RFC: live-to-final replies for long-running sessions

[franksong2702]

Thinking Path

• Redesign live-to-final assistant replies for running agent sessions #3400 is an umbrella product discussion for redesigning live-to-final assistant replies.
• Long-running sessions exercise the hardest states: long waits, many tool calls, long final answers, compression, retry/tool limits, reconnects, and user intervention.
• The repo needs a public RFC that defines product semantics before follow-up implementation slices expand beyond the first lifecycle work.
• This PR keeps the change docs-only and avoids coupling the RFC identity to one implementation PR.

Refs #3400.

What Changed

• Added docs/rfcs/live-to-final-assistant-replies.md.
• Added the RFC to docs/rfcs/README.md.
• Added the RFC to the runtime/durability index in docs/CONTRACTS.md.
• Defined goals, non-goals, representative public issue signals, live/settled/recovery/terminal semantics, long-running edge cases, delivery slices, review checklist, related contracts, and open questions.

Why It Matters

This gives #3400 a neutral product-semantics anchor for long-running sessions instead of treating #3401 as the whole solution. It also gives future PRs a review checklist for Auto Compression, no-final outcomes, tool/retry ceilings, replay/reconnect, very long final answers, and queue/steer/interrupt adjacency.

Contract Routing

• Contract family: RFC / runtime-durability / product semantics for live assistant reply rendering.
• Touched contracts: docs/rfcs/live-to-final-assistant-replies.md, docs/rfcs/README.md, docs/CONTRACTS.md.
• Evidence used: CONTRIBUTING.md, docs/CONTRACTS.md, existing RFC index shape, and public Redesign live-to-final assistant replies for running agent sessions #3400 issue context.
• Contract change type: new proposed RFC; no runtime behavior change.

Verification

• git -C /Users/xuefusong/hermes-webui-worktrees/3400-live-to-final-rfc diff --cached --check
• git -C /Users/xuefusong/hermes-webui-worktrees/3400-live-to-final-rfc show --check --stat --oneline HEAD
• Grepped the RFC/index files for private/local context and implementation-only terms; no matches.
• Checked referenced public GitHub issue/PR links; all returned 200 after retrying transient curl connection failures.
• Not run: pytest, because this is a docs-only RFC PR with no runtime code changes.

Risks / Follow-ups

• This RFC is Proposed; implementation remains split across Redesign live-to-final assistant replies #3401 and follow-up slices.
• Open questions remain around very long final-answer affordances, repeated compression display, and whether queue/steer/interrupt should receive a dedicated public control-surface RFC.

Model Used

AI-assisted: OpenAI GPT-5 Codex via Codex desktop. A user-provided Claude Code read-only review was evaluated before finalizing the RFC wording.