Prompt Logging

Talemate can write every prompt it sends to a language model — together with the model's response and a bundle of metadata — to a JSON Lines file on disk. This is intended for debugging prompt issues, comparing behaviour across clients, or feeding collected prompts into offline analysis.

Prompt logging is off by default and is enabled with a single environment variable.

Enabling

Set TALEMATE_LOG_PROMPTS=1 before starting the server. Any truthy value enables logging (Talemate checks for the variable being set to a non-empty string).

Linux

Prefix the start command:

TALEMATE_LOG_PROMPTS=1 ./start.sh

Or when running manually:

TALEMATE_LOG_PROMPTS=1 uv run src/talemate/server/run.py runserver --host 0.0.0.0 --port 5050

Windows

SET TALEMATE_LOG_PROMPTS=1
start.bat

Disabling

Unset the variable (or set it to an empty string) and restart Talemate:

Linux

unset TALEMATE_LOG_PROMPTS
./start.sh

Windows

SET TALEMATE_LOG_PROMPTS=
start.bat

Output file

Setting	Value
Path	logs/prompt_log.jsonl in the Talemate project root
Format	JSON Lines (one JSON object per line)
Write mode	Append — never truncated or rotated
Flush	Every record is flushed immediately, so partial runs are not lost

The file is opened the first time a prompt is logged after startup and kept open for the lifetime of the process. It is never rotated or truncated by Talemate, so the file will keep growing as long as the variable is set. If you only need a short capture, enable the variable, reproduce the problem, disable it, and delete or move the file afterwards.

logs/prompt_log.jsonl vs logs/prompt_log.json

The JSON-Lines file described here is written by the server while it runs.

The similarly named logs/prompt_log.json is a separate, one-shot snapshot produced by the Export button in the Debug Tools → Prompts tab and is independent of this environment variable.

Record schema

Each line in prompt_log.jsonl is a single JSON object with the fields below. Field order inside the object is not guaranteed.

Field	Type	Description
kind	string	Prompt kind identifier (e.g. conversation, narrate_scene, summarize). Corresponds to the template/system-prompt kind used for the call.
prompt	string	The full finalized prompt text sent to the model, after template rendering and any client-side formatting.
response	string	The full model response text, after stop-string trimming and smart-quote normalization.
prompt_tokens	int	Prompt token count. Uses the client's own counter unless the backend returned an explicit prompt-token count, in which case that value is preferred.
response_tokens	int	Response token count. Uses the backend-reported count when available, otherwise the client's tokenizer.
client_name	string	Name of the client that produced the prompt (as configured on the Clients screen).
client_type	string	Client type identifier (e.g. openai, anthropic, koboldcpp).
time	number	Wall-clock seconds spent on the generation, measured around the backend call.
agent_stack	list of strings	The agent call stack at the time of the prompt, outermost first. The final entry is the agent that actually issued the call (e.g. ["director", "conversation"]). Empty if no agent context was active.
generation_parameters	object	Final generation parameters passed to the backend for this call (temperature, top-p, max tokens, etc. — contents vary by client type).
inference_preset	string or null	Name of the active inference preset, if any.
preset_group	string or null	Preset group the client is using, if any.
reasoning	string or null	The extracted reasoning / thinking trace for this response, when the client supports reasoning tokens.
template_uid	string or null	UID of the Jinja prompt template that produced the prompt. Useful for correlating a log line back to a specific template in the Prompt Manager.

The record is the same PromptData structure that Talemate emits over the websocket to populate the in-app Debug Tools Prompts tab, so anything visible there is also in the log file.

Quick inspection

Because each line is a self-contained JSON object, the file works well with standard JSON tools. A few examples:

Pretty-print the last prompt:

tail -n 1 logs/prompt_log.jsonl | jq

Count prompts per agent (top of the stack):

jq -r '.agent_stack[-1] // "none"' logs/prompt_log.jsonl | sort | uniq -c

Extract only the prompts that took longer than five seconds:

jq 'select(.time > 5)' logs/prompt_log.jsonl

Related

• Debug Logging — enable DEBUG-level logging and error-log file output with TALEMATE_DEBUG=1.
• Debug Tools › Prompts — in-app viewer for the same prompt records, with a one-shot export button.