Add multi-turn-consistency-check skill for grounding corrected long threads

[hanyueqiang]

Summary

This adds multi-turn-consistency-check, a skill for long multi-turn threads where earlier assumptions can leak into later answers.

The skill guides Codex to:

• rebuild the current state from the latest user instructions and fresh evidence
• keep a compact internal fact ledger
• detect conflicts across turns such as changed requirements, stale plans, and mismatched names, IDs, paths, or dates
• separate verified facts from inference before concluding

Why

Long coding, debugging, review, and planning threads often accumulate stale assumptions after corrections or scope changes.

This skill packages a narrow workflow for those cases. It does not try to promise perfect hallucination prevention. Instead, it reduces factual drift and unsupported carry-over by making the agent:

• re-check the active state
• call out conflicts explicitly
• verify unstable claims before answering
• leave unresolved items unresolved instead of guessing

Scope

This version is intentionally single-mode:

• it performs an internal consistency pass at answer time
• it does not create or maintain a persistent markdown log by default
• it only writes a durable note when the user explicitly asks for one or the task clearly requires a handoff

The skill is also configured for explicit use with allow_implicit_invocation: false to keep the scope narrow for an initial contribution.

Files

• skills/.curated/multi-turn-consistency-check/SKILL.md
• skills/.curated/multi-turn-consistency-check/references/fact-ledger.md
• skills/.curated/multi-turn-consistency-check/references/conflict-signals.md
• skills/.curated/multi-turn-consistency-check/agents/openai.yaml

Notes

If maintainers think this fits better in .experimental first, I am happy to move it.