Long-Lived Sessions Finally Catch Up: Compaction Now Always Rebuilds the System Prompt
Hermes used to restore a long-lived session's system prompt byte-for-byte from its snapshot at every compaction — so guidance improvements, new prompt blocks, and tool renames never reached a session that ran for months. A change merged August 30 (PR #98426) makes the compaction commit always run the live prompt builder: byte-equal output keeps the original object (prefix caches intact), a drifted output wins and logs the diff.
Read more
Context Compression Gets 5x Faster: One Auxiliary Request Instead of 19
Old lean compaction called the auxiliary model once per history chunk to build digests — up to 28 sequential calls on a big session, turning compaction into 7-11 minute stalls on slow auxiliary routes. A change merged August 30 (PR #98628) deletes the digest loop: one auxiliary request per compaction attempt. Measured on a real 500K-token transcript: 196.5s to 39.6s, and ~11K fewer tokens after compaction.
Read more
Give Hermes a time machine: search the web as it was with BackSearch
Imagine asking an LLM to predict next month’s news and realizing it peeked at today’s headlines. BackSearch by General Reasoning freezes the news archive so searches only return documents crawled on or before a date you choose. An official standalone Hermes plugin (NousResearch/hermes-plugin-backsearch, v1.0.0, announced Aug 29) adds two tools — backsearch and backfetch — gated on OPENREWARD_API_KEY, with zero tool footprint when the key is absent. This post covers the concept, install steps, tool reference, and use cases for leak-free forecasting backtests.
Read more
Three new model providers: Ramp Router, Nebius Token Factory, and Tencent TokenPlan
On August 29 Hermes merged three new model providers: Ramp Router (router.com) — one OpenAI Responses-compatible endpoint that routes to many upstreams with server-side fallbacks and spend controls; Nebius Token Factory — Nebius pay-as-you-go inference; and Tencent TokenPlan — the Tencent Cloud Hunyuan token-plan subscription on an Anthropic-compatible endpoint, which also brings the hy4-preview model into the official catalog. PRs #97915/#97916/#97917 are on main, not yet in a release.
Read more
Three new official optional skills: questionnaires, setup wizards, and a plan interrogation
On August 29 Hermes landed three new optional skills at once: decision-questionnaire turns a decision only one person can make into an async questionnaire, setup-wizard-generator produces interactive bash wizards for human-only setup steps, and grill-me v2.0.0 upgrades to frontier-rounds interviewing that interrogates every unexamined assumption in your plan. All three are Hermes-native ports from the 240k-star mattpocock/skills repo, installable with a single hermes skills install.
Read more
Task lists can nest now: Hermes todo grows subtasks
When you hand Hermes a complex job, it keeps a todo list in the session to track its own progress. That list used to be flat — "package the release" and "make sure tests pass" could only sit side by side. PR #97921 (merged Aug 29) adds a parent field to todo items: tasks can nest into trees, surfaces render the indentation automatically, and the whole feature costs only ~45 tokens on the cached tool schema.
Read more
Need to ask a question mid-task? /btw answers it right now, without interrupting
You used to have two bad options when a question popped up mid-run: interrupt the agent, or open a fresh session that knows nothing about the conversation. PR #97937/#97974 (merged Aug 29) turns /btw into a real side-question command: it answers from the current conversation with full context at cache-read prices — no interruption, no pollution, no extra spend. /background retires in the same change; /bg becomes the canonical background-task command.
Read more
One Command Opens a Whole Session: hermes chat -q Now Stays Interactive on a TTY
hermes chat -q "question" used to answer once and exit; getting an interactive session meant running it, saving the session file, then resuming — a clunky two-step dance. The change merged on August 28 (PR #97121) makes -q seed a live interactive session on a real TTY, with the first message submitted literally — never parsed as a slash command or shell escape.
Read more
Hermes Stops Guessing Your Context Size: Usage Anchoring Aligns Compression Thresholds with Real Numbers
Hermes used to re-estimate the whole conversation's token count every turn with heuristics (chars/4, flat 1500-token images), compounding error as history grew — causing surprise compressions and 413 payload-too-large errors. The change merged on August 28 (PR #97206) anchors context accounting on provider-reported usage, shrinking the estimation window from the whole transcript to a single turn.
Read more
Session Temp Files No Longer Blow Up Your RAM Disk: terminal.temp_dir and 72-Hour Auto-Cleanup
Many Linux distros mount /tmp as a RAM-backed tmpfs, and heavy Hermes sessions fill it up. The fix merged on August 28 moves session temp files to real storage at ~/.hermes/cache/terminal by default, adds the terminal.temp_dir config key and TERMINAL_TEMP_DIR env var, and auto-prunes files older than 72 hours.
Read more
A Quiet "Tool Diet": 6 Tool Definitions Slimmed Down, Saving 17%–44% Tokens Per Call
Every time Hermes calls a tool, the model must re-read that tool's JSON parameter definition (schema) along with the request — the bigger the definition, the higher the fixed per-turn overhead. Between August 27 and 28, the maintainers merged a wave of "schema diet" changes: process, todo, read_file, skill_manage, video_generate, and browser_exec all slimmed down, cutting 17%–44% of tokens per call with zero behavior change.
Read more
Search Once, Ask Thrice: Multi-Query tool_search With Stemming
Hermes' tool_search just got a serious upgrade: it now takes an array of queries searched in parallel, returns results grouped per query with one shared tools map, and Snowball stemming matches 'browsing' to 'browser' and 'issues' to 'create_issue'. tool_describe batches names too, so one bad name no longer fails the whole call. Fewer round-trips, cheaper discovery, better recall.
Read more
No More Keychain Nag: Opt-In OS-Keychain Encryption for Stored Secrets
Electron safeStorage used to park a per-app key in the macOS login keychain — and on machines with a locked or missing keychain, every Hermes Desktop launch turned into a blocking 'Keychain Not Found' password prompt. v0.20.6 makes keychain-backed encryption an explicit opt-in: the default path never touches safeStorage at all, a one-shot migration converts old encrypted blobs to plain 0600 files, and flipping the toggle re-encodes every stored secret in place.
Read more
Browse as You: Real-Profile Browsing in Hermes Agent
Hermes' browser used to start every session logged out — a wall the moment a task needed your logins. Consent-gated real-profile browsing copies your default browser's active profile into a managed snapshot and drives it with Hermes' packaged Chromium: the agent browses as you, with your cookies and saved logins, always under explicit consent and off by default. Here's how it works, how to enable it, and the Windows caveats.
Read more
Ship It With One Command: the publish-site Skill for Versioned Website Deploys
The new publish-site skill turns 'put this website online' into a disciplined five-step pipeline: build, preview locally (or via a shareable tunnel), version every deploy with a git tag, deploy through a provider ladder (GitHub Pages → Cloudflare Pages → Netlify), and verify the live URL returns HTTP 200 before reporting success. Rollback is one command away.
Read more
Hermes v0.20.6: The Agent Learns to Browse as You, a 50+ Server MCP Catalog, and Secrets Without the Keychain Nag
v0.20.6 is a steady roll of substance: consent-gated real-profile browsing lets the agent use your real browser logins, the desktop Browser gets its own OS window plus a managed SSH remote-update engine, the remote MCP catalog passes 50 live-verified vendor servers, web_search/web_extract gain TTL result caching, lean-tail compression becomes the default, tool_search learns multi-query searches with stemming, and stored secrets can opt into OS-keychain encryption — no more per-launch Keychain prompts.
Read more