PR #23346: refactor(cron): DRY media extensions from gateway.platforms.base

• Repository: NousResearch/hermes-agent
• Author: mehmetkr-31
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/23346

Description (problem / solution / changelog)

Moves _VIDEO_EXTS and _IMAGE_EXTS to the canonical location in gateway/platforms/base.py and imports them from cron/scheduler.py instead of duplicating. This prevents drift when new formats are added (e.g. .heic images) and removes the "keep in sync" comment hazard.

Related to #23324.

Changed files

• cron/scheduler.py (modified, +3/-4)
• gateway/platforms/base.py (modified, +3/-3)
• scripts/release.py (modified, +1/-0)

---

PR #23510: fix(agent): cap auxiliary LLM concurrency per task

• Repository: NousResearch/hermes-agent
• Author: konsisumer
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/23510

Description (problem / solution / changelog)

Cap concurrent in-flight auxiliary LLM calls per task to prevent retry/fallback storms from amplifying provider incidents.

What changed and why

• agent.auxiliary_client.call_llm and async_call_llm now read auxiliary.<task>.max_concurrency from config and acquire a per-task semaphore around the entire call (including retries and provider fallback). Background work like title generation and context compression can spawn one call per active session — during provider degradation those calls fan out across retries and trip downstream rate limits, so the cap stops the amplification.
• Sync limits use threading.BoundedSemaphore; async limits use asyncio.Semaphore keyed by (task, event-loop) so each loop gets its own. Semaphores are rebuilt when the configured limit changes, so live config edits take effect.
• Default behavior is unchanged (no limit) so existing setups don't regress. The session-search task already has its own concurrency cap inside tools/session_search_tool.py; leaving that untouched.
• cli-config.yaml.example and website/docs/user-guide/configuration.md document the knob with recommended values for title_generation and compression.

How to test

• pytest tests/agent/test_auxiliary_concurrency.py -q — the new tests cover: config parsing (invalid / zero / negative / missing values), semaphore caching + rebuild on limit change, sync concurrency enforcement under 6 concurrent threads with limit=2, async enforcement under 6 concurrent tasks with limit=2, and semaphore release on exception so the cap doesn't deadlock after failures.
• pytest tests/agent/test_auxiliary_client.py -q — full existing auxiliary suite still passes (two unrelated Codex-timeout tests are pre-existing flakes; verified they fail on main without these changes too).
• Manual: set auxiliary.title_generation.max_concurrency: 2 in ~/.hermes/cli-config.yaml, open multiple sessions that all trigger first-exchange title generation, and watch the provider logs — only two title calls run at a time.

What platforms tested on

• macOS on darwin-arm64 (local)

Fixes #23324

<!-- autocontrib:worker-id=issue-new-daf1ce33 kind=pr-open -->

Changed files

• agent/auxiliary_client.py (modified, +160/-0)
• cli-config.yaml.example (modified, +19/-0)
• tests/agent/test_auxiliary_concurrency.py (added, +288/-0)
• website/docs/user-guide/configuration.md (modified, +30/-0)