openclaw - 💡(How to fix) Fix Runtime parity needs 20-turn and 100-turn soak coverage [1 comments, 2 participants]

Part of #80171 and follow-up to #80323.

The Phase 1-5 harness can compare Pi and Codex cells, but current runtime-pair depth is mostly short-turn coverage.

Audit snapshot from #80323:

• Across the 97 parsed QA markdown scenarios, the maximum runAgentPrompt count in a scenario is 4.
• Most scenarios are 0-1 direct agent prompts; only 10 scenarios have 2+ direct agent prompts.
• Curated JSONL replay fixtures currently cover 2, 2, and 3 user turns respectively.
• There is no 20-turn or 100-turn Codex-vs-Pi runtime soak lane.

Why this matters:

The known Codex-vs-Pi risk class is not only single-turn shape drift. First-hour usage can include repeated follow-ups, compaction pressure, tool-result replay, memory recall, subagent delivery, plugin availability after install/update, and auth/profile reuse. A short lane will catch immediate tool-call shape drift but can miss delayed runtime drift across accumulated session state.

Acceptance sketch:

• Add a deterministic 20-turn runtime-pair scenario that mixes read-only tools, one safe write/edit, follow-up references, memory recall, and at least one subagent/tool-result continuation.
• Add a slower 100-turn soak tier that can run outside the default PR gate but is available for release/testbox/scheduled validation.
• Capture first drift turn and per-turn tool/usage deltas in the runtime parity report.
• Keep the 100-turn lane optional or scheduled so release PR latency stays bounded.