Changelog

HERMES_DASHBOARD_INSECURE is now a deprecated no-op. Non-loopback dashboard binds require an auth provider (basic auth, Nous Portal OAuth, or self-hosted OIDC). Updated Scene 4 from 5 cards to 4 cards (removed resource limits card, changed dashboard card to "supervised dashboard + auth"), added note about the no-op flag and auth requirement. Updated narration to explain the auth gate. 7 scenes, 158s, 9.0 MB. Quality gates passed (0 errors). Frame inspection confirmed 4 cards and no-op note visible.

discord.allow_any_attachment is now a no-op, any file type is always accepted. Updated Scene 7 (sessions) subtitle to mention that any file type uploaded is accepted and cached for the agent to inspect. Updated narration to mention the no-op flag. 9 scenes, 177s, 9.6 MB. Quality gates passed (0 errors). Frame inspection confirmed file-type text visible in sessions scene.

Added fetch_patch_from_commits() function that fetches actual diffs from individual commits when the GitHub compare API returns empty patches (+0/-0). This was the root cause of all changed docs being classified as "review" with "No patch data available" instead of being actionable. Verified: all 4 changed docs now correctly classify as re-cut or new-episode.

Second re-cut to sync script.md with narration.txt (two sync fixes in Scene 5 and Scene 7). The /learn command was already present from the Jun 24 re-cut. 8 scenes, 155s, 8.9 MB. Quality gates passed (0 errors). Frame inspection confirmed /learn text visible in Scene 5.

Second re-cut to add fail-closed provider drift behavior (commit a4e61ddf). Unpinned cron jobs now snapshot the provider at creation. If the global default changes, the job fails closed: skips the run and alerts you to pin the provider/model explicitly. Added to Scene 4 narration and Scene 3 concept sub text. Font size reduced from 72 to 66 to fit longer text. The continuable jobs feature (mirror_delivery) was already present from the earlier Jun 25 re-cut. 8 scenes, 195s, 10.6 MB. Quality gates passed (0 errors). Frame inspection confirmed fail-closed text visible in Scene 4.

Three material signals from telegram.md. (1) Command menu cap raised from 30 to 60 so built-in commands and common skills stay visible in the Telegram slash picker (commits dbe14ce3, d4be583d). (2) New command_menu.priority_mode config for custom ordering (prepend, append, replace). (3) New rich_drafts config (default false) to gate experimental rich draft previews during DM streaming (commit 17beb55e). Updated Scene 8 bonus cards from 3 to 4 cards (added command_menu: 60 card), added note about rich_drafts: false. Updated narration, script.md header, and how-to.md. Em-dash scan clean. 9 scenes, 192s, 10.7 MB. Frame inspection confirmed all 4 cards and rich_drafts note visible, no layout issues.

reference/environment-variables.md: HERMES_ENABLE_PROJECT_PLUGINS description updated to reflect that project plugin Python API files are no longer auto-imported even when the var is enabled (commit c42d44cb). This is a security wording clarification, not a new env var or changed default. SHA bumped for Env Var Cheat Sheet.

user-guide/features/extending-the-dashboard.md: Material change to plugin backend route security model (commit c42d44cb). User-installed plugins can now register backend API routes (previously bundled-only). Plugin discovery priority reordered (user-installed now wins on conflict). No existing episode covers this doc directly. Needs manual authoring.

New /learn slash command added upstream (commit e32ebc6a). /learn distills a reusable skill from anything you describe: a directory, a URL, or the workflow you just walked through. Updated Scene 5 terminal to show /learn the REST client in ~/sdk, focus on auth, updated note to mention /learn distills skills, updated narration to describe slash learn as the fast way to create a skill. Also updated how-to.md with /learn in the skills section. Em-dash scrubbed from meta.json. 7 scenes, 144s, 8.4 MB. Frame inspection confirmed /learn visible in terminal and note.

New /learn feature added to the skills doc (commit e32ebc6a). /learn is the fast way to turn anything you can describe into a reusable skill without hand-writing the SKILL.md. Updated Scene 5 (procedural memory) to feature /learn alongside skill_manage as the two ways to create skills from the agent side. Reduced font size from 72 to 64 to fit the longer sub text. Updated narration, script.md, and how-to.md. Em-dashes scrubbed from script.md scene headers, how-to.md, and meta.json. 8 scenes, 158s, 9.4 MB. Frame inspection confirmed /learn text visible and layout clean.

12 docs had pip install replaced with cd ~/.hermes/hermes-agent && uv pip install -e "." (commit a911bcda). These are install-path updates with no narrative impact: aws-bedrock.md, use-voice-mode-with-hermes.md, faq.md, tts.md, voice-mode.md, web-dashboard.md, dingtalk.md, matrix.md, sms.md, weixin.md.

cli-commands.md had hermes update and hermes postinstall descriptions simplified (pip references removed). No new commands or flags for Five Slash Commands/Pick Your Brain/Sessions to teach.

sessions.md had a clarification warning box added about sessions.json being the gateway routing index, not the session list. Clarification only, no new feature.

skills-catalog.md had a new petdex skill entry added to the catalog table. Catalog addition, not a feature change.

(new): Petdex mascots feature. Animated pet sprites that react to agent activity across CLI, TUI, and desktop. Off by default, purely cosmetic. Needs manual authoring.

(new): Skill catalog page for the petdex skill. Bundled skill, not a standalone episode topic.

pip install path removed entirely from updating.md. The hermes update section was restructured (git/pip split removed, unified path). Material but deferred to next run (2 per cap).

New command_menu config section with max_commands cap raised from 30 to 60, plus priority_mode and priority list. Material but deferred (Hermes on Telegram was already re-cut on 2026-06-23 for rich messages opt-in).

The kanban-worker and kanban-orchestrator bundled skills were folded into auto-injected system-prompt guidance. Workers and orchestrators no longer need a skill installed; the lifecycle is injected at spawn. Updated Scene 6 narration and on-screen text to say "No skill to install: the worker lifecycle is auto-injected into every worker's system prompt at spawn." 11 scenes, 282s, 15.9 MB.

Rich messages changed from enabled-by-default to opt-in. The default is now the legacy MarkdownV2 path because current Telegram clients make rich text hard to copy. Updated the bonus tricks card from status_indicator: true to rich_messages: true with "opt in for native tables, task lists, math. off by default for copy fidelity." script.md created (episode predated v2 spec). 9 scenes, 173s, 9.7 MB.

36 additional docs with detected changes are classified and deferred to subsequent runs (2 per run cap). These include material changes to: hooks (subagent_start, manual allowlisting), build-a-plugin (kanban lifecycle hooks, dispatch_tool), memory-providers (Mem0 v3 API, Honcho OAuth), memory (background_review auxiliary model), configuration (worktree_sync, session limit 400 to 5000, email pairing), security (email pairing opt-in), docker (dashboard insecure deprecated, tool-loop hard stops), slash-commands (new /suggestions, /blueprint, /credits, /billing), providers (Google Gemini OAuth removed), configuring-models (mid-session switch warning), computer-use (cross-platform, telemetry), and others.

removed the "New Guides" section. Every episode now lives in its real thematic module (Get Running, Daily Driver,... Locking It Down) instead of a catch-all that grew forever.

from the menu, video title cards, and pages. The series reads as a topic library, not a rigid numbered list. "What's new" lives in this changelog.

a collapsible module drawer (the current section auto-opens), a sticky top bar showing where you are, and a thumb-reachable bottom bar (Menu · Search · Updates). Much easier to get around on a phone.

straight to the next episode.

every page now publishes structured data (VideoObject, HowTo, BreadcrumbList), Open Graph/Twitter cards, and canonical URLs; the site now serves sitemap.xml, robots.txt, and an llms.txt index so search engines and AI assistants can find and cite the episodes.

added (used for social/preview images and video structured data).

Connect Hermes to SimpleX, the decentralized messenger with no persistent user IDs. Daemon over WebSocket, allowlist or DM pairing.

LINE Messaging API channel for Hermes. Channel token/secret, webhook tunnel, U/C/R ID allowlist.

Google Chat via a Cloud Pub/Sub pull subscription and a service account.

Lightweight ntfy push-notification channel (httpx only, no SDK); topic-is-identity model and its limits.

Teams bot via the Teams CLI and AAD app, DM vs @mention behavior, Adaptive Card approvals.

Teams meeting pipeline (transcripts/recordings) via MS Graph webhook + the teams_pipeline plugin.

Let Hermes drive a desktop in the background. Enable, permissions, first session, and the safety guardrails.

The X (Twitter) search tool via xAI, config (model/timeout/retries) and the degraded-result flag.

How Hermes produces real file artifacts, supported types, and the absolute-path mechanic.

Admin-pinned config and secrets in /etc/hermes that users cannot override; precedence and limits.

Run multiple profiles behind one gateway with gateway.multiplex_profiles and /p/<profile>/ routing.

The browser dashboard and Chat-tab session switcher; install, pages, and the auth gate.

Opt-in workspace checkpoints and /rollback list/diff/restore for safe recovery.

Model list now cached one hour then refreshed (model_catalog.ttl_hours default 24 to 1); provider-add example slugs refreshed. Reason: reference/cli-commands.md + reference/model-catalog.md.

--clone now copies skills too; new --clone-from flag; s6-overlay per-profile gateway supervision in Docker; dashboard profile switcher. Reason: user-guide/profiles.md.

Single-query worktree flag changed from -q to -z; MESSAGING_CWD replaced by the terminal.cwd config key. Reason: user-guide/git-worktrees.md.

The context-file 20K truncation cap is now configurable via context_file_max_chars. Reason: user-guide/features/context-files.md.

New STT providers (Mistral voxtral, xAI grok-stt) and TTS providers (mistral, gemini, xai, kittentts, piper); Docker Linux desktop audio bridge. Reason: user-guide/features/voice-mode.md.

New discord.voice_fx config (ambient thinking bed while tools run + spoken acknowledgement). Reason: user-guide/messaging/discord.md.

New !approve / !deny text-fallback approval prompts when interactive buttons can't render. Reason: user-guide/messaging/slack.md.

Pointer to the official WhatsApp Cloud API path; new message-batching debounce config (text_batch_delay_seconds, text_batch_split_delay_seconds). Reason: user-guide/messaging/whatsapp.md.

Tool-progress bubbles are suppressed on Signal (no message editing); final reply still delivered. Reason: user-guide/messaging/signal.md.

New distinction between the Email gateway adapter (zero-dep imaplib/smtplib) and the Himalaya email skill. Reason: user-guide/messaging/email.md.

SMS_WEBHOOK_HOST default changed from 0.0.0.0 to loopback 127.0.0.1 (security). Reason: user-guide/messaging/sms.md.

New troubleshooting (env vars apply on restart; REST 401 means use a Long-Lived Access Token). Reason: user-guide/messaging/homeassistant.md.

New hermes skills opt-out/opt-in/--sync/--remove, --no-skills flag, skills.write_approval gate with /skills review, NVIDIA trusted source. Reason: user-guide/features/skills.md.

Name-based (case-insensitive) job lookup for pause/resume/run/remove/edit; new hermes cron edit verb; broadened [SILENT] matching. Reason: user-guide/features/cron.md.

Silence-token matching broadened to [SILENT], SILENT, NO_REPLY, NO REPLY; new telemetry correlation fields (telemetry_schema_version, turn_id, api_request_id). Reason: user-guide/features/hooks.md.

Cron jobs now addressable by name; new hermes cron edit verb shown in the test step. Reason: user-guide/features/cron.md.

HERMES_* env passthrough locked to four vars; new terminal.env_passthrough allowlist; required_environment_variables skill frontmatter. Reason: user-guide/features/code-execution.md.

Default model claude-sonnet-4 to claude-sonnet-4.6; model constructor param now resolves from config at runtime. Reason: guides/python-library.md.

Added AWS Bedrock, Azure Foundry, NVIDIA, Ollama Cloud; new 64K minimum context floor (smaller windows rejected at startup). Reason: integrations/providers.md.

fallback_providers (list) replaces fallback_model; subagents and cron jobs now inherit the chain; per-task fallback_chain for auxiliary tasks. Reason: user-guide/features/fallback-providers.md.

New usage-limit rotation (rotate immediately, no retry); auth.json gains secret_source/secret_fingerprint; borrowed secrets are reference-only. Reason: user-guide/features/credential-pools.md.

Auxiliary task slots expanded 8 to 11 (Triage Specifier, Kanban Decomposer, Profile Describer, Curator); dual model_aliases formats. Reason: user-guide/configuring-models.md.

Newer platforms (Matrix, WhatsApp Cloud) use _ACCESS_TOKEN rather than _BOT_TOKEN; 30+ new vars fit the existing buckets. Reason: reference/environment-variables.md.

What Hermes Is, One-Line Install, Your First Chat, Termux, Nix, Updating, Five Slash Commands re-rendered with the install-URL, provider-list, skills-hub, and updating/safety changes already reconciled in the June doc-sync.

Tools, SOUL.md, Memory, Memory Providers, Browser, Web Search, Matrix, Delegation, MCP, Plugins Tour, Build a Plugin, API Server, Local LLM on Mac, Security, Docker re-rendered to current content.

Re-rendered; sessions doc itself changed only cosmetically, commands re-verified against current CLI.

Local Ollama, Nemotron 3 Ultra Free, Photon iMessage, WhatsApp Cloud API, Run with Nous Portal, Tool Search, TTS re-rendered to current content.

8 docs with link-path/Portal-tip-only changes got a docSHA bump in coverage.json so doc-watch stops re-flagging them: daily-briefing-bot.md (Daily Briefing Bot · Capstone bonus), use-voice-mode-with-hermes.md (Voice Mode), batch-processing.md (Batch Processing), goals.md (Five Slash Commands/Set a Goal, Walk Away), personality.md (SOUL.md/AGENTS.md), provider-routing.md (Pick Your Brain/Provider Routing), vision.md (Vision · Paste), sessions.md (Sessions).

user-guide/features/rl-training.md: Removed upstream. No published episode covers it.

user-guide/skills/godmode.md: Removed upstream. No episode impact.

reference/cron-recipes-catalog.mdx: Removed upstream (auto-generated catalog). No episode impact.

developer-guide/environments.md: Removed upstream (contributor-facing). No episode impact.

user-guide/skills/optional/productivity-shop-app.md plus several deprecated skill docs (creative-concept-diagrams, mlops-atropos-environments, blockchain-base): Removed upstream. A coverage.json cross-check confirms none of these map to a published episode, so no re-cuts are required.

53 material doc changes detected in the 2026-06-18 audit (brand/docs-audit-report.json). 5 re-cut so far (The Curator, Image Generation, Hermes on Telegram, Kanban Board, Configuration Deep Dive); ~48 episodes remain.

8 docs need only a docSHA bump, no re-cut (Daily Briefing Bot · Capstone, Voice Mode, Batch Processing, Set a Goal, Walk Away, SOUL.md/AGENTS.md, Provider Routing, Vision · Paste, Sessions).

guides/automate-with-cron.md (Daily Briefing Bot · Capstone) and user-guide/features/kanban-tutorial.md (Kanban Board) — reworded guidance, no code/flag/config change.

25+ new end-user docs above are not yet built.

local clone + watch-state sit at upstream HEAD cf58f1a5 (2026-06-19). A fresh gh-backed audit (python brand/doc-watch.py --audit) is still needed to capture any 2026-06-20 → 2026-06-22 changes.

Full build. 5 content scenes covering what the changelog is, the Last Updated box, three sections per date group (re-cuts, new episodes, cosmetic), material vs cosmetic distinction, and where to find the changelog from any episode page. 7 scenes total (title + 5 content + outro), 71s, 4.6 MB. TTS via ElevenLabs. Frame inspection confirmed all scenes render cleanly with no overlaps or black frames.

Added the optional Online/Offline status indicator (status_indicator: true config flag, off by default). Expanded bonus tricks from 2 cards to 3 cards. 9 scenes, 166s, 9.4 MB. Frame inspection confirmed the new status_indicator: true card renders correctly. MP4 deployed to site/videos/.

Added title_generation to the universal pattern scene (another model slot example alongside auxiliary, compression, and fallback). Added managed scope/org deployment mention to the new-defaults scene (admins can pin config and secrets users cannot override). 8 scenes, 230s, 12.9 MB. Frame inspection confirmed "title generation" and "managed scope" text visible. MP4 deployed to site/videos/.

(user-guide/managed-scope.md): New doc. Needs manual authoring.

(user-guide/messaging/raft.md): New doc. Needs manual authoring.

(user-guide/features/web-dashboard.md): New session switcher panel feature. No existing episode covers the dashboard directly.

(user-guide/multi-profile-gateways.md): New gateway.multiplex_profiles opt-in config and contract changes. No existing episode covers this directly.

reference/cli-commands.md: hermes debug report now includes GUI/dashboard logs. No new commands or flags. SHA bumped for Five Slash Commands, Pick Your Brain, Sessions.

Full build. 8 content scenes covering hardware requirements, Ollama install, model pull, Hermes setup, first chat, optimization tips, and cost comparison. 167s, 10.3 MB. Previous render was stuck on title slide (broken composition). Rebuilt episode.mjs with correct data format, regenerated TTS (10 clips, ElevenLabs), rendered via HyperFrames. Frame inspection confirmed all scenes transition correctly.

Full build. 6 content scenes covering the free Nous Portal offer, desktop app install, CLI install, the colon-free variant detail, switching models, and troubleshooting. 104s, 6.7 MB.

Full build. 6 content scenes covering what Photon is, the persistent-connection architecture, prerequisites, setup command, starting the gateway, and free-tier limits. 111s, 6.7 MB.

Full build. 7 content scenes covering the official API path, the setup wizard, prerequisites, credentials from Meta dashboard, webhook setup with cloudflared, the 24-hour conversation window, and security notes. 138s, 8.3 MB.

All 4 new episode HTML pages deployed to actionableops.com. All return 200.

All 4 new MP4 video files deployed and return 200.

Site sidebar now lists 66 episodes across 10 modules.

Fixed TTS and page.json next-field format (string to array) for build-page-html.mjs compatibility.

Fixed local-ollama-setup page.json cmd field names (text to code) and source URLs (relative to absolute).

Image-to-image editing added upstream. The image_generate tool now edits existing images when the active model supports it. New inputs: image_url (source photo) and reference_image_urls (style/composition cues). Four backends support editing: FAL (edit-capable models), OpenAI gpt-image-2, xAI Grok Imagine, Krea 2. Added a new concept scene (Scene 5) covering the edit mode. Expanded from 8 to 9 scenes. Rendered 202s, 10.7 MB. Frame inspection confirmed all scenes clean. Reason: upstream commit 9a2f275 added the image-to-image editing section to image-generation.md.

New config key auto_subscribe_on_create (default true). When a worker calls kanban_create from a session with a persistent delivery channel, the originating session is auto-subscribed to the new task's events. Set false to require explicit kanban_notify-subscribe calls. Updated the human-in-the-loop scene sub text and narration to name the config key. 11 scenes, 271s, 15.3 MB. Frame inspection confirmed the new text renders cleanly. Reason: upstream commit 9a2f275 added auto_subscribe_on_create to the kanban config table.

1 new-episode candidate detected (reference/tools-reference.md, image_generate tool entry). Logged for manual authoring. New episodes are not auto-created.

user-guide/skills/optional/payments/payments-stripe-projects.md: borderline change, logged for human review.

Built the doc-to-episode ledger. 59 episodes, 90 source references, 87 blob SHAs. Every source doc is now machine-traceable to its episode.

Change detection script. Monitors upstream NousResearch/hermes-agent (branch main, path website/docs/) via gh api. Compares last-seen SHA against HEAD. Falls back to per-doc SHA audit when the compare API is capped at 300 files. Classifies changes: added, modified, removed, renamed. Guardrails: single-flight lock, kill switch, per-run cap (2), notification hook.

Material vs cosmetic decision tree with 5 worked examples from real upstream commits.

Full pipeline orchestration. Detects changes, builds episodes (TTS + composition + check + render + ffprobe), updates coverage.json, git backup, notifies. Dry-run mode for testing.

Windows Task Scheduler setup. Daily at 9 AM with catch-up enabled.

Seeded to May 24 baseline (when episodes were built) so the first real run detects all changes between episode creation and now.

LLM consolidation is now opt-in (was always-on). New config key curator.consolidate: false. New flag hermes curator run --consolidate. Updated script, narration, composition, description, how-to. Rendered via HyperFrames (1920x1080, 195s, 10.7 MB). TTS via ElevenLabs. Reason: upstream commit 7bbffce made consolidation opt-in.

Fixed 4 broken source URLs (404 dead links on published site):

Updating Hermes (updating): cli-reference/doctor -> reference/cli-commands

One-Line Install (one-line-install): getting-started/windows -> user-guide/windows-native

Vision · Paste (vision): user-guide/context-references -> user-guide/features/context-references

Sessions (sessions): user-guide/memory -> user-guide/features/memory

All 4 deployed live to actionableops.com.