No badge is set, because a later callback replaces it. It's usually INFO: harnesses drop callbacks when they redraw or exit. It's a WARN only when the caller had waited at least half its time budget, which points at a slow daemon.broker: slow brokered requestis written when a callback takes over a second, and breaks the time down by phase.
Performance¶
⏱ Profile daemon… captures CPU, heap, goroutine, block and mutex profiles and a trace of this agentd. Each capture lists its CPU split and links to pprof's web UI. zip downloads it, and share… sends its files to an agent. See Profiling agentd.
Below that is poll latency per endpoint: p50, p90, p99, max, a sparkline and a breakdown by phase. It resets when the daemon restarts or with the reset button.