Hermes Agent Roundup: Durable Delegations, Codex Usage Reset, DeepInfra Provider
X API v2 remains fully paywalled (search and lookup both return 402). This roundup is sourced from the GitHub API - 13 feature PRs merged in the last 48 hours, with 6 landing substantial changes across delegation reliability, provider expansion, and desktop UX.
Durable background delegations
The most impactful change of the window: background delegation completions now survive process restarts. teknium1's #63494 (merged July 13, +1,323 / -77 in 11 files) adds three layers of durability to the delegation pipeline.
Dispatch ownership is persisted to the profile's state.db before daemon work is submitted. Terminal results and completion events are persisted before publishing to the fresh-turn queue. On restart, undelivered events are restored through the normal ownership-checked consumer rail. The PR adds cross-process delivery tests - 494 lines of new test coverage in test_completion_delivery.py and 179 lines in test_async_delegation.py.
The practical effect: if Hermes crashes or is restarted mid-delegation, the delegated work completes and the result waits in the queue rather than being silently dropped. No configuration changes needed - it works through the existing state database.
DeepInfra provider
kshitijk4poor salvaged the DeepInfra integration from a prior PR (#63767) onto current main in #63969 (merged July 13, +2,539 / -163 in 40 files). The integration preserves original commit authorship from Georgi Atsev.
DeepInfra joins the provider roster with chat, image generation, video generation, TTS, and STT support. Paid image and TTS backends are gated behind explicit user selection. The PR also strips authorization headers across cross-origin catalog redirects for security and filters the chat picker to DeepInfra's own model catalog.
The new agent/video_gen_provider.py (+291 lines) wires up DeepInfra's video generation API. Provider additions in agent/model_metadata.py (+18 lines) and agent/auxiliary_client.py (+66 / -6) handle catalog and pricing integration.
Codex usage reset from within Hermes
#64280 by teknium1 (merged July 14, +579 / -7 in 24 files) lets Codex users see and redeem their banked usage-limit resets directly from Hermes via /usage reset. Previously this required opening the Codex CLI, app, or website.
The core logic lives in agent/account_usage.py (+198 / -4): a new parser reads rate_limit_reset_credits.available_count from the Codex /usage payload, and redeem_codex_reset_credit() hits POST .../rate-limit-reset-credits/consume. The /usage command now shows "You have N resets banked - use /usage reset to activate."
Gateway slash command routing in gateway/slash_commands.py (+24 lines) and CLI wiring in cli.py (+48 / -1) complete the path. Locale strings were added across 20 language files.
Reasoning effort levels: max and ultra
#62650 (merged July 12, +175 / -65 in 47 files) adds max and ultra as generic reasoning-effort choices across every surface: CLI, gateway, dashboard, desktop, delegation, batch, Nous Portal, OpenRouter, ChatGPT OAuth, and direct OpenAI routes.
GPT-5.6's API exposes max; Codex exposes ultra as the product tier and sends max on the wire. The PR normalizes this: GPT-5.6 ultra maps to the API wire value max on both Chat Completions and Responses routes. Compatibility is preserved by clamping the new top tier on providers that only support low/medium/high.
The change touched 47 files across the stack - shared parsing, slash completion, settings, pickers, labels, docs, and translations. Desktop i18n strings in English and Japanese were added. The broad component footprint (agent, CLI, gateway, desktop, dashboard, OpenAI provider) signals how reasoning effort passes through the architecture end-to-end.
Dashboard and desktop improvements
Two more PRs from July 13 round out the window:
#63699 by kshitijk4poor (+996 / -24 in 9 files) adds a safe session import flow to the dashboard, salvaging shannonsands's work from #54586. Imports validate metadata and SQLite-bound scalar fields before writes, with malformed batches rolling back atomically. Bounds on sessions, messages, and serialized import size prevent oversized payloads from destabilizing the UI.
#63520 by teknium1 (+714 / -74 in 24 files) adds profile-aware Smart/Manual/Off approval-mode control to the desktop app, salvaging Tortugasaur's #43205. The control replaces the binary approval bypass with a three-way statusbar menu that reads and writes through the focused gateway profile, reconciling optimistic state through backend confirmation.
Across these six PRs: 6,326 additions across roughly 150 files from 5 contributors (teknium1, kshitijk4poor, vishal-dharm, ethernet8023, OutThisLife). The cluster centers on resilience (durable delegation, safe session imports) and provider optionality (DeepInfra, reasoning effort levels).
[^1]: teknium1. "feat(delegation): make background completions durable." NousResearch/hermes-agent. July 13, 2026.
[^2]: kshitijk4poor. "feat(providers): add hardened DeepInfra integration." NousResearch/hermes-agent. July 13, 2026.
[^3]: teknium1. "feat(codex): redeem banked usage-limit resets via /usage reset." NousResearch/hermes-agent. July 14, 2026.
[^4]: teknium1. "feat(reasoning): add max and ultra effort levels." NousResearch/hermes-agent. July 12, 2026.
[^5]: kshitijk4poor. "feat(dashboard): add safe session import flow." NousResearch/hermes-agent. July 13, 2026.
[^6]: teknium1. "feat(desktop): add profile-aware approval mode control." NousResearch/hermes-agent. July 13, 2026.