phi-prompt-inspect

name	phi-prompt-inspect
description	Pull the exact rendered system + user prompts phi worked from on a past agent run, via the logfire MCP. Use when investigating phi's behavior — why she said that, what her context actually contained, is a block stale or feedback-looping. Read-only, no PDS writes.

phi is instrumented by pydantic-ai → logfire. every agent.run() writes an agent run span whose attributes.pydantic_ai.all_messages carries the full conversation as sent to the model: system prompt with each dynamic block as a separate text part, then user / assistant / tool turns.

the two queries

Both run via mcp__logfire__query_run(project="phi", query=..., start_timestamp=..., end_timestamp=...). Project is always required. Default window is 30min; max 14d. Narrow it aggressively — phi fires an agent run every ~10s (notification poll), so data is dense.

Step 1 — find the span behind a specific behavior. final_result is phi's own brief string summary of what she did on that run; that's the easiest way to pick the right one out of a dense window:

SELECT span_id, trace_id, start_timestamp,
       attributes->>'final_result' AS final_result
FROM records
WHERE span_name = 'agent run'
  AND attributes->>'gen_ai.agent.name' = 'phi'
ORDER BY start_timestamp DESC
LIMIT 5

Filter on gen_ai.agent.name = 'phi' to exclude sub-agents — phi-inner-critic, phi-extractor, observation-reconciler each fire their own agent run spans with short purpose-built prompts. Useful when you want them, distracting when you don't.

Step 2 — pull the rendered prompt once you have a span_id:

SELECT attributes->'pydantic_ai.all_messages' AS messages
FROM records
WHERE span_id = '<span_id>'
LIMIT 1

structure of pydantic_ai.all_messages

It's an array of messages:

• [0] = system — parts is the array of dynamic blocks in registration order. parts[0] is the static base (personalities/phi.md + cross-cutting operational rules); the rest are the dynamic blocks: [YOUR INFRASTRUCTURE], [OPERATOR], [NOW], [OPERATIONAL HISTORY], [KNOWN RELAYS], [GOALS AND INTERESTS] + [SELF-AWARENESS] + [SELF STATE], [RECENT OPERATIONS], [DISCOVERY POOL], [NEW NOTIFICATIONS], per-author memory ([PHI'S SYNTHESIZED IMPRESSION OF @h] / [OBSERVATIONS ABOUT @h] / [PAST EXCHANGES WITH @h] / [BACKGROUND RESEARCH ON @h]), [RELEVANT MEMORIES — synthesized for this query], [ATLAS], [DOCKET], [OWNED FEEDS], [SEMBLE]. See docs/system-prompt.md for what each one is for.
• [1] = user — the path-specific task prompt (notifications batch / cycle / reflection) plus any appended blocks: [WORKFLOW STATE], [RECENT FLOW MENTIONS], [RECENT CONVERSATIONS], [FIRST INTERACTION WITH @h], [SERVICE HEALTH].
• [2..] = assistant + tool roundtrips inside the agent loop. Inspect these to see which tools phi actually called and with what arguments.

what to look for

• Personality dilution. parts[0] should contain phi.md verbatim. Compare against the file.
• Feedback loops. Several dynamic blocks reflect phi's own past output back at her — [SELF-AWARENESS] (haiku description of recent posts), [RECENT OPERATIONS] (verbatim recent writes), [DOCKET] titles (often derived from past observations). If they're all in a register phi.md tells phi not to use, the model sees that register reinforced even when the constitution forbids it.
• Block bloat. Compare token weight of abstract rules (top of parts[0]) against concrete current blocks. Concrete + current usually dominates abstract + general.
• Tool roundtrips in [2..]. Watch for recall / search_memory calls and what was returned — sometimes phi's drift comes from what private memory surfaced.

gotchas

• agent_name is also present in attributes (pydantic-ai's own), but gen_ai.agent.name is the OTel-standard key the sub-agents also set — filter on that.
• A trace can span multiple agent run spans (the main run plus sub-agents triggered during it). trace_id groups them; span_id identifies one.
• The JSON column accepts -> for object access and ->> for text extraction. Array index access (->0) and chaining (->'foo'->0->'bar') work in DataFusion just like Postgres.