feat(messaging): mirror local-surface turns (Desktop/TUI/dashboard) back to the platform chat

[Maziyang2]

Why this matters

Keep the record complete when you continue a messaging conversation from the Hermes Desktop app (or TUI / dashboard).

Hermes sessions are cross-surface: you can reopen your Telegram DM in the Desktop app and keep chatting. But turns served by a local surface exist only on that surface — the platform's own chat never sees them, so the history on your phone has a hole exactly where you kept talking (and the bot cannot replay your side — it can only post as itself).

This PR adds an opt-in mirror: every completed turn served locally for a gateway-owned session is forwarded to that session's own platform chat — your message as a quote block, then the assistant's reply:

📲

> the message you sent from the Desktop app
> (quoted line by line)

the assistant's reply

• Off by default — display.platforms.<platform>.mirror_local_turns: true enables it per platform (Telegram shown; any platform whose plugin exposes a standalone sender works).
• Exact target — delivered to the chat_id / thread_id stored on the session row: the same chat the gateway itself delivers to.
• No gateway connection needed — uses the same standalone sender chain as hermes send / cron delivery; failures are logged and never affect the turn.
• Quiet by default — sent with disable_notification (configurable), so a phone that already followed along on the desktop isn't pinged twice.
• Never noisy — synthetic turns (compaction notices, auto-continue, background-process notes) and silent ([SILENT]) responses are skipped.

Companion PR — a two-part messaging upgrade

This is the second half of a coherent messaging/Telegram upgrade; both PRs are designed to land and be used together:

	PR	What it adds
1	#110611	single-message streaming for long replies + Bot API 10.3 extras (stop button, message effects, ephemeral notices) — each with an opt-in switch
2	this PR	cross-surface continuity — local-surface turns are mirrored to the platform chat, so the record stays complete

They are split into two PRs only for compatibility / independent reviewability — neither depends on the other's code, and they can merge in any order. The long-text handling here deliberately follows the same parameter semantics introduced in #110611 (streaming_single_message_4096_split):

• mirror_local_turns_4096_split: false (default) — keep one message where possible; past the platform cap the remainder paginates (page markers kept)
• mirror_local_turns_4096_split: true — split into standalone ≤4096 messages (markers stripped)

so both features behave consistently when enabled together.

What changed

• tui_gateway/turn_mirror.py (new) — decision point, message composition, detached delivery (daemon thread; never blocks or breaks the turn it mirrors).
• tui_gateway/prompt_turn.py — dispatch after message.complete (no-op unless enabled). One hook covers Desktop + TUI + dashboard, which share this backend.
• gateway/display_config.py — mirror_local_turns, mirror_local_turns_silent, mirror_local_turns_4096_split, mirror_local_turns_label (+ normalisers).
• tools/send_message_senders.py — silent + chunk_indicators kwargs on the standalone Telegram sender; _registry_standalone_send forwards them only to senders that declare them (other platforms keep their existing call shape).
• plugins/platforms/telegram/adapter.py — _standalone_send pass-through.
• Docs (website/docs/user-guide/messaging/index.md), cli-config.yaml.example, tests.

How to test

bash scripts/run_tests.sh tests/tui_gateway/test_turn_mirror.py tests/gateway/test_display_config.py \
  tests/tools/test_send_message_tool.py tests/tui_gateway/test_auto_continue.py

• New suite tests/tui_gateway/test_turn_mirror.py — 24 passed (composition, decision table, target resolution, delivery options, contained failures).
• Scoped suites across display config / senders / tui_gateway server+contracts / caption / proxy / relay — 990 passed, 0 failed, 3 skipped.
• Live check: set display.platforms.telegram.mirror_local_turns: true, continue a Telegram session from the Desktop app — each completed turn arrives in the Telegram chat as quote + reply.

Tested on macOS (Darwin 27), Python 3.11.

Type of Change

• ✨ New feature (non-breaking; opt-in, default off)

Checklist

Code

• I've read the Contributing Guide
• Commit messages follow Conventional Commits
• Searched for existing PRs — no duplicate; companion to feat(telegram): single-message streaming + Bot API 10.3 extras (stop button, effects, ephemerals), all opt-in (#110564) #110611 (independent; split for reviewability)
• Only changes related to this feature
• Tests added for the changes (scoped suites above; CI covers the full matrix)
• Tested on my platform: macOS (Darwin 27)

Documentation & Housekeeping

• Docs updated (messaging guide + cli-config.yaml.example)
• Cross-platform impact considered — generic by design; Telegram path exercised; other platforms degrade gracefully (silent/chunk_indicators forwarded only when the sender declares them)
• Additive, default off — no behavior change without config

[Maziyang2]

Follow-up commit 80d0c24b48 adds reliable delivery for the opt-in local-turn mirror path.

Why this matters

A Desktop/TUI/dashboard turn should not disappear from the platform chat because the standalone send raced a process exit or gateway restart. This keeps the mirror asynchronous while making the delivery obligation durable and recoverable.

What changed

• Records a pending obligation before spawning the mirror sender.
• Uses the turn request id for atomic deduplication; an already-recorded turn is not sent again.
• Persists the transport-owning profile and binds both ledger access and standalone credentials to that profile.
• Preserves adapter send metadata across recovery and records the final platform message id.
• Reuses the existing bounded retry/recovery sweeps; pending/attempting/failed rows recover after restart with the existing duplicate-warning markers.
• Leaves the existing opt-in flag and the feat(telegram): single-message streaming + Bot API 10.3 extras (stop button, effects, ephemerals), all opt-in (#110564) #110611 Telegram streaming work unchanged.

Verification

• 118 focused gateway/ledger/mirror tests passed, plus the real two-profile state-store scope test.
• ruff, compileall, git diff --check, and compat-pointer checks passed.
• No live Telegram send/edit/stop test was run.

This is intentionally a focused follow-up commit on #111631 rather than adding more scope to #110611.