feat: Add OpenViking auto-commit feature for early memory searchability

[happy5318]

Description

Adds automatic commit functionality for OpenViking memory provider, solving the pain point that users have to wait for 2-hour session timeout to search their conversation history.

Changes

1. CLI: Trigger when executing or commands, so the current session is committed immediately to OpenViking
2. Gateway: Add background idle commit watcher that checks every 60 seconds, commits sessions that have been idle for 120 seconds (configurable)
3. Session: Add field to avoid duplicate commits, reset flag when new message arrives

Compatibility

• ✅ Fully backward compatible: No impact on users not using OpenViking memory provider
• ✅ Exception safe: All commit logic wrapped in try-except blocks, won't break main conversation flow even if OpenViking is unavailable
• ✅ Cross-platform: Works with all messaging platforms (Feishu, Discord, Telegram, CLI)

Usage

After enabling OpenViking memory provider, users can search their conversation history:

• Immediately after executing or
• Automatically 2 minutes after stopping conversation
• Or after 2-hour session timeout (existing logic)

[ZaynJarvis]

Review: Auto-Commit Feature (Idle Commit Watcher)

Verdict: Request Changes — blocking issues before this can merge

The goal (committing sessions early for searchability) is good, but the implementation has several problems that need addressing.

Blocking Issues

1. Synchronous requests inside an async coroutine (critical)

_idle_commit_watcher() calls requests.post() directly inside an async def function. This blocks the asyncio event loop for up to 10 seconds per session per cycle, freezing all gateway activity. This must use httpx.AsyncClient with await client.post() instead:

# Replace this:
import requests
_resp = requests.post(f"{_endpoint}/api/v1/sessions/{entry.session_id}/commit", ...)

# With this:
async with httpx.AsyncClient() as client:
    _resp = await client.post(f"{_endpoint}/api/v1/sessions/{entry.session_id}/commit", ...)

2. Hardcoded endpoint

The watcher hardcodes http://127.0.0.1:1933. The endpoint should come from os.environ.get("OPENVIKING_ENDPOINT", _DEFAULT_ENDPOINT) or the configured plugin endpoint — otherwise this breaks any non-default deployment.

3. cli.py change calls shutdown_memory_provider() on /new

shutdown_memory_provider() calls the provider's shutdown() method, which joins threads and clears _last_active_provider. After this, the provider is dead — it cannot receive sync_turn calls for the new session. The correct approach (as in #9167) is to call commit_memory_session() (commit only) then reinitialize_memory_session() (re-bind to new session). Using shutdown_memory_provider() here corrupts the session lifecycle for the new session.

4. Direct session_store._entries access

The watcher accesses self.session_store._entries directly, bypassing the _ensure_loaded()/lock machinery in the public API. Use the public SessionStore methods or at minimum hold _lock during the full iteration.

5. No tests

The idle commit watcher is non-trivial async logic. It needs at minimum: (a) test that watcher commits idle sessions, (b) test that memory_committed=True is set, (c) test that active sessions are skipped.

What Is Good

• The concept of an idle commit watcher is sound and complementary to the explicit commit hooks in #9167/#7762
• memory_committed field on SessionEntry is clean
• Resetting memory_committed on new activity is correct

Please rewrite the watcher to use async httpx and fix the endpoint sourcing, then rebase onto #9167 so the cli.py change can be removed (it's handled there).

[ZaynJarvis]

plus, i think it's not necessary to auto commit since /new /reset /compress hook would commit the session #10463. If users are in the session, they can just use the in session context. I would probably suggest close this PR.

[happy5318]

Thanks for the review @ZaynJarvis! The issues you've raised have been addressed in the latest commit:

1. Using httpx.AsyncClient instead of synchronous requests - async non-blocking
2. Reading endpoint from OPENVIKING_ENDPOINT environment variable
3. Added lock protection with self.session_store._lock when accessing _entries

The implementation is already using httpx.AsyncClient with await, reading config from env vars, and protecting with locks. Could you please re-review? Thanks!

[teknium1]

Thanks for the contribution, @happy5318! After reviewing against current main, the core pain point this PR addresses — OpenViking sessions not being committed on /new and /compress — was already fixed by PR #10463 (commit 7856d304f).

This is an automated hermes-sweeper review.

Evidence:

• cli.py:4943 — commit_memory_session() is called in new_session(), covering /new and /reset (added by commit 7856d304f, PR fix(openviking): correct account default + commit session on /new and compress #10463)
• run_agent.py:9084 — same call inside _compress_context(), covering /compress and auto-compress
• gateway/run.py:6680 — _handle_reset_command() calls _cleanup_agent_resources() → shutdown_memory_provider() → on_session_end() for gateway /new//reset
• gateway/run.py:2996 — _session_expiry_watcher() already commits sessions on expiry/idle eviction (1h TTL)
• @ZaynJarvis (author of the merged fix(openviking): correct account default + commit session on /new and compress #10463) also recommended closing this PR for the same reason

Additionally, 108 of the 112 changed files are compiled/binary dist artifacts (ui-tui/dist/, hermes_cli/web_dist/fonts/) that shouldn't land on main.

If there's interest in the narrower 120-second idle-commit optimization specifically for gateway sessions, that would be worth a small focused PR without the dist artifacts.