phi-prompt-inspect

Use query_traces(sql, start, end) with a narrow ISO 8601 time window and a LIMIT. The tool caps output at 6,000 characters. Inspect one field in slices; requesting an entire conversation or all attributes will truncate it.

Find the trace behind the behavior using self-traces if its identity is unknown. Polling does not imply an agent run occurred. A trace may contain Phi and several supporting agents; identify the model request you intend to inspect rather than assuming the first span belongs to Phi.

For a known trace, list model requests:

SELECT start_timestamp, span_id, span_name,
       attributes->>'gen_ai.usage.input_tokens' AS input_tokens
FROM records
WHERE trace_id = '<trace_id>' AND span_name LIKE 'chat %'
ORDER BY start_timestamp LIMIT 30

For the chosen span, measure the logged fields before reading them:

SELECT length(attributes->>'gen_ai.system_instructions') AS instruction_chars,
       length(attributes->>'gen_ai.input.messages') AS input_chars,
       length(attributes->>'gen_ai.tool.definitions') AS tool_chars
FROM records WHERE span_id = '<span_id>' LIMIT 1

Read a bounded slice of one field; advance the offset only if needed:

SELECT substring(attributes->>'gen_ai.system_instructions', 1, 3500) AS excerpt
FROM records WHERE span_id = '<span_id>' LIMIT 1

Use the same pattern for gen_ai.input.messages (task and preceding turns) and gen_ai.tool.definitions (exposed names, descriptions, schemas). Offsets start at 1. The text is serialized JSON; a slice can split a JSON value. Compare the first request with a later request when investigating tool-result growth. A listed tool was available; only a tool-call span proves it was called.

These fields were verified against September 2026 model-request spans. Older agent spans may use pydantic_ai.all_messages; inspect the actual shape rather than assuming fixed system/user array positions. Null fields mean absent telemetry, not absent instructions. Scrubbed values and truncated output limit what you can conclude. A trace summary is not proof a claimed write succeeded.

Report the trace/span, relevant text, and the inference separately. Provider input totals include cached subsets; do not add cache reads/writes again.

/diagnostic is a fresh scheduled-context preview. It omits the entry-point task, notification-specific material, and exposed tool schemas; it is useful for comparison but cannot reconstruct a past run. docs/system-prompt.md maps blocks to their producers. Inspect a suspicious producer or tool contract through own-source rather than turning a trace inference into a new rule.