| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354 |
- name: alans-way
- repo: https://github.com/capthvnsen/alans-way-agents
- sha: 0f65549ac1353f6b74b8ac6a22d58f3eb6dea0a3
- subdir: alans-way
- version: "0.21.9"
- title: "Alan's Way Plugin"
- description: "Idle check-ins from the primary bot of a Hermes gateway on a Linux machine or a macOS guest VM: when the bound Telegram chat has been quiet for two hours, during 08:00-22:00 in the user's timezone, the bot checks in once (does one safe step and reports it, suggests one, or asks one question), backing off when unanswered. The model tool can only make check-ins quieter; more often, resume and wider hours are operator-only via /proactivity or the CLI. After install, the workspace-setup skill walks the agent through setting up Alan's Workspace (a separate desktop app for the user's computer: macOS or Windows) and Tailscale; desktop input goes only through Hermes's approval-gated computer_use tool. Disclosure — once the gateway connects Telegram, a background thread checks every 60 seconds. When the bound Telegram DM has been quiet (2 h by default, doubling per unanswered check-in, 08:00-22:00 user time), it injects an internal user-role prompt into that chat (needs plugins.entries.alans-way.allow_gateway_injection, which setup sets), so the bot runs a turn on its own and may take one \"safe, reversible\" step; the user still approves tool calls in Telegram. Its pre_llm_call/post_llm_call hooks see every session's turns and store the bound chat's activity timestamps in plugin state; optional cross-bot export adds the separate storage disclosed below. On first load it reads $HERMES_HOME/companion/proactivity/proactivity.sqlite3 once, read-only. The plugin itself makes no network calls. The workspace-setup skill has the agent fetch this repo at the catalog pin into ~/alans-way-agents (or reuse a clean clone there) and hand a JSON setup plan to the human in Alan's Workspace. The app runs setup.sh through an external SSH process outside the gateway, checks the exact repoRef and desktopRef and rejects tracked checkout changes; the agent must never run setup or even --verify in its gateway terminal or bypass the lifecycle guard. Windows guests require the human's external Git Bash shell. The human pastes the app's output for review. The external setup does the following. (1) Installs long-running services (Alan's Tools: a managed Chrome/Chromium on a loopback-only debugging port, the first free port from 9223 up, plus a browser broker and an optional availability watcher): systemd units hermes-alans-way-chromium, -browser and mac-watch with loginctl enable-linger where systemd is running; supervisord programs alans-way-chromium, alans-way-browser and alans-way-mac-watch, written to a conf.d file, where it is not (VM images and containers); a KeepAlive LaunchAgent on a macOS guest; logon Scheduled Tasks on Windows. When no X display is running it also installs and runs Xvfb on :99 (unit hermes-alans-way-xvfb or program alans-way-xvfb, -nolisten tcp). (2) Installs packages when they are missing, only as root or with passwordless sudo, through apt, dnf or yum: Google Chrome stable (the current .deb downloaded from dl.google.com on amd64 apt hosts, which also adds Google's apt repository, so the host keeps taking Chrome updates from Google), else the distro chromium package, else Playwright 1.49.1 Chromium through npx (root only); and xvfb. It only warns, and installs nothing, for xdotool, scrot, at-spi2-core and python3-gi, which the VM desktop path needs. (3) Appends the user's computer's public key to ~/.ssh/authorized_keys (administrators_authorized_keys on a Windows admin account) with a from= restriction to the Tailscale ranges 100.64.0.0/10 and fd7a:115c:a1e0::/48, rewriting an existing unrestricted line for the same key; pins its host key in known_hosts and adds a ControlMaster block to ~/.ssh/config for Tailscale hosts only (with a tailscale nc ProxyCommand when Tailscale runs in userspace-networking mode). (4) Adds a TimeoutStartSec=180 drop-in to the hermes-gateway unit. Hermes refuses plugin file changes while a gateway is running, so setup drain-stops that profile's gateway (SIGUSR1 under supervisord, else systemctl stop or hermes gateway stop), replays the refused plugin install or update, and starts it again.
- With --restart it restarts the gateway only if setup changed something the gateway loads (or an earlier restart never happened), detached about ten seconds after printing its summary: by SIGUSR1 drain under supervisord, else a transient systemd-run timer or hermes gateway restart. (5) For the main profile and every profile with its own Telegram bot, writes a managed mcp_servers.workspace_browser block with connect_timeout: 25 seconds (removing other entries that run this router, with a backup; a re-run keeps the computer address and OS it already holds, and --mac-ssh none removes them), turns off the built-in browser toolset for Telegram and cron (--keep-browser only skips disabling it; it does not re-enable a previously disabled toolset), turns computer_use on for Telegram and cron, and sets computer_use.backend to alans-way-computer when it installed or found that provider (under the skill's --skip-plugin flow it only warns and prints the config command). It reads each profile's TELEGRAM_BOT_TOKEN only to derive the numeric bot id. It warns, and changes nothing, when a copy of the workspace-operations skill in a profile's skills directory shadows the plugin's. The workspace router reaches the user's computer over SSH on their tailnet; on the user's computer it calls the local API of Alan's Workspace on 127.0.0.1 with that app's bearer token. The skill's moved-VM checklist has the agent read the app's preferences.json on the user's computer over the same SSH path and tell the human where the VNC password lives without pasting it. The Alan's Workspace setup page has the human run the alans-way connect script from its main branch on their own computer, and the agent may install Tailscale with its official install script. The manual --allow-desktop-actions setup flag, which the skill offers but never adds itself, adds cua:*:background entries to command_allowlist. On a Hermes without the computer-use provider API, including 0.21.5, it also exposes workspace_computer_action, which drives the user's desktop, Terminal included, with no approval. The repo also holds bootstrap-orgo.sh and migrate.sh, which the plugin, setup.sh and the skills never run: a user runs them by hand to provision a fresh supervisord Linux computer (it fetches the Hermes and Tailscale installers over HTTPS, pipes setup.sh from this repo's main branch into bash unless --repo-ref is given, installs openssh-server if absent, writes supervisord programs, and POSTs the Tailscale login URL with a shared secret to a --callback URL when one is given) or to copy ~/.hermes, including .env, auth.json and state.db, to another host over SSH (it also looks up each profile's bot username at api.telegram.org with that profile's bot token). Since 0.21.0, every enabled profile also gets a default-on Telegram texting voice: an additional pre_llm_call hook reads that profile's voice mode and adds a tone note to Telegram turns, and transform_llm_output can remove a known leading assistant opener from the final Telegram reply. Messages containing backticks, quotes, indentation or silence markers are left unchanged by that polish. /tone texting|default|status and hermes tone store the mode under voice in Hermes plugin state; setup.sh --tone applies it to each configured profile. It layers over the user's persona without writing SOUL.md or agent.system_prompt. Check-ins now name one concrete open thread. /proactivity now requests one immediately from the bound chat even when paused or outside active hours, without changing the settings or unanswered-check-in count. Morning briefs are off by default: /proactivity brief on (or the operator CLI) enables one daily internal user-role prompt near active_start, with 0-20 minutes of jitter, at most two hours late and skipped after recent user activity. It asks the model to consult calendar and email read-only if those tools exist. The existing proactivity tool may disable the brief but cannot enable it; settings.morning_brief and last_brief_date are stored in the existing idle plugin state. The workspace router adds a commit object to
- the existing browser open/action tools for purchases, bookings, sends, posts, submissions, deletions and account changes. It checks control names, selectors, refs, action-looking URLs, batches, scripts, keys and dialogs against page state; missing page knowledge needs a snapshot or explicit commit. Shared commit-name patterns allow benign Buy it again, Book a demo and Subscribe & Save labels without excusing other commit-bearing words. Tax schedule names are treated as benign labels without exempting one-on-one bookings from approval. It asks through MCP elicitation (60-second default wait, configurable with HERMES_ROUTER_APPROVAL_MS), refuses unavailable, declined, timed-out or stale approvals, drops agent-supplied approval receipts and adds a single-use receipt bound to the approved tab and epoch. An approved commit is never replayed after a lost connection or host change. It appends commit decisions to commit-decisions.jsonl beside the mac-watch state file, creating the log with mode 0600: timestamp, bot id, page origin, decision and any supplied kind, summary, amount, currency and counterparty, or the detected target. Page/control/challenge and pending-approval state is held in memory. The same two skills now prefer Hermes web_search/web_extract for plain lookups and the user's computer for interactive search, hand CAPTCHAs to the human, poll at most every 30 seconds for up to 30 minutes, and switch sources after two challenges in ten minutes. The router stops remembering challenge pages for failover and reports changed tab ids in result metadata. The operations skill covers the app's automatic takeover after its grace period; CAPTCHA and recording holds and Take-back cooldowns still keep the bot out. It asks the agent to save only lasting general facts and preferences to memory; procedures belong in skills and project/task records in the user's chosen file or tracker, asking where if none is chosen. Existing memory is not migrated or deleted, and live tab state is never saved there. The setup skill pins the companion app source to f110115ff23367e4e21eeaf00e658a8706369652 and adds a long-chat reference: report effective compression and summary routing read-only, then only with explicit consent for each profile enable compression.proactive_prune_tokens at 48000, with 8000-character eligibility and 4096-token minimum reclaim, using the profile-scoped config CLI and keeping custom values, micro_compact and summary routing. This writes the approved profile's config.yaml and records prior values outside the repo for undo; setup.sh does not run it. A weekly time-saved report is a user-created Hermes cron recipe, not an installed job. setup.sh --verify now checks the running service manager, display, SSH ProxyCommand and managed browser's loopback /json/version endpoint, and reports noVNC reachability from its actual bind address. Manual supervisord restart guidance drains with USR1 first. The repo also adds scripts/voice_eval.py, user-run only: it sends fixed fictional scenarios and replies to Anthropic models through the SDK or claude -p for scoring and optionally writes a JSON result; the plugin, setup and skills never run it. Since 0.21.5, the plugin registers two more bundled skills, web-errands and integrations, with no new tools, hooks, commands or required environment variables. web-errands is guidance for shopping, booking, applications, filings, cancellations and claims in the workspace browser: every purchase, booking, send, submission and deletion still goes through the commit approval gate described above, CAPTCHA, sign-in and card-entry walls go back to the human, and card numbers are never stored. It has the agent record each application's company, role, date, posting URL and confirmation in the user's chosen project/task record, asking where if none is chosen. For a price, restock or status watch the user asks for or explicitly accepts, it has the agent create a Hermes cron job with Hermes's cronjob_manage tool (or ask the operator), polling no more often than every 10 minutes and stopping after three consecutive
- failures; cron runs cannot answer approvals, so a watch that reaches its target stages the result for the user. Watches use attach_to_session: true with no deliver override and continuity: true; each check is a fresh session, uses its prior completed output as a baseline and replies [SILENT] unless something meaningful changed, failed or reached its target. An optional safe read-only deterministic monitor URL or script can skip the model and delivery on unchanged output; probe errors still need reporting/backoff. Cron cannot pause/remove itself unless the operator enabled cron.allow_agent_scheduling. In-chat monitoring is a separate user-entered /heartbeat every command, one per conversation replacing the previous one; it runs only while idle with the chat process/gateway alive, costs a model turn per check and needs /heartbeat clear from the user. The integrations skill has the agent check that an account is connected, with a small read, before promising a task that needs it, and reuses Hermes's Google Workspace, Himalaya, Home Assistant, Honcho and X routes first. The Google route uses Hermes's own setup, whose fixed scope list includes write access; the skill says so and stops if the user declines. Only after the user agrees in chat does it guide an install: a Hermes catalog plugin (Home Assistant or Honcho), a separately reviewed Himalaya v1.2.0 binary, rclone v1.75.2 (the release zip from downloads.rclone.org, checked against that release's SHA256SUMS before extraction) or Taskwarrior 3.5.0 (a source build of upstream commit 3419d5ba1e6a780fbbb0a8d68d8fd31e675c1bf8 cloned from GitHub, after checking the CMake and Rust versions and asking before changing any toolchain), the last two into versioned directories under ~/.local/share/alans-tools/. Secrets are entered by the human in a private terminal outside the agent session, never in Telegram; where a recipe needs a secret in the agent's terminal (a mail credential, RCLONE_CONFIG_PASS), it explains the grant and adds only that variable to the profile's terminal.env_passthrough. The recipes keep rclone.conf (mode 0600) and Taskwarrior's taskrc and data under the profile's HERMES_HOME/integrations. Scheduled reads are read-only cron jobs with a profile-local failure-count file, and pausing goes through the owner conversation or an explicitly authorized hermes cron pause. API writes (send, post, delete, share) need fresh explicit approval in the live conversation, and a declined, timed-out or unavailable browser commit is never completed through another tool. Mastodon is not bundled: its recipe is human-run documentation in this repo (docs/integrations/mastodon.md) that the plugin and skills never run. None of this runs at plugin load or from setup.sh. Telegram replies, check-ins and briefs share one status-first style: concrete status, evidence or caveat, plain limits, and no tool names, file paths, scripts or job and session IDs in user-facing text unless asked. Prices, availability and policies must come from a page read this turn or be labeled unconfirmed. Multi-step browser work continues in short chunks until done or blocked, with one short start line and one-line updates only at milestones, routed-around failures or noticeable waits; routine reads, actions and transitions stay quiet, a new mid-task message is acknowledged and folded in, and Hermes's clarify buttons may be used for decisions the user owns, never for approvals; a whole checkout batch still sits under one commit. For a message of 200 characters or fewer that is wholly a recap request, greeting, thanks or acknowledgment, the voice hook appends a three-sentence limit; the check reads only the current message, with no model call, history or network. The check-in menu now ends with asking about a decision the user owns. The operations skill adds tab hygiene: the agent closes finished tabs it owns with the existing close tool, never another bot's or the user's and never one with a pending handback or recording hold. The router adds requestingBotId to tab-list replies and forwards a reply over 4 MiB unchanged,
- without that decoration, writing a content-free line to stderr. The skill also tells the agent to report installed versions by running hermes --version, reading the plugin's install metadata, running local git describe and rev-parse on the Alan's Tools checkout, and calling the app's existing status tool. setup.sh --bind now also makes the bound DM Hermes's Telegram home by running hermes -p PROFILE config set for TELEGRAM_HOME_CHANNEL and TELEGRAM_HOME_CHANNEL_THREAD_ID (30-second limit each; on failure it warns to send /sethome); it loads on gateway restart. setup.sh --verify exits 0 for a complete local audit, 1 for failed required checks and 2 for a partial install (warnings, or skipped browser or services). The setup skill's checkout step now initializes ~/alans-way-agents if absent and fetches the pin (depth 1) from the repo's GitHub origin, or reuses a clean clone there after two read-only local git helpers (5-second limit) confirm one recorded and effective origin URL naming this repo; for file and direct Git installs it uses the installer-recorded revision, and it never falls back to a tag when that record is invalid. Its computer-use API probe separates an absent API from a failed or timed-out probe and ends the probe's process group or tree on timeout. The setup skill also offers, each only after an explicit answer, with readback and a remembered-choice marker in the profile's config: one grouped autonomy choice for display.busy_input_mode: steer and agent.intent_ack_continuation: true, for every platform in that profile after restart. It reads effective values, including defaults, proposes changes only from stock interrupt/auto and preserves deliberate choices, custom or unknown values and unreadable config. General setup/proactivity consent is insufficient; an explicit no preserves both settings, silence writes nothing. Both keys are read back, including unchanged keys, before writing and reading back skills.config.workspace-setup.autonomy_choice as chosen or keep-current. Failed writes/readbacks leave the marker untouched; it records the answer only and never restores settings. Legacy busy_input_choice: keep-interrupt skips only the busy key; unknown legacy state is reported and skipped. agent.verify_on_stop is left alone unless the coding user requests stricter checks and one grouped Telegram question covering platforms.telegram.typing_indicator, display.platforms.telegram.interim_assistant_messages, tool_progress, notifications, long_running_notifications and busy_steer_ack_enabled, platforms.telegram.reactions, display.memory_notifications (hides the Memory updated and Self-improvement review notices on every platform in that profile), cron.wrap_response (profile-wide: removes scheduled-message headers, job IDs and footers for all delivery targets) and, as a separate city-confirmed question, timezone. Markers written through the same CLI are skills.config.workspace-setup.autonomy_choice and telegram_feel_choice; the former busy_input_choice is read but no longer written; a legacy busy_steer_ack_choice marker is only read, never written. Silence and non-interactive runs write nothing, and neither setup script writes any of these keys. When the VNC viewer binds loopback only, the skill can, with explicit consent, have the human run a temporary SSH forward to the computer's loopback (local port 16080, ends after 30 minutes) and verify it with curl from that computer. For a manual or file install, SAFE requires no scanner confirmation; setup and dependency consent remain separate. A later CAUTION revision needs review and explicit consent to the exact tree before retrying hermes plugins install with --ref SHA --enable --force; --enable alone does not confirm a warning, and --force confirms CAUTION and can replace an install but cannot override DANGEROUS. Hermes's unmodified scanner reports SAFE for this plugin: 0 HIGH, 2 MEDIUM (python_subprocess: the two git helpers in skills/workspace-setup/SKILL.md) and 3 LOW (hardcoded_ip_port: the loopback forward examples in the same file). This pin also extends the existing
- texting-note and minimal final-opener polish to public session source tags desktop, tui and bot_room, alongside Telegram; CLI, cron, webhook and API-server sources keep their default voice. A tone-independent pre_llm_call task note asks the agent to track and verify mini goals with Hermes's todo tool or working context, continue independent authorized work through completion, recheck historical blockers against newer messages and verify every requested criterion with each item's own source before replying. It requests a short start line and milestone-only updates, keeps routine work quiet and preserves questions, approvals, failures and honest completion. Recap replies are capped at three sentences total including any detail offer. Follow-up cron jobs are saved only after the user asks or says yes, attached to this chat with attach_to_session: true and no deliver override; no later message may be promised without a saved job. A not-yet reply allows one offer at a named local time, with no saved repeat until consent, and jobs are removed when their reason ends. /goal is suggested once for a big multi-step job, not small asks. Fixed Hermes progress/error/approval messages are unchanged. Scripted tool conversations prove hook delivery; consistent real-model quiet behavior is unverified. Browser guidance starts normal work in the user's Workspace even while they are away; host vm is for explicit VM requests or router-reported outages, never login/CAPTCHA/approval errors. Existing VM tasks finish there when the computer returns; no third-tier browser is substituted. Reasoned human handbacks hold VM tabs as well as computer tabs for up to 30 minutes until Give back; a chat reply alone cannot bypass awaiting_human. Resume uses the host, tab ID and fresh epoch from one reply. The human-only Move button can carry URL/scroll and chosen drafts; cookies move only when a restore is reported, while passwords, uploads and in-memory state stay behind, and the agent has no move tool. Router changes bind approvals to exact observed page targets, reject missing/malformed action success receipts, bound recovery and owned-child shutdown, and project agent-facing replies without exposing internal approval proof. The router never replays a sent write after connector retry, timeout or host change, including writes before the first response; only reads and unsent calls can replay within their original remaining budget. Cancellation is forwarded and stale-host checks reject queued desktop input before sending it; uncertainty requires checking the outcome before any repeat. Browser and desktop replies add a content-free diagnostic, on stderr and appended to the profile's logs/workspace-router.log, with fixed tool name, serving host, default/request/fallback selection and reason, and elapsed time; it contains no URLs or page/task content. Idle direct SSH probes can discover a recovered computer sooner without moving an active VM task. Cross-bot proactivity adds the cross_bot_ask tool and pre_tool_call/on_session_end hooks; no middleware, required secrets or direct network destinations are added. Optional cross-bot proactivity is off by default, configured with the same explicit absolute shared root and designated primary in each participating profile, then enrolled per bot through /proactivity cross on (activation in a running gateway needs a restart). Only completed user/final-assistant exchanges from each enrolled bot's bound owner Telegram DM are exported: at most four recent exchanges and 2 KiB per digest, expiring after 24 hours per exchange. There is no historical backfill or export of groups, other chats, attachments, reasoning, tool history, cron, check-ins, briefs, child or imported turns. Exchanges that appear to contain credentials or payment details, or match operator redact terms, are skipped entirely before truncation; sanitization cannot detect every secret. Only the designated primary receives other enabled profiles' fresh digests during permitted check-ins, including user-requested /proactivity now; its own digest is excluded and exports
- are JSON-escaped as untrusted observations, bounded to 32 KiB across 16 profiles. Ordinary turns receive no digests or dispatch permit; separately consented helper results can arrive there as disclosed below. Briefs and cron receive neither. The primary may automatically schedule one read_only or draft_only cross_bot_ask with fresh target evidence per check-in into another enrolled bot's existing bound DM; replies appear there with primary attribution, and only separately consented short outcomes can return to the primary. The receiver needs its own gateway-injection permission and verified routing, observes pause/hours/busy settings, and starts no spontaneous secondary check-ins. Queued asks are at most 800 characters, expire after two hours and are deleted before injection; sent means scheduled, not delivered or completed, with no replay, delivery-receipt store or result polling. Result sharing is separately off by default, including for existing cross-bot users: /proactivity cross results on must be chosen in both bound owner DMs. Only future asks can return a sanitized Done/Blocked summary of at most 280 UTF-8 bytes, correlated to one ask, with the same two-hour expiry measured from ask creation. The private shared results.json stores up to 64 pending/running/completed records with profile identities, enrollment generations, ask IDs, timestamps, prompt/session-correlation hashes, status and summary. Results are consumed once before exposure at the primary's next ordinary owner turn or permitted check-in, as untrusted helper reports rather than independently verified facts. There is no wake, retry or delivery guarantee; a crash or helper restart can lose an outcome. /proactivity cross results off purges pending outcomes; cross off also revokes the choice. Copied outcomes can persist in chat/provider context after expiry or opt-out. Attempts are limited to four per rolling day gateway-wide, two per target and 60-minute target spacing. The plugin-owned POSIX shared directory uses mode 0700 and files 0600: digests and queued bodies, persistent enrollment/generation and eligibility/heartbeat records, body-free turn/attempt hashes retained for 48 hours, one-use check-in permits and imported-session locks expiring within 30 minutes. Session locks contain the exact session key and ask ID and survive restart until session end, expiry or opt-out. Pruning runs on activated startup, before store access and every 60 seconds while connected; offline expired files can remain until an activated profile prunes them. Dormant profiles do no cross-bot store access. /proactivity cross off immediately stops sharing, receiving and dispatch and purges that profile's digest, permits, locks and pending asks; body-free records remain until expiry and scheduler roles revert on restart. Only the operator can enable or expand sharing; the model may disable it. Launch-gateway profile routing policy is read through public configuration/home-scope APIs; no Hermes session DB or credential store is read for export. Imported sessions block terminal/code, desktop actions, delegation, messaging, scheduling, further cross-bot asks and unknown connectors; reviewed reads/drafting and existing gated browser actions remain allowed. Browser approval, CAPTCHA and ownership gates still apply; this is not universal read-only enforcement. Copied context goes through the configured model provider and target Telegram delivery and may remain in Hermes history, API-content sidecars, compaction and memory-provider records, including Honcho when configured, and model-provider retention after expiry or opt-out. The instruction not to save other bots' excerpts is model policy, not enforced deletion; the plugin performs no memory-provider writes. No new direct network endpoint or telemetry is introduced. Optional integrations use the user's selected Google Workspace API, IMAP/SMTP mail service, Home Assistant instance, Honcho API, X API or rclone cloud/WebDAV/S3 endpoint; browser errands contact user-selected sites. Honcho can receive conversation turns, derived conclusions,
- peer/session IDs and version/OS/plugin headers. Native tools retain their own profile-scoped credentials, including Google client/token/pending OAuth files and encrypted rclone provider tokens; the plugin does not create a separate credential store. Setup downloads also reach GitHub/raw.githubusercontent.com, the official Tailscale installer, configured distro package repositories and the npm/Playwright distribution, in addition to Google and rclone downloads already disclosed."
- maintainer: capthvnsen
- tier: community
- category: automation
- requires_hermes: ">=0.21.5"
- docs_url: https://github.com/capthvnsen/alans-way-agents/tree/0f65549ac1353f6b74b8ac6a22d58f3eb6dea0a3/alans-way
- readme: true
- platforms: [linux, macos]
- known_issues:
- - "Install this on the machine that runs Hermes (a Linux host or a macOS guest VM), not on the user's primary computer. Windows guests are new and untested on real hardware. If the plugin is already installed from the catalog, setup leaves it as it is, so this pin stays."
- - "Setup turns off Hermes's stock browser toolset for Telegram and cron so the agent uses the workspace browser (pass --keep-browser to skip disabling it; this does not re-enable it). It never turns off computer_use: on Hermes with the computer-use provider API, desktop input goes through the gated computer_use tool and its alans-way-computer provider."
- - "Check-ins need plugins.entries.alans-way.allow_gateway_injection: true and a bound Telegram DM (setup does both). A background thread checks every 60 seconds once the gateway connects Telegram."
- - "Setup may install Chrome, Chromium, Playwright's Chromium or Xvfb with root or passwordless sudo, and may drain-stop and start the profile's gateway to apply a plugin change Hermes refuses while it runs. Pass --skip-services to skip the services and their packages."
- - "The computer's SSH key is only accepted from the tailnet (from= restriction). An existing unrestricted line for the same key is rewritten with it."
- - "A copy of the workspace-operations skill under a profile's skills directory shadows the plugin's and never gets its updates. Setup only warns; move anything worth keeping out of it, then delete it."
- - "The texting voice is default-on for Telegram, desktop, TUI and bot_room source tags. /tone default disables its note and final-opener polish per profile. CLI, cron, webhook and API-server sources retain the default voice. Check-ins and morning briefs keep their own short style. Fixed Hermes progress/error/approval messages, direct sends and cron reports are outside these hooks; interim messages get the tone note but no polish, and streaming can show raw chunks before a final edit. Consistent real-model quiet behavior remains unverified."
- - "Morning briefs are opt-in (/proactivity brief on), use the existing bound DM and gateway-injection permission, and may read calendar/email through tools already available to the model. /proactivity now is user-only from that chat and bypasses pause/hours for one requested check-in."
- - "Telegram browser Approve buttons need Hermes v0.21.6 or newer including commit 239740a19e (first in date tag v2026.6.19), despite the generic >=0.21.5 install floor. Without MCP form elicitation support commits fail with approval_unavailable; unattended cron declines them. The gate protects against agent mistakes and malicious pages, not a deliberately hostile bot with same-user terminal/file access."
- - "The router appends approval details and detected targets to commit-decisions.jsonl beside mac-watch state, shared by routers using that path; no retention or rotation is implemented. Keep this local log private. CAPTCHA handback, recording holds and Take-back cooldowns still block takeover. If the router reports human_has_control before a commit approval, wait or claim the tab through the normal app control flow first."
- - "Long-chat pruning is an agent-guided, separately consented config option per profile, not a setup.sh default. It preserves customized values and summary routing, rewrites sent history and can break the prompt-cache prefix; undo affects future pruning and cannot restore tool bodies already pruned from active context."
- - "web-errands and integrations are agent guidance, not setup steps. Any Himalaya, rclone, Taskwarrior, Home Assistant or Honcho install happens only after the user agrees in chat, never from setup.sh or plugin load. Taskwarrior is built from source (CMake >=3.24, Rust >=1.91.1 and several GiB of temporary space). The Google Workspace route requests Hermes's fixed scope list, which includes write access."
- - "Guided setup can change a profile's busy-input, intent-ack continuation, Telegram display, notification, reaction, memory-notice, cron-wrapping and timezone settings, but only after explicit answers. Busy mode, intent-ack continuation, memory-notice hiding, cron.wrap_response and timezone reach beyond Telegram within that profile; the autonomy choice takes effect after restart and preserves existing deliberate choices. Silence and non-interactive runs write nothing; neither setup script writes these keys."
- - "setup.sh --bind also sets TELEGRAM_HOME_CHANNEL for the bound profile through hermes config set and clears TELEGRAM_HOME_CHANNEL_THREAD_ID; it loads on gateway restart. If that fails, send /sethome in the primary DM."
- - "Desktop control needs a Hermes with tools.computer_use.backend.ComputerUseProvider (commit 63f5e0ea45726c5d8d7d207d140613997730aab3 or newer); a version string alone does not prove it. setup.sh --verify exits 2, not 0, for a partial install."
- - "This pin needs the paired workspace/remote browser host with approval-binding-v1 support for approved browser actions. Older hosts fail closed when that proof is unavailable; update and verify the paired workspace before rollout."
- - "Cross-bot proactivity is off by default and requires operator configuration and per-bot /proactivity cross on; activation needs a gateway restart. It exports bounded completed owner-DM excerpts across enabled profiles on one gateway. /proactivity cross off stops sharing/receiving/dispatch immediately and purges local shared text and permits; body-free hashes remain for 48 hours. Sanitization skips credential/payment-looking exchanges but cannot guarantee secrecy. Digest TTL is 24 hours, queued asks two hours; offline expired files remain until an activated profile prunes them. Copied context in Hermes history, sidecars, compaction, memory-provider records and model-provider retention is not erased by expiry or opt-out. The no-memory instruction is model policy. Shared storage requires POSIX owner-only directories/files. Dormant profiles do no store access."
- - "During a permitted primary check-in the model may automatically schedule one read-only/draft-only ask into an enrolled target bot's existing owner DM, with primary attribution. sent means scheduled, not delivered or completed; there is no delivery receipt or replay. A separate off-by-default results opt-in in both owner DMs permits one sanitized summary (280 UTF-8 bytes) per future ask, held in private results.json for at most two hours from the ask and consumed once at the primary's next owner turn or permitted check-in. There is no wake/retry; missing means unknown. Result opt-out purges pending outcomes and cross opt-out revokes the choice; already copied chat/provider context remains. Imported-session restrictions last at most 30 minutes and survive restart; terminal/code, desktop input, delegation, messaging, scheduling and unknown connectors are blocked, while existing gated browser calls remain possible. Existing approval/CAPTCHA/ownership gates apply; this is not universal read-only enforcement. Targets require gateway-injection permission and verified same-bot routes; missing permission, routing, busy/pause/hours or expiry can prevent delivery."
- - "Admission validation passes with warnings: cross_bot_ask is declared but only registers for an activated primary cross-bot profile; the default isolated probe does not activate it. proactivity.py and voice.py read gateway.session_context.get_session_env, which current Hermes flags as private imports that may break across releases."
- - "0.21.9 router changes: when the computer's ssh link drops and mac-watch reports it online, the router reconnects to it at most once a minute, re-initializes and replays only read-only and never-sent calls; an action that may already have run is answered as uncertain, never repeated. Only a failed reconnect or an offline/stale computer fails over to the VM, and after a failover a call that names a computer tab is refused rather than served by the VM. The approval gate also learns page names from a tab-open snapshot. A pre_llm_call note on every turn names the workspace tools and routine Mac moves."
- capabilities:
- provides_tools:
- - proactivity
- - cross_bot_ask
- provides_hooks:
- - pre_tool_call
- - pre_llm_call
- - post_llm_call
- - on_session_end
- - transform_llm_output
- provides_middleware: []
- requires_env: []
|