Three code paths constructed temporary AIAgent instances without
skip_memory=True, causing the Honcho memory provider to initialise
and create empty sessions that were never written to:

1. gateway/run.py — hygiene auto-compress and /compress temp agents
   lacked skip_memory=True (flush agent at line 717 already had it)
2. plugins/memory/honcho/__init__.py — per-session strategy ran
   migrate_memory_files on every new session, uploading MEMORY.md /
   USER.md / SOUL.md into short-lived sessions that were immediately
   abandoned after session rotation
3. run_agent.py — _spawn_background_review created a full AIAgent to
   review conversation history for local memory extraction, but did
   not need Honcho (it uses the shared _memory_store directly)

All three now pass skip_memory=True or short-circuit appropriately,
consistent with the existing convention across 20+ call sites.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…isolation

The plugin architecture refactor (924bc67, 2026-04-02) removed the
honcho_session_key parameter from AIAgent, breaking the gateway →
Honcho session key plumbing. This caused all Honcho sessions to use
bare timestamp IDs (e.g. 20260412_171002_xxx) instead of the stable
per-chat key (e.g. agent-main-telegram-dm-8439114563), fragmenting
conversation history across disposable sessions.

Fix: thread gateway_session_key through the full path:
1. gateway/run.py — pass gateway_session_key=session_key to AIAgent
2. run_agent.py — accept, store, and forward via _init_kwargs
3. plugins/memory/honcho/__init__.py — extract and pass to resolver
4. plugins/memory/honcho/client.py — add gateway_session_key as a
   new priority level in resolve_session_name (after /title override,
   before per-session fallback)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…supplement, trivial skip

Inspired by claude-honcho's architecture. Fixes the parroting problem
where dialectic output was the sole injection source, causing stale
observations and roleplaying responses in auto-injected context.

New injection architecture:
  - Base layer: peer.context() result (representation + card), cached
    with 5-minute TTL. Stable, cheap, always available.
  - Supplement: dialectic result fires every N turns (cadence-gated),
    cached until next refresh. Adds synthesis when fresh.
  - Trivial prompt skip: "ok", "yes", "continue", slash commands, and
    empty prompts skip injection entirely. Zero token waste.

Session start pre-warm now:
  - Fetches peer.context() synchronously and caches it
  - Fires dialectic with targeted prompt ("Summarize what you know about
    this user. Focus on preferences, current projects, and working style.")
    instead of generic "What should I know about this user?"

The dialectic is no longer the sole injection source. It supplements
the stable representation. honcho_reasoning tool remains the only way
to get direct dialectic output on demand.

- Skip dialectic at session start when peer has no context (empty
  representation + card = new peer). Avoids generating slop from
  nothing.
- Change warm prompt from "Focus on preferences, current projects,
  and working style" to "Focus on preferences, goals, and style" —
  Hermes isn't only a work assistant.