Skip to content

Built-in system context

Every agent run through the shared gateway wiring receives a turn-start clock snapshot and a bounded host summary, including Dashboard, Telegram, CLI, scheduled runs and delegates. No agent-file edits or approval grants are needed.

  • system.time reads the clock and owner profile again on every call.
  • system.info returns the same time information plus OS/version, kernel, architecture, hardware model when detectable, and host timezone. Host facts are cached for five minutes; unknown hardware or virtualization is not guessed.
  • The owner’s zone is the one in Settings → Profile. BUDDI_TZ only fills an empty profile at start and is the fallback while the profile names none (or is unreadable), then New York. The same zone drives the turn clock, {{today}} in agent prompts, tools, widgets (“here” on the World clock), reminders, digests and the dashboard’s footer and lock clocks.
  • A change in Settings → Profile (or through owner.set_profile, or MCP’s buddi.profile_update once approved) applies at once, with no restart: the server keeps the profile’s zone in memory, updates it on every save and re-reads it every minute. The three check a change with the same rules (owner-profile-edit.ts), and every field — name, full name, pronouns, language, birthday, about, formats, places — reaches every agent’s context at its next turn, since the block below is built fresh for each turn.
  • Every schedule records whether its zone was named on purpose (core.schedule_specs.timezone_explicit). One made without a zone — the recap, the learning digest, a plugin’s mission that names none, an agent’s schedule.propose without timezone — follows the owner’s zone: a Profile change moves it as a new schedule revision, so its next run is 8 AM in the new zone. One whose creator named a zone (the tool’s timezone, a mission that declares one, a zone chosen through POST /api/missions/:id/schedule) keeps it. At every start, schedules that follow the owner and sit in another zone move to it. Schedules from before the flag are settled once at start, by provenance: one buddi itself made without a zone (a plugin’s default mission that names none, the learning digest, the first-run arc, a starter’s or plugin agent’s declared mission, matched by mission id and cron) that sits in the default zone of the time (BUDDI_TZ, else New York) or the Profile’s follows the owner; any other, schedule.propose and dashboard ones included, keeps its zone, since nothing recorded whether it was named. Missions on the dashboard say “follows your timezone” instead of the zone; buddi missions adds it to the schedule line. A one-off reminder is an instant and does not move.
  • buddi doctor (a checkout’s table and a packaged install’s Timezone line) names the zone and where it comes from — Settings → Profile, BUDDI_TZ, or the default — and this machine’s own zone when it differs.
  • How the owner reads times and dates (Settings → Profile: 12-hour or 24-hour, “Thu, Oct 1”, “Thursday, 1 October” or ISO) is one line under “About the owner” for every agent, so a reply writes “2:05 PM” to someone who reads 12-hour time. Auto adds nothing.
  • Who the owner is — full name, pronouns, birthday (said as today, or “in N days” within a week, with the age when the year is set) — is in the same block, from Settings → Profile.
  • The front desk alone is also told the owner’s places — each label, the address as typed, the town it was matched to and its zone — as context, the way it is told the timezone. Other agents are not; a plugin reads them only through its declared owner:places.

Every run’s context carries three lines beside the try-first rule (GROUNDING_LINES, “Read, never recall”): anything about today, the news, mail, calendar, money, prices, or anything a tool or a colleague can read is read or delegated, never answered from memory; no source, figure or quote the conversation did not read is named, and what was read is attributed to where it came from; when nothing can reach it, the agent says so in one line. The native-search paragraph (packages/runtime/src/search.ts) keeps its own, search-specific “cite as you go”.

The runtime checks it (packages/runtime/src/grounding.ts). When the final answer of a turn that read nothing — no tool call or delegation that came back successfully (an unknown, refused, truncated or failed call reads nothing), no native search, not a decided approval coming back — cites sources, it is not delivered: not stored, not sent, and its streamed words are withdrawn (onRetract; the web chat gets live.settle with retracted: true and drops the live text at once, Telegram replaces the streamed message with “Checking that…” after any edit in flight, then streams the retry into it). The model gets one turn, not kept in the transcript: “You cited sources without reading anything. Verify with your tools or a colleague, or remove the claims.” If that answer still reads nothing, it is delivered with unchecked: true on the run’s result and run.finished, and the surfaces draw “Answered from memory, not checked” as a quiet line under it. Each firing writes a run.grounding event (stage: retried, then unchecked when it comes to that), visible in the activity log.

“Cites” is deliberately narrow, since a false alarm is worse than a miss: an outlet from a known list (CBS, AP, Reuters, Le Monde…) or a domain counts only in a citation position (in parentheses, after “according to” / “selon” / “d’après” / “per” / “via”, before “reports” / “said” / “writes”, or as a link); a capitalised name after “according to” / “selon” / “d’après” counts too. A bare URL outside a citation position is not a source; it only backs up numbered markers or a “Sources:” line. An outlet and its own domain (“Reuters”, reuters.com) are one source. It fires on two distinct sources, or one (or a bare link) with numbered markers ([1]) or a “Sources:” line. Sources the conversation already holds (the owner named them, an earlier tool returned them) or the agent carries (its persona, its memory preamble, the platform context) are not counted; code and email addresses are ignored; figures alone never count.

Every prompt, with or without a run, carries one line (REPLY_LANGUAGE_LINE): answer in the language the owner’s message is written in, unless they ask for another; when the message is too short to tell, in the profile language, else English; documents, tool results, colleagues’ answers and summaries in other languages never change the answer’s language, they are quoted as they are. The profile’s language, when set, shows under “About the owner” as the fallback for a message too short to tell, and rides on the platform context as language for the runtime. When it is unset, owner.profile_gaps lists it, so the front desk asks it once like any other gap.

The runtime checks it (packages/runtime/src/language.ts), with no model: a stopword count for English, French, Spanish, German, Portuguese and Italian. Code, links, blockquotes and complete quoted passages (any length, line breaks included) are not counted; a word several languages share counts for each by a fraction; a language is named only with three hits at least (two for the owner’s message), twice the runner-up’s score, and a tenth of the words. The owner’s message is read from its opening (the text before its first blank line, at most forty words), so an instruction over a pasted article counts, not the article; and only the owner’s own words: a mission passes its prompt (ownerText), a reminder its text, not the English instructions buddi wraps around them. When the final answer of a turn is confidently in a language other than the anchor (the message’s language when it can be told; else, for a message under eight words, the profile language), it is held back the way the grounding guard holds an answer back (not stored, not sent, streamed words withdrawn through onRetract) and the model gets one turn, not kept in the transcript: “Answer in .” (the message’s language, else the profile’s). Whatever it answers then is delivered, without a flag; an empty rewrite delivers the held-back answer. One retry per turn across this guard and the grounding guard: after a grounding retry the language is not checked, and a language rewrite that cites unread sources is delivered flagged unchecked rather than sent back again. Each firing writes a run.language event (stage: retried, or kept when the run had no budget left for a rewrite; reply and target).

It never fires on a reply under 12 words or one that is mostly code, when the owner’s message cannot be told and no profile language buddi can read is set, when the owner’s message (or one of their last few) tells buddi the answer language (“answer in French”, “a Spanish summary”, “en français s’il te plaît”; not “French politics”) or asks for a translation, in a delegate’s run (its answer is quoted into its caller’s), or on a decided approval coming back.

When the turn’s opening words name a headline from today’s edition — case and accents ignored, at least 70% of the headline’s words in order, or five words of it (the whole of a shorter one) word for word — the context gets one more line naming the story, the agent that delivered it and its plugin (“a story from today’s edition by @anchor (News)”), and how to read it: the agent’s own tool of that plugin, else a delegation to that agent, else the handle to name. The edition is found in core’s own record, in one query: a mission.report whose link names a saved edition (#/p/<plugin>/…?edition=…) in a run with a mission.delivered event today. A message of a word or two does not query at all.

Host means the machine/environment running the Buddi server, not the browser or Telegram client. macOS product version and hardware model identifier are read using fixed, bounded system commands. Windows/Linux report available OS facts; hardware model can be null. A hardware identifier is not guessed into a retail name.

No hostname, username, serial number, credentials, environment dump or private filesystem inventory is included. These facts confer no browser, shell or file access. Use separately granted status tools to check live host/computer permissions.

The two platform tools are separate from configurable agent grants. Their schemas are appended by the shared runtime, and the actual tool set is recorded in run events. The time context is authoritative over older dates in agent personas/history; long-running tasks should call system.time again.