Ensure a system message always precedes the first user message in LLM requests

Summary

This is a tracking issue for a repo-wide invariant: every LLM request built in this repository must place a system message before the first user message.

The main agent loop already enforces this via Agent.init_state() (openhands-sdk/openhands/sdk/agent/agent.py) — it guarantees the SystemPromptEvent sits at index 0/1 and raises AssertionError if a user message appears before it. Standalone LLM calls (condenser, goal judge, pre-flight pings, security analyzers) do not all follow the same rule today.

Why it matters

• Consistent model steering: standalone calls that omit a system message lose the shared context/format expectations other paths get.
• Correct transport serialization: on the OpenAI Responses API, Message.to_responses_value() returns a string for system (→ instructions) and a dict list for user (→ input). Sending steering instructions as a lone user message conflates "how" with "what", and a lone system message would serialize to instructions with empty input.
• Subscription/Codex transport prepends system chunks onto the first user message; a missing system lead-in produces a bare, unsteered request.

Subissues

• #5142 — Condenser summarization: split into system + user (PR #5143)
• #5145 — Goal judge: split judge prompt into system + user messages
• #5146 — Agent-server profile pre-flight ping: prepend a system message
• #5147 — GraySwan analyzer: guarantee system message first when history window drops the SystemPromptEvent
• #5148 — Condenser: ensure the SystemPromptEvent is never forgotten (keep_first=0 / summary-first edge case)
• #5149 — Condenser hard context reset: preserve the leading system prompt instead of summarizing it at offset 0
• #5150 — Document the invariant (this pattern, as a repo convention)

Each subissue must add (or verify) a system message before the first user message in the request it builds.

This issue was created by an AI agent (OpenHands) on behalf of juanmichelini.

Start with Agent.init_state() in openhands-sdk/openhands/sdk/agent/agent.py and review the linked subissues #5145–#5150 for the standalone request paths they identify. Check each condenser, goal judge, pre-flight, and GraySwan request, plus the documentation task, to verify a system message precedes the first user message and that relevant tests pass.