Fix/honcho local mode gateway session

[adavyas]

What does this PR do?

This PR is a focused follow-up to pr-736 that corrects two runtime integration issues in the current Honcho changes.

First, it enforces memoryMode=local at runtime. In the current PR branch, local mode is described as disabling Honcho, but Hermes can still initialize remote Honcho state, create sessions, and prewarm memory paths. This change makes local mode behave consistently with that contract when both active peers resolve to local.

Second, it fixes gateway ownership of session-scoped Honcho state. The gateway currently creates short-lived AIAgent instances per request, but writeFrequency="session" requires a Honcho manager that survives across requests within the same gateway session. This PR moves that ownership to the gateway session layer so the same Honcho manager is reused across messages and flushed at real session boundaries.

This change is intentionally narrow. It does not add new configuration, change public schemas, or bundle unrelated cleanup. It keeps the PR aligned with the Hermes contributing guidance by focusing on one logical bug-fix area: making the current Honcho integration behave correctly.

Related Issue

Follow-up to #736

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Security fix
• Documentation update
• Tests (adding or improving test coverage)
• Refactor (no behavior change)
• New skill (bundled or hub)

Changes Made

• Updated run_agent.py to:
  • prevent remote Honcho activation when both active peers are configured as local
  • allow the gateway to inject an existing Honcho manager and config instead of always creating a new one
  • keep the first-turn Honcho warmup path compatible with prompt-prefix caching by avoiding dynamic dialectic prewarm in the cached prompt path
• Updated gateway/run.py to:
  • persist Honcho managers per gateway session key
  • reuse those managers across short-lived request handlers in the same session
  • shut them down at actual gateway session boundaries rather than at request boundaries
• Added regression tests in tests/test_run_agent.py covering:
  • memoryMode=local skipping Honcho activation
  • injected Honcho managers bypassing fresh client initialization
• Added regression tests in tests/gateway/test_honcho_lifecycle.py covering:
  • gateway Honcho manager reuse across requests
  • local mode skipping gateway Honcho manager creation
  • session reset shutting down the persistent gateway Honcho manager

How to Test

1. Configure Hermes with Honcho enabled and set both active peers to memoryMode=local.
2. Run a Hermes session and verify that remote Honcho is not initialized or prewarmed.
3. Start the gateway and send multiple messages in the same session.
4. Verify that the same Honcho manager is reused across requests for that session.
5. Reset or end the session and verify that the persistent Honcho manager is shut down at that boundary.

Validation performed for this branch:

1. pytest tests/test_run_agent.py -q
2. pytest tests/gateway/test_honcho_lifecycle.py -q
3. pytest tests/ -q
4. python -m hermes_cli.main doctor
5. python -m hermes_cli.main config check
6. python -m hermes_cli.main chat -q "test message"

Test result summary:

• pytest tests/ -q -> 2692 passed, 5 skipped, 23 deselected

Checklist

Documentation & Housekeeping

• N/A: no user-facing documentation changes were required
• N/A: no config keys changed, so cli-config.yaml.example did not require updates
• N/A: no contributor workflow or architecture guide changes were required
• I've considered cross-platform impact per the compatibility guide
• N/A: no tool schema or description changes were required

Screenshots / Logs

Commits included:

• 7ffaaa0 fix(honcho): enforce local mode and cache-safe warmup
• 353a9e7 fix(gateway): persist Honcho managers across session requests

[erosika]

Consolidated into feat/honcho-async-memory — cherry-picked both fix commits (7ffaaa0, 353a9e7) on top of the rebased feature branch. This PR can be closed once the feature branch lands.

[teknium1]

Closing this PR — the branch is 258 commits behind main, and merging it would delete ~27,000 lines of current code (skin engine, email gateway, skills, dozens of test files, etc.).

Additionally, the parent PR #736 (Honcho integration) is still open and hasn't been merged, so there's no Honcho code on main to fix yet.

If these Honcho fixes are genuinely needed, they should be applied to PR #736's branch directly rather than submitted as a separate PR targeting main. Thanks for the contribution!