self-traces

name	self-traces
description	Query your own execution traces (logfire) to answer "what did I actually do" — every tool call, argument, error, and silent failure across all your runs. Load before using query_traces. Use for postmortems, retro receipts, auditing your own trading discipline, and noticing failures your runs never surfaced. Not for deciding what to post.

your traces are the ground truth of your own behavior. your memory holds what you chose to write down; your PDS holds what you published; the traces hold what you actually did — including everything that failed silently. when you misremember (a wrong handle, a tool you're sure you called, a trade you think you pre-registered), the trace settles it.

the actual situation — read this before querying

• query_traces(sql, start, end=None) hits a real analytics database holding weeks of your history, millions of rows. an unbounded query is a firehose that will blow out your context. move carefully: every query gets a tight start/end window, a LIMIT, and only the columns you need.
• the time window comes from the start/end tool arguments (ISO 8601), not from SQL — a WHERE start_timestamp > ... alone does not bound the scan.
• the columns this skill names below are the ones that matter; they exist. don't guess at others — select what you see here.
• this is read-only. you cannot break anything; you can only waste context.

span shapes that matter

• span_name = 'running tool' — one row per tool call you made. attributes->>'gen_ai.tool.name' is the tool, attributes->>'tool_arguments' is the JSON args you passed.
• span_name = 'agent run' — one row per run of you (a batch, a cycle, a scheduled pass).
• is_exception = true — things that broke. exception_type, exception_message. many of these were swallowed so the run could continue; you never saw them at the time. this is where your blind spots live.
• trace_id groups one run's spans; filter on it to reconstruct a single incident end to end.

recipes

what did I do in a given run (start from the tool call you remember):

SELECT start_timestamp, attributes->>'gen_ai.tool.name' AS tool,
       left(attributes->>'tool_arguments', 200) AS args
FROM records WHERE span_name = 'running tool'
ORDER BY start_timestamp LIMIT 50

what failed on me lately (weekly hygiene, or when something feels off):

SELECT exception_type, left(exception_message, 150) AS msg, count(*) AS n
FROM records WHERE is_exception
GROUP BY 1, 2 ORDER BY n DESC LIMIT 20

receipts for a claim about yourself (tool mix over a period — pass a month-long window via timestamps):

SELECT attributes->>'gen_ai.tool.name' AS tool, count(*) AS n
FROM records WHERE span_name = 'running tool'
GROUP BY 1 ORDER BY n DESC LIMIT 30

discipline

traces are for postmortems, retro receipts, and audits — answering "what happened", "why did I do that", "is this claim about myself true". they are not an input for deciding what to post or trade next; reading your own reasoning back in ordinary cycles is a mirror, and you already know where mirrors lead. cite what you find the way you cite any incident: timestamp, what the trace shows, what you concluded.