Rabbit Chat Studio records model requests and tool calls locally, so you can answer "why did the model say that" and "what did the tool actually do". There are three surfaces: the Interaction Monitor, the Agent Mission Monitor and diagnostic logs.
Interaction Monitor
Open Settings → Runtime & Debug → Interaction Monitor.
The entry appears when a chat session exists; it queries the current
session_id.
UI
The panel follows the light/dark theme and the active palette like every other surface. The left side is the interaction list; the right side shows details, and the bottom third is a session report over the current session.
List items show:
- interaction type badge;
- actor (model name, tool name, MCP server);
- action name;
- timestamp;
- error message (if any);
- output preview.
Details show:
| Field | Description |
|---|---|
| Type | llm_api / llm_error / tool_output / mcp_call / mcp_response, etc. |
| Actor | Model or tool |
| Action | e.g. chat_completion, memory, read_file |
| Timestamp | Local time |
| Duration | Milliseconds |
| Error | Failure reason |
| Input | Request body / tool arguments, JSON-formatted |
| Output | Response / tool result |
| Metadata | Extra information |
The session report aggregates the current session in SQL (get_session_stats over api_requests + interaction_log): total tokens, cache hit/miss with the weighted hit rate, the share of the context window in use, request count, thinking time per round, tool traffic and the busiest tools. The context window is looked up from the newest request's model and shows a dash when unknown; a round without thinking stores NULL for its reasoning time, so averages are not diluted by fake 0 ms rounds.
Top controls:
- Auto-refresh: every 2 seconds;
- Clear logs: delete all rows from
api_requestsandinteraction_log(not filtered by session); chat history is untouched; - Compact database: apply the retention window and
VACUUMto reclaim disk space.
What is logged
| Type | Trigger |
|---|---|
llm_api |
Each successful Chat Completions request; full request body, response content, duration |
llm_error |
Failed request; error and duration (request body is not duplicated) |
tool_output |
Built-in tool result |
mcp_call |
MCP tool invocation |
mcp_response |
MCP tool result |
api_requests also stores a metrics row: model, finish_reason, prompt/completion tokens, duration, reasoning time (reasoning_ms, the first delta after streamed reasoning) and time to first token (first_token_ms), plus error. The Interaction Monitor mainly displays interaction_log; a log row can read "· thinking 3.6s" from those columns.
Retention
Settings → Runtime & Debug → Log retention days:
- default 90 days;
0= keep forever;- on startup, rows older than the window are deleted from
api_requests/interaction_log; - Compact database applies the policy again and runs
VACUUM; - chat history
historyis never pruned automatically.
Privacy
llm_api input includes the full request body: system prompt, conversation, attachment text and tool results may all be there. Request bodies are the main driver of chat.db growth and may contain secrets or business data. Compact/clear logs before sharing chat.db or an export.
Agent Mission Monitor
Open Settings → Runtime & Debug → Agent Missions Monitor.
It shows the current session's mission records:
| Field | Description |
|---|---|
| Mission | Root task description |
| Agent | Agent that executed it |
| Status | running / completed / stopped, etc. |
| Created/updated | Database timestamps |
| Context | Root task context |
| Active tasks | Each task's id, name, description, status |
| Deliverables | Files the finished mission left in the exchange area |
| Episodic summary | Summary saved during execution |
| Final report | Report written by mark_mission_accomplished |
Refreshes every 2 seconds, can be refreshed manually, and has a Deliverables folder button that opens <workspace>/.data/deliver/<session-id> — the sub-agent deliverable exchange area — for manual cleanup.
Tables:
agent_missionsagent_tasks
See Sub-Agents.
MCP diagnostic logs
MCP panel → Logs on a server:
- in-memory ring buffer; cleared on restart;
- records save, start,
initialize,tools/list,tools/calland errors; - refreshes every 1.5 seconds;
- can be refreshed or cleared;
- testing opens the log automatically.
This is the first place to look when MCP fails: npx download, PATH, auth and protocol handshake errors all appear here.
app-<date>.log
Location (app-<YYYY-MM-DD>.log, one file per local day):
| Platform | Path |
|---|---|
| Windows | %APPDATA%\com.leonard.aichat\app-<date>.log |
| macOS | ~/Library/Application Support/com.leonard.aichat/app-<date>.log |
| Linux | ~/.local/share/com.leonard.aichat/app-<date>.log |
Notes:
- Under the Microsoft Store (MSIX) build, Windows redirects this path to
%LOCALAPPDATA%\Packages\XLBBB.RabbitChatStudio\LocalCache\Roaming\com.leonard.aichat\app-<date>.log. - Only debug builds (
cargo/tauri dev) write it; release builds disable it by default. - Settings → Runtime & Debug → Log output selects file vs terminal (
println); debug builds only. - It contains startup, config apply, workspace switch, MCP start, auto-accept, context compression, goal-loop continuations and request success/failure.
- Every line carries a process tag (
[primary-<pid>]/[shadow-<pid>]), so several app windows (multi-instance) writing the same file stay attributable. - The date is resolved per log line, so a window running across midnight starts the next day's file automatically.
- Old files are swept at startup by the same log retention days setting that prunes the database log rows (
0keeps them forever). The legacyapp.logis no longer written and is left in place.
Debugging workflows
The model's answer is wrong
- Interaction Monitor → latest
llm_api; - inspect the system message and history in Input: system prompt, skills, todos;
- check finish_reason, tokens and duration;
- for
llm_error, read the error and HTTP status.
A tool did not run
- Is there a
tool_outputentry? - None: the model did not call a tool. Check the tool toggle and whether the model supports tools.
- Present but wrong: inspect the tool input/output, paths, arguments and permissions.
- Stuck: is a confirmation dialog waiting? L0–L3 commands time out after 120 seconds.
MCP tool unavailable
- MCP panel → Logs;
- check whether
initializeandtools/listsucceeded; - check command/PATH/URL/token;
- run Test connection.
Sub-agent failed
- Agent Mission Monitor: mission status and task errors;
- Interaction Monitor: the sub-agent's tool calls;
- CLI: process output and timeout; cloud: HTTP error and polling status.
FAQ
| Symptom | Explanation |
|---|---|
| Monitor is empty | No interactions in this session, or logs were cleared |
| Disk usage did not shrink after clearing | Deleting rows does not shrink the file; run Compact database |
| Compaction is slow | VACUUM on a large database takes time; do not click repeatedly |
Release build has no app-<date>.log |
By design; use the Interaction Monitor and MCP logs |
| Logs contain secrets | Request bodies include prompts and tool results; clean before exporting/sharing |
| Mission monitor has no missions | No sub-agent ran in this session, or orchestration fell back |