Hermes Agent Roundup: max/ultra Reasoning, Workspace Filter, Fireworks AI
The X API v2 search and lookup endpoints became fully paywalled today (HTTP 402 on every query). Today's roundup is sourced from GitHub, where development velocity continues at pace. Four feature PRs merged in the last 48 hours, plus several fixes on the same day.
Max and ultra reasoning effort levels
PR #62650 by teknium1 adds max and ultra as reasoning-effort choices across every Hermes surface. The patch touches 47 files (+175/-65) spanning CLI, gateway, dashboard, desktop, delegation, batch processing, Nous Portal, OpenRouter, ChatGPT OAuth, and direct OpenAI routes.
This closes a gap that existed since at least April 2026. The Anthropic adapter already mapped max to Claude's output_config.effort = max, but every input validator - from config parsing to slash commands to the desktop picker - silently rejected the value and fell back to the default medium. A user who set agent.reasoning_effort: max was downgraded two levels below their previous xhigh without an obvious error.
The new code normalizes max at the provider boundary. For GPT-5.6, the API's max maps directly. For Codex, the product tier ultra sends max on the wire. Non-Anthropic providers that do not support the top levels get clamped to their strongest supported effort.
Five community PRs and issues led up to this. #49580 (June 20) documented the validator gap. #12211 (April 18) attempted an earlier fix. #61634 and #61648 (July 9) tackled GPT-5.6 support. #61996 (July 10) exposed max for GPT-5.6 specifically. The merged PR consolidates all of them.
CLI workspace filter and cwd restore
PR #63091 by OutThisLife (+141/-0 in 5 files) adds a --workspace filter to the CLI and auto-restores the working directory when resuming a session.
The feature builds on session git metadata that main already records - git_repo_root and git_branch captured through the desktop and gateway paths. It supersedes an earlier PR #48591 by professorpalmer that predated the metadata layer and required a parallel git tracking model.
The new implementation is scoped narrowly: it reads from the existing metadata, exposes a CLI filter, and restores cwd on resume. No new database columns, no shelling out to git. The test file covers session workspace binding end to end.
Gateway runtime readiness checks
PR #62645 by teknium1 (+316/-5 in 6 files) adds bounded readiness checks to the authenticated gateway health endpoint.
The public /health liveness probe stays unchanged - cheap, no auth, binary answer. A new authenticated /health/detailed endpoint reports real subsystem state: config parsing success, read-only state DB access, configured model presence, disk pressure, gateway and platform state, active API runs, pending process completions, and active delegations.
The readiness module is self-contained in gateway/readiness.py (117 lines) with a matching test suite (103 lines). The API server integration is 36 lines. This is the kind of operational surface that matters when you run Hermes on a VPS and need to know whether a restart resolved the issue without SSHing in.
Fireworks AI joins as bundled provider
PR #62593 by teknium1 (+457/-8 in 22 files) makes Fireworks AI a first-class provider across all model selection surfaces.
The integration uses Fireworks' OpenAI-compatible endpoint with live catalog, PAYG-safe fallbacks, and auxiliary model routing. It credits nine earlier community PRs that attempted the same integration: #3564, #3757, #6054, #15332, #23541, #28983, #51479, #51496, and #61182.
No partner-attribution headers, no promoted placement - just a standard provider plugin.
Also merged today
- #63123: Kanban persistent workspace guidance - task creation now defaults project-backed boards to persistent workspaces and makes scratch deletion explicit (+84/-14, 3 files)
- Desktop fallback editor resync after config reload
- API stream lifetime fixes - separating run control from stream lifetime, keeping live runs tracked past stream TTL
The repo sits at 213K stars, 390 contributors, and 27,457 open issues. The gap between tagged releases continues to widen - v0.18.2 shipped July 7, and several hundred PRs have merged since then. The next curated release (v0.19.0) will document the full window.