Pick any line to see why it is the way it is: the commit, the pull request and issue it came from, and what the agent was thinking.
| Chat and workspace agents: channels, DMs and named agents you talk to | 1 | /** |
| The docs folder is gone, and what it held lives where people read it: how a self-hosted g1t runs and how to deploy g1t to Cloudflare are pages on docs.g1t.sh under Run g1t yourself, and speed, rate limits and operating g1t.sh are sections of CONTRIBUTING.md; code that cited a file in docs/ now points to the page or section that covers it, or says what it means itself, and applied migrations and the runner images are left as they were. | 2 | * The agents service: a workspace's own agents |
| 3 | * (docs.g1t.sh/guides/agents/). Their definitions and every version of | |
| 4 | * them, the templates they start from, each agent's desk, and their replies | |
| 5 | * in chat. | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 6 | * |
| 7 | * `@g1t`, the platform's own agent, is not one of these. | |
| 8 | * | |
| 9 | * Reached through service bindings: `POST /rpc/<method>`, with snake_case | |
| 10 | * JSON (@g1t/contracts workspace-agents.ts). | |
| 11 | */ | |
| 12 | import { | |
| Cards you act on in chat; agents comment and review as themselves; names shown cleanly; commits on the calendar | 13 | type AgentCardAction, |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 14 | type AgentProposal, |
| 15 | type AgentRedraft, | |
| 16 | type DraftReply, | |
| 17 | MAX_PERSONAL_AGENTS, | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 18 | type AgentDelivery, |
| Cards you act on in chat; agents comment and review as themselves; names shown cleanly; commits on the calendar | 19 | type CardActionResult, |
| Routines run when something happens: a pull request ready for review or merged, checks or a deploy failing, an issue opened | 20 | type G1tEvent, |
| The Marketplace is where a workspace finds what to add, reached from Apps: the agent catalog of roles with who already does each job and Add to workspace; every integration with whether it's connected; g1t's own extensions, listed as Soon; and requests, where members ask and owners add or turn down, each told by a notification. Underneath, an extension is a manifest (publisher and tier, source and version from a tag, scopes, the domains its data goes to, its page on g1tusercontent.com, no price) and an install pins its version with an off switch, a budget and the audit log. Apps lists only what was added from the Marketplace, with pins; g1t's own pages stay in their sidebars, and AI Gateway joins Workspace's. The marketplace guide says how. | 21 | type InstallRequest, |
| 22 | type InstallRequests, | |
| Each agent has an Abilities tab: g1t's built-ins, always on within the asker's access; its computer, coming; each connected integration's actions one row each, read, import, comment or resolve, with a level for each, alone, alone when the person asked for it, ask first or never, and whose connection it runs on; and MCP servers an owner adds. Reading is alone, writing inside g1t follows today's choices, anything that leaves g1t asks first, and production deploys can't go above ask. The agents service enforces every level: ask first posts a card to allow or deny and parks a session until it's answered, a refusal names its rule in the transcript and the audit log, and a missing ability posts a request to owners. The agent abilities guide says how. | 23 | type McpServer, |
| Pricing says it plainly: models at the provider's price, the agent rate for what g1t runs around every model call (the gateway, secrets, routing, context and pass-through to your own provider, so your own keys too), everything else at cost plus 20%, your own runners free, no seats; Security and quality comes with the plan with no separate fee, and live activations end. Each agent has an effort setting, Auto to Max, with what a typical task has cost at each level, and Spend's Spend less, keep quality suggests a lower level only when the agent's own past work shows quality held, to apply or dismiss. The pricing, spend and agents guides say how. | 24 | type AgentEffortCosts, |
| An archived agent can be found, brought back, or deleted for good. The Agents page ends with Archived, folded, where owners see every archived agent and members their own personal ones, each with when and by whom it was archived and what deleting it would take; Restore brings it back, under a new handle when a live agent has taken its old one, and Delete permanently asks for the handle to be typed and then removes its sessions, replies, memories, routines, versions, drafts, skills and its computer's home, leaving its messages in chat under its saved name, the workspace's spend totals and the audit log. In Chat an archived agent is deactivated the way people expect: never offered as someone to message or mention, never woken, a conversation with it kept in the list, dimmed and marked Archived, open to read with a note in place of the composer and a way to restore it for owners; after deletion the conversation still reads, its messages marked as a deleted agent's. The agents, chat and agent budgets guides say so. | 25 | type ArchivedAgent, |
| Pricing says it plainly: models at the provider's price, the agent rate for what g1t runs around every model call (the gateway, secrets, routing, context and pass-through to your own provider, so your own keys too), everything else at cost plus 20%, your own runners free, no seats; Security and quality comes with the plan with no separate fee, and live activations end. Each agent has an effort setting, Auto to Max, with what a typical task has cost at each level, and Spend's Spend less, keep quality suggests a lower level only when the agent's own past work shows quality held, to apply or dismiss. The pricing, spend and agents guides say how. | 26 | type AgentRecommendation, |
| 27 | type AgentRecommendations, | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 28 | type NewWorkspaceAgent, |
| 29 | type Result, | |
| 30 | type ServiceBinding, | |
| 31 | type User, | |
| 32 | type WorkspaceAgent, | |
| 33 | UNVERIFIED, | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 34 | askerAccess, |
| Chat and workspace agents: channels, DMs and named agents you talk to | 35 | awaitsConfirmation, |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 36 | chatClient, |
| The Marketplace is where a workspace finds what to add, reached from Apps: the agent catalog of roles with who already does each job and Add to workspace; every integration with whether it's connected; g1t's own extensions, listed as Soon; and requests, where members ask and owners add or turn down, each told by a notification. Underneath, an extension is a manifest (publisher and tier, source and version from a tag, scopes, the domains its data goes to, its page on g1tusercontent.com, no price) and an install pins its version with an off switch, a budget and the audit log. Apps lists only what was added from the Marketplace, with pins; g1t's own pages stay in their sidebars, and AI Gateway joins Workspace's. The marketplace guide says how. | 37 | cleanRequestNote, |
| Chat and workspace agents: channels, DMs and named agents you talk to | 38 | fail, |
| 39 | identityClient, | |
| 40 | newId, | |
| The Marketplace is where a workspace finds what to add, reached from Apps: the agent catalog of roles with who already does each job and Add to workspace; every integration with whether it's connected; g1t's own extensions, listed as Soon; and requests, where members ask and owners add or turn down, each told by a notification. Underneath, an extension is a manifest (publisher and tier, source and version from a tag, scopes, the domains its data goes to, its page on g1tusercontent.com, no price) and an install pins its version with an off switch, a budget and the audit log. Apps lists only what was added from the Marketplace, with pins; g1t's own pages stay in their sidebars, and AI Gateway joins Workspace's. The marketplace guide says how. | 41 | notifyClient, |
| Chat and workspace agents: channels, DMs and named agents you talk to | 42 | ok, |
| 43 | openD1, | |
| 44 | } from "@g1t/contracts"; | |
| 45 | ||
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 46 | import { MANAGE_REFUSAL, NOT_YOURS_REFUSAL, canArchive, canChange, canManage, canSee, canSeeAgent, creatableScope } from "./access.ts"; |
| 47 | import { DRAFT_OUTPUT_TOKENS, draftSystem, jsonIn, proposalFrom, redraftFrom, redraftSystem, startingBudget, trySystem, tryTurns, wordsOf } from "./builder.ts"; | |
| 48 | import { callBuilder } from "./builder-call.ts"; | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 49 | import { type Definition, applyChanges } from "./definition.ts"; |
| Each agent has an Abilities tab: g1t's built-ins, always on within the asker's access; its computer, coming; each connected integration's actions one row each, read, import, comment or resolve, with a level for each, alone, alone when the person asked for it, ask first or never, and whose connection it runs on; and MCP servers an owner adds. Reading is alone, writing inside g1t follows today's choices, anything that leaves g1t asks first, and production deploys can't go above ask. The agents service enforces every level: ask first posts a card to allow or deny and parks a session until it's answered, a refusal names its rule in the transcript and the audit log, and a missing ability posts a request to owners. The agent abilities guide says how. | 50 | import { listMcpTools } from "./mcp-client.ts"; |
| 51 | import { MAX_MCP_SERVERS, checkMcpName, checkMcpUrl } from "../../../packages/contracts/src/abilities.ts"; | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 52 | import { builtinChanges } from "./orchestrator.ts"; |
| Chat and workspace agents: channels, DMs and named agents you talk to | 53 | import type { Desk } from "./desk.ts"; |
| 54 | import type { ReplyEnv } from "./reply.ts"; | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 55 | import { ensureBuiltin } from "./builtin.ts"; |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 56 | import { type Row, definitionOf, insertAgent, isPersonal, periods, selectAgents, toAgent, updateAgent, versionStatement } from "./store.ts"; |
| Chat and workspace agents: channels, DMs and named agents you talk to | 57 | import { TEMPLATES, TEMPLATE_IDS } from "./templates.ts"; |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 58 | import { readPolicy } from "./policy.ts"; |
| 59 | import { runDue } from "./routines.ts"; | |
| Routines run when something happens: a pull request ready for review or merged, checks or a deploy failing, an issue opened | 60 | import { onEvents } from "./triggers.ts"; |
| @g1t welcomes you when you join a workspace, and that is the one introduction any agent makes. Identity now says when a person becomes a member, workspace.member_joined, by invitation, by an invite code, or by creating the workspace, and webhooks can hear it; the agents service hears it too, and for anyone who joined, never the creator, the workspace's @g1t opens your direct message and says in its own voice who it is, what to ask it, how agents answer, and where things are, once per person per workspace and charged to @g1t like any reply, with a fixed welcome when no model can be used. The agents, workspaces, authentication and webhooks guides say so. | 61 | import { welcome } from "./welcome.ts"; |
| Cards you act on in chat; agents comment and review as themselves; names shown cleanly; commits on the calendar | 62 | import { cardAction } from "./cards.ts"; |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 63 | import { type SessionEnv, sweep } from "./sessions.ts"; |
| Pricing says it plainly: models at the provider's price, the agent rate for what g1t runs around every model call (the gateway, secrets, routing, context and pass-through to your own provider, so your own keys too), everything else at cost plus 20%, your own runners free, no seats; Security and quality comes with the plan with no separate fee, and live activations end. Each agent has an effort setting, Auto to Max, with what a typical task has cost at each level, and Spend's Spend less, keep quality suggests a lower level only when the agent's own past work shows quality held, to apply or dismiss. The pricing, spend and agents guides say how. | 64 | import { EFFORT_NAMES, checkDue, effortCostsOf, markResolved, outcomesSince, readRecommendations, recommendationRow, sinceWindow, toRecommendation } from "./recommend.ts"; |
| An agent is on teams the way a person is: through team membership, with no team or department of its own. Add it to teams when you create it, from its profile or from the team's page; the Agents sidebar groups agents by the teams they're on, with those on none last; what an agent is told about its teammates, team budgets, skills reach, Spend's By team and the People pages all read memberships. Agents with an old team name matching a team are moved onto it once; the rest are left off a team. Templates name a role only. The agents, teams and people guides say how. | 65 | import { moveAgentTeams } from "./team-move.ts"; |
| Pricing says it plainly: models at the provider's price, the agent rate for what g1t runs around every model call (the gateway, secrets, routing, context and pass-through to your own provider, so your own keys too), everything else at cost plus 20%, your own runners free, no seats; Security and quality comes with the plan with no separate fee, and live activations end. Each agent has an effort setting, Auto to Max, with what a typical task has cost at each level, and Spend's Spend less, keep quality suggests a lower level only when the agent's own past work shows quality held, to apply or dismiss. The pricing, spend and agents guides say how. | 66 | import { effortOf } from "./routing.ts"; |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 67 | import * as views from "./views.ts"; |
| An archived agent can be found, brought back, or deleted for good. The Agents page ends with Archived, folded, where owners see every archived agent and members their own personal ones, each with when and by whom it was archived and what deleting it would take; Restore brings it back, under a new handle when a live agent has taken its old one, and Delete permanently asks for the handle to be typed and then removes its sessions, replies, memories, routines, versions, drafts, skills and its computer's home, leaving its messages in chat under its saved name, the workspace's spend totals and the audit log. In Chat an archived agent is deactivated the way people expect: never offered as someone to message or mention, never woken, a conversation with it kept in the list, dimmed and marked Archived, open to read with a note in place of the composer and a way to restore it for owners; after deletion the conversation still reads, its messages marked as a deleted agent's. The agents, chat and agent budgets guides say so. | 68 | import { computerCommands, computerView, eraseComputer, forgetComputer, keepComputer, resetComputer, sleepComputer, wakeComputer } from "./computer.ts"; |
| 69 | import { destroyRefusal, destroyStatements, listArchived, restoreHandle, restoreRefusal } from "./archived.ts"; | |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 70 | import { monthKey } from "./budget.ts"; |
| Merge branch 'customize' into main: Customize replaces the Marketplace, the Artifacts trash and spaces changes, the Security-free cleanup, with the desktop app's own title bar kept in the shell. | 71 | import { customizeRpc } from "./customize-rpc.ts"; |
| Agents → Skills is the workspace's skill library in the open SKILL.md format: write a skill in g1t, import a SKILL.md or zip, link a repository whose .g1t/skills folders publish on every push, or save a finished session as a skill once a person reviews it; attach a skill to an agent, a team or the whole workspace, each attachment pinned to a version. Agents see each skill's name and when to use it, and read the rest with use_skill when a request needs it; a skill never adds a permission, and scripts wait for an agent's own computer. The agent skills guide says how. | 72 | import { skillPushes, skillRpc } from "./skill-rpc.ts"; |
| The Marketplace is where a workspace finds what to add, reached from Apps: the agent catalog of roles with who already does each job and Add to workspace; every integration with whether it's connected; g1t's own extensions, listed as Soon; and requests, where members ask and owners add or turn down, each told by a notification. Underneath, an extension is a manifest (publisher and tier, source and version from a tag, scopes, the domains its data goes to, its page on g1tusercontent.com, no price) and an install pins its version with an off switch, a budget and the audit log. Apps lists only what was added from the Marketplace, with pins; g1t's own pages stay in their sidebars, and AI Gateway joins Workspace's. The marketplace guide says how. | 73 | import { type Answered, type Person, answerLine, findListing, listRequests, listingPath, openRequest, requestsPath, resolveListing, resolveRequest } from "./installs.ts"; |
| Chat and workspace agents: channels, DMs and named agents you talk to | 74 | |
| 75 | export { Desk } from "./desk.ts"; | |
| 76 | ||
| 77 | type Env = ReplyEnv & { | |
| 78 | IDENTITY: ServiceBinding; | |
| 79 | /** The audit log. */ | |
| 80 | EVENTS: ServiceBinding; | |
| Every agent can have its own computer. A session that needs one wakes it: a home of its own on g1t cloud, one per agent and never shared, where it runs commands, reads and writes files and keeps what it made, with each session working in its own folder under a shared home; after ten idle minutes it sleeps, its home kept as a snapshot and restored when it wakes, and Reset wipes the home while memory and artifacts stay. Its shell and files are abilities with the usual choices, Alone, Alone when asked, Ask first or Never, offered only inside sessions and never to a chat reply; every command shows on the session with its output, and the agent's new Computer tab shows the state, the disk used of the five gigabytes included, the recent commands, and Wake, Put to sleep and Reset. Machine time counts only while it is awake, on the sandbox lines of the ledger that name the agent and who asked, held to the same spend caps as the session; the disk itself costs nothing in this version. The runner gained a long-lived supervisor that answers the computer's requests inside the container, and the runner service a computer per agent that keeps its snapshot in the agent homes bucket when one is attached, and says so when none is. The REST API and the agent tool can read a computer, wake it, put it to sleep and reset it. The agents, abilities, sessions, runners, billing and deploy guides say how it works and what an operator sets up; pinning a computer to your own runner, its browser and take-over come next. | 81 | /** The runner, for agents' own computers (./computer.ts). */ |
| 82 | RUNNER?: ServiceBinding; | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 83 | DESKS: DurableObjectNamespace<Desk>; |
| 84 | }; | |
| 85 | ||
| 86 | /** Workspace ids by slug, kept a minute: every call names a workspace by slug. */ | |
| 87 | const workspaceIds = new Map<string, { id: string | null; until: number }>(); | |
| 88 | ||
| 89 | class Agents { | |
| 90 | private readonly env: Env; | |
| 91 | private readonly db: D1Database; | |
| 92 | private readonly defer: (work: Promise<unknown>) => void; | |
| 93 | ||
| 94 | // Plain fields, not parameter properties: Node's type stripping does not take those. | |
| 95 | constructor(env: Env, defer: (work: Promise<unknown>) => void) { | |
| 96 | this.env = env; | |
| 97 | this.db = env.DB; | |
| 98 | this.defer = defer; | |
| 99 | } | |
| 100 | ||
| 101 | private async workspaceId(slug: string): Promise<string | null> { | |
| 102 | const key = slug.toLowerCase(); | |
| 103 | const kept = workspaceIds.get(key); | |
| 104 | if (kept && kept.until > Date.now()) return kept.id; | |
| 105 | const workspace = await identityClient(this.env.IDENTITY).getWorkspace(key); | |
| 106 | if (workspaceIds.size > 5_000) workspaceIds.clear(); | |
| 107 | workspaceIds.set(key, { id: workspace?.id ?? null, until: Date.now() + 60_000 }); | |
| 108 | return workspace?.id ?? null; | |
| 109 | } | |
| 110 | ||
| 111 | /** The workspace's id, if the viewer may see its agents. */ | |
| 112 | private async seen(workspace: string, viewer: User | null): Promise<Result<string>> { | |
| 113 | if (!canSee(viewer, workspace)) return fail("not_found", "There is no such workspace."); | |
| 114 | const id = await this.workspaceId(workspace); | |
| 115 | return id ? ok(id) : fail("not_found", "There is no such workspace."); | |
| 116 | } | |
| 117 | ||
| Cards you act on in chat; agents comment and review as themselves; names shown cleanly; commits on the calendar | 118 | /** Internal, from chat: a person pressed an action on one of agents' cards. */ |
| 119 | async cardAction(a: AgentCardAction): Promise<Result<CardActionResult>> { | |
| 120 | if (a?.viewer && awaitsConfirmation(a.viewer)) return UNVERIFIED; | |
| 121 | if (!a?.viewer || !canSee(a.viewer, a.workspace ?? "")) return fail("not_found", "No such card."); | |
| 122 | return cardAction(this.env as unknown as SessionEnv, a); | |
| 123 | } | |
| 124 | ||
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 125 | /** What the views need: the workspace, the viewer, and whether they own it. */ |
| 126 | private async context(workspace: string, viewer: User | null): Promise<Result<views.ViewContext>> { | |
| 127 | if (viewer && awaitsConfirmation(viewer)) return UNVERIFIED; | |
| 128 | const seen = await this.seen(workspace, viewer); | |
| 129 | if (!seen.ok) return seen; | |
| 130 | return ok({ | |
| 131 | env: this.env as unknown as SessionEnv, | |
| 132 | db: this.db, | |
| 133 | slug: workspace.toLowerCase(), | |
| 134 | workspaceId: seen.value, | |
| 135 | viewer: viewer!, | |
| 136 | owner: canManage(viewer, workspace), | |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 137 | audit: (action, handle, message) => this.audit(viewer!, workspace, action, handle, message), |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 138 | }); |
| 139 | } | |
| 140 | ||
| 141 | /** Runs `view` with the context, or answers why it can't. */ | |
| 142 | async view<T>(a: { workspace: string; viewer: User | null }, view: (ctx: views.ViewContext) => Promise<Result<T>>): Promise<Result<T>> { | |
| 143 | const ctx = await this.context(a?.workspace ?? "", a?.viewer ?? null); | |
| 144 | if (!ctx.ok) return ctx; | |
| 145 | return view(ctx.value); | |
| 146 | } | |
| 147 | ||
| Chat and workspace agents: channels, DMs and named agents you talk to | 148 | /** The workspace's id, if the viewer may change its agents. */ |
| 149 | private async managed(workspace: string, viewer: User | null): Promise<Result<string>> { | |
| 150 | if (awaitsConfirmation(viewer)) return UNVERIFIED; | |
| 151 | const seen = await this.seen(workspace, viewer); | |
| 152 | if (!seen.ok) return seen; | |
| 153 | return canManage(viewer, workspace) ? seen : fail("forbidden", MANAGE_REFUSAL); | |
| 154 | } | |
| 155 | ||
| 156 | private async row(workspaceId: string, handle: unknown): Promise<Row | null> { | |
| 157 | if (typeof handle !== "string") return null; | |
| 158 | const now = new Date(); | |
| 159 | return this.db | |
| 160 | .prepare(selectAgents("a.workspace_id = ?3 AND a.handle = ?4 AND a.archived_at IS NULL")) | |
| 161 | .bind(...periods(now), workspaceId, handle.trim().replace(/^@/, "").toLowerCase()) | |
| 162 | .first<Row>(); | |
| 163 | } | |
| 164 | ||
| An agent is on teams the way a person is: through team membership, with no team or department of its own. Add it to teams when you create it, from its profile or from the team's page; the Agents sidebar groups agents by the teams they're on, with those on none last; what an agent is told about its teammates, team budgets, skills reach, Spend's By team and the People pages all read memberships. Agents with an old team name matching a team are moved onto it once; the rest are left off a team. Templates name a role only. The agents, teams and people guides say how. | 165 | /** |
| 166 | * The teams a new agent joins as it is made, by slug: each must be one the | |
| 167 | * viewer manages (an owner, or the team's maintainer), as adding anyone | |
| 168 | * to a team asks. Membership lives on the team, in identity. | |
| 169 | */ | |
| 170 | private async joinable(workspace: string, viewer: User, asked: unknown): Promise<Result<string[]>> { | |
| 171 | if (asked === undefined || asked === null) return ok([]); | |
| 172 | if (!Array.isArray(asked) || asked.some((slug) => typeof slug !== "string")) return fail("invalid", "Teams are a list of team slugs."); | |
| 173 | const slugs = [...new Set((asked as string[]).map((slug) => slug.trim().replace(/^@/, "").toLowerCase()).filter(Boolean))]; | |
| 174 | if (slugs.length > 20) return fail("invalid", "Add a new agent to at most 20 teams."); | |
| 175 | const identity = identityClient(this.env.IDENTITY); | |
| 176 | const found = await Promise.all(slugs.map((slug) => identity.getTeam(viewer, workspace.toLowerCase(), slug).catch(() => null))); | |
| 177 | for (const [i, team] of found.entries()) { | |
| 178 | if (!team?.ok) return fail("invalid", `${workspace} has no team called ${slugs[i]}.`); | |
| 179 | if (!team.value.can_manage) return fail("forbidden", `Only owners and ${team.value.name}'s maintainers add agents to it.`); | |
| 180 | } | |
| 181 | return ok(slugs); | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 182 | } |
| 183 | ||
| Chat and workspace agents: channels, DMs and named agents you talk to | 184 | private async handleTaken(workspaceId: string, handle: string, except: string | null): Promise<boolean> { |
| 185 | const found = await this.db | |
| 186 | .prepare("SELECT id FROM agents WHERE workspace_id = ? AND handle = ? AND archived_at IS NULL") | |
| 187 | .bind(workspaceId, handle) | |
| 188 | .first<{ id: string }>(); | |
| 189 | return !!found && found.id !== except; | |
| 190 | } | |
| 191 | ||
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 192 | /** |
| 193 | * The workspace's agents, and personal agents when asked for: `mine`, the | |
| 194 | * viewer's own; `all`, every member's for an owner. | |
| 195 | */ | |
| 196 | async list(a: { workspace: string; viewer: User | null; personal?: unknown }): Promise<Result<WorkspaceAgent[]>> { | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 197 | const seen = await this.seen(a.workspace, a.viewer); |
| 198 | if (!seen.ok) return seen; | |
| 199 | const now = new Date(); | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 200 | await this.ensureBuiltin(seen.value); |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 201 | const all = a.personal === "all" && canManage(a.viewer, a.workspace); |
| 202 | const mine = (a.personal === "mine" || a.personal === "all") && (a.viewer?.kind ?? "user") === "user" ? (a.viewer?.id ?? "") : ""; | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 203 | const rows = await this.db |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 204 | .prepare( |
| 205 | `${selectAgents( | |
| 206 | `a.workspace_id = ?3 AND a.archived_at IS NULL AND (a.scope = 'workspace'${all ? " OR a.scope = 'personal'" : mine ? " OR (a.scope = 'personal' AND a.owner_id = ?4)" : ""})`, | |
| 207 | )} ORDER BY a.builtin DESC, a.handle`, | |
| 208 | ) | |
| 209 | .bind(...periods(now), seen.value, ...(mine && !all ? [mine] : [])) | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 210 | .all<Row>(); |
| 211 | return ok(rows.results.map((row) => toAgent(row, now))); | |
| 212 | } | |
| 213 | ||
| 214 | async get(a: { workspace: string; handle: string; viewer: User | null }): Promise<Result<WorkspaceAgent>> { | |
| 215 | const seen = await this.seen(a.workspace, a.viewer); | |
| 216 | if (!seen.ok) return seen; | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 217 | await this.ensureBuiltin(seen.value); |
| Chat and workspace agents: channels, DMs and named agents you talk to | 218 | const row = await this.row(seen.value, a.handle); |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 219 | // Another member's personal agent is theirs: it isn't there for anyone else but owners. |
| 220 | return row && canSeeAgent(a.viewer, a.workspace, row) ? ok(toAgent(row, new Date())) : fail("not_found", `There is no agent called @${a.handle}.`); | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 221 | } |
| 222 | ||
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 223 | /** The agent by handle, if the viewer may see it. */ |
| 224 | private async visible(workspace: string, viewer: User | null, handle: unknown): Promise<Result<{ workspaceId: string; row: Row }>> { | |
| 225 | if (awaitsConfirmation(viewer)) return UNVERIFIED; | |
| 226 | const seen = await this.seen(workspace, viewer); | |
| 227 | if (!seen.ok) return seen; | |
| 228 | const row = await this.row(seen.value, handle); | |
| 229 | if (!row || !canSeeAgent(viewer, workspace, row)) return fail("not_found", `There is no agent called @${String(handle ?? "")}.`); | |
| 230 | return ok({ workspaceId: seen.value, row }); | |
| 231 | } | |
| 232 | ||
| Chat and workspace agents: channels, DMs and named agents you talk to | 233 | /** Internal: agents by id, archived ones too, so old messages still show who wrote them. */ |
| 234 | async byIds(a: { ids: string[] }): Promise<WorkspaceAgent[]> { | |
| 235 | const ids = [...new Set((Array.isArray(a.ids) ? a.ids : []).filter((id) => typeof id === "string"))].slice(0, 100); | |
| 236 | if (!ids.length) return []; | |
| 237 | const now = new Date(); | |
| 238 | const rows = await this.db | |
| 239 | .prepare(selectAgents(`a.id IN (${ids.map((_, i) => `?${i + 3}`).join(", ")})`)) | |
| 240 | .bind(...periods(now), ...ids) | |
| 241 | .all<Row>(); | |
| 242 | return rows.results.map((row) => toAgent(row, now)); | |
| 243 | } | |
| 244 | ||
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 245 | /** |
| 246 | * Who may create an agent, and of which scope: the workspace's id and | |
| 247 | * policy, and the scope it gets (access.ts `creatableScope`). | |
| 248 | */ | |
| 249 | private async creatable(workspace: string, viewer: User | null, asked: unknown): Promise<Result<{ workspaceId: string; scope: "workspace" | "personal"; policy: Awaited<ReturnType<typeof readPolicy>> }>> { | |
| 250 | if (awaitsConfirmation(viewer)) return UNVERIFIED; | |
| 251 | const seen = await this.seen(workspace, viewer); | |
| 252 | if (!seen.ok) return seen; | |
| 253 | const policy = await readPolicy(this.db, seen.value, monthKey(new Date())); | |
| 254 | const scope = creatableScope(viewer, workspace, asked, policy.members_create_agents); | |
| 255 | if (!scope.ok) return fail("forbidden", scope.message); | |
| 256 | return ok({ workspaceId: seen.value, scope: scope.scope, policy }); | |
| 257 | } | |
| 258 | ||
| An agent is on teams the way a person is: through team membership, with no team or department of its own. Add it to teams when you create it, from its profile or from the team's page; the Agents sidebar groups agents by the teams they're on, with those on none last; what an agent is told about its teammates, team budgets, skills reach, Spend's By team and the People pages all read memberships. Agents with an old team name matching a team are moved onto it once; the rest are left off a team. Templates name a role only. The agents, teams and people guides say how. | 259 | async create(a: { workspace: string; viewer: User | null; input: NewWorkspaceAgent; teams?: unknown }): Promise<Result<WorkspaceAgent>> { |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 260 | const allowed = await this.creatable(a.workspace, a.viewer, a.input?.scope); |
| 261 | if (!allowed.ok) return allowed; | |
| 262 | const { workspaceId, scope, policy } = allowed.value; | |
| 263 | const personal = scope === "personal"; | |
| 264 | const viewer = a.viewer!; | |
| 265 | if (personal) { | |
| 266 | const count = await this.db | |
| 267 | .prepare("SELECT COUNT(*) AS n FROM agents WHERE workspace_id = ? AND owner_id = ? AND scope = 'personal' AND archived_at IS NULL") | |
| 268 | .bind(workspaceId, viewer.id) | |
| 269 | .first<{ n: number }>(); | |
| 270 | if ((count?.n ?? 0) >= MAX_PERSONAL_AGENTS) { | |
| 271 | return fail("invalid", `You have ${MAX_PERSONAL_AGENTS} personal agents here, the most one person keeps. Archive one to make another.`); | |
| 272 | } | |
| 273 | } | |
| 274 | // A new agent starts with a budget, unless one was given: a member's own | |
| 275 | // $20 a month and $2 a session, or the workspace's default for its agents. | |
| 276 | const given = a.input?.budget ?? {}; | |
| 277 | const start = startingBudget(scope, policy.default_agent_monthly_micros); | |
| 278 | const budget = { | |
| 279 | monthly_micros: given.monthly_micros !== undefined ? given.monthly_micros : start.monthly_micros, | |
| 280 | daily_micros: given.daily_micros !== undefined ? given.daily_micros : start.daily_micros, | |
| 281 | task_micros: given.task_micros !== undefined ? given.task_micros : start.task_micros, | |
| 282 | }; | |
| 283 | const { scope: _scope, ...rest } = a.input ?? ({} as NewWorkspaceAgent); | |
| An agent is on teams the way a person is: through team membership, with no team or department of its own. Add it to teams when you create it, from its profile or from the team's page; the Agents sidebar groups agents by the teams they're on, with those on none last; what an agent is told about its teammates, team budgets, skills reach, Spend's By team and the People pages all read memberships. Agents with an old team name matching a team are moved onto it once; the rest are left off a team. Templates name a role only. The agents, teams and people guides say how. | 284 | const input = { ...rest, budget }; |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 285 | const checked = applyChanges(null, input, TEMPLATE_IDS); |
| Chat and workspace agents: channels, DMs and named agents you talk to | 286 | if (!checked.ok) return fail("invalid", checked.message); |
| 287 | const definition = checked.value; | |
| An agent is on teams the way a person is: through team membership, with no team or department of its own. Add it to teams when you create it, from its profile or from the team's page; the Agents sidebar groups agents by the teams they're on, with those on none last; what an agent is told about its teammates, team budgets, skills reach, Spend's By team and the People pages all read memberships. Agents with an old team name matching a team are moved onto it once; the rest are left off a team. Templates name a role only. The agents, teams and people guides say how. | 288 | // A personal agent is on no team: teams are shared, and it answers only its member. |
| 289 | const teams = personal ? ok([] as string[]) : await this.joinable(a.workspace, viewer, a.teams); | |
| 290 | if (!teams.ok) return teams; | |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 291 | if (await this.handleTaken(workspaceId, definition.handle, null)) { |
| Chat and workspace agents: channels, DMs and named agents you talk to | 292 | return fail("conflict", `${a.workspace} already has an agent called @${definition.handle}.`); |
| 293 | } | |
| 294 | const now = new Date().toISOString(); | |
| 295 | const id = newId("agt"); | |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 296 | const by = viewer.username; |
| 297 | const owned: Record<string, string> = personal ? { scope: "personal", owner_id: viewer.id, owner_username: viewer.username.toLowerCase() } : {}; | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 298 | try { |
| 299 | await this.db.batch([ | |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 300 | insertAgent(this.db, id, workspaceId, definition, { version: 1, created_by: by, created_at: now, updated_at: now, ...owned }), |
| Chat and workspace agents: channels, DMs and named agents you talk to | 301 | this.versionStatement(id, 1, definition, by, now), |
| 302 | ]); | |
| 303 | } catch (error) { | |
| 304 | if (String(error).includes("UNIQUE")) return fail("conflict", `${a.workspace} already has an agent called @${definition.handle}.`); | |
| 305 | throw error; | |
| 306 | } | |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 307 | this.audit(viewer, a.workspace, "create_agent", definition.handle, `Created ${personal ? "the personal agent " : ""}@${definition.handle} (version 1)`, "agents", personal ? "member" : "owner"); |
| A new agent says nothing until you do. Creating one still opens your direct message with it, but no introduction is written there, nothing is charged to its budget and nothing waits in Chat; the desk's special first message, its fixed fallback and the prompt that asked for it are gone. The agents guide says so. | 308 | // On its teams as it is made, so it knows them from its first message. |
| An agent is on teams the way a person is: through team membership, with no team or department of its own. Add it to teams when you create it, from its profile or from the team's page; the Agents sidebar groups agents by the teams they're on, with those on none last; what an agent is told about its teammates, team budgets, skills reach, Spend's By team and the People pages all read memberships. Agents with an old team name matching a team are moved onto it once; the rest are left off a team. Templates name a role only. The agents, teams and people guides say how. | 309 | if (teams.value.length) { |
| 310 | const identity = identityClient(this.env.IDENTITY); | |
| 311 | await Promise.all( | |
| 312 | teams.value.map((slug) => | |
| 313 | identity | |
| 314 | .setTeamAgent(viewer, a.workspace.toLowerCase(), slug, id) | |
| 315 | .then((added) => (added.ok ? null : console.error("agents: a new agent was not added to its team", id, slug, added.error.message))) | |
| 316 | .catch((error: unknown) => console.error("agents: a new agent was not added to its team", id, slug, String(error))), | |
| 317 | ), | |
| 318 | ); | |
| 319 | } | |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 320 | const row = await this.row(workspaceId, definition.handle); |
| Chat and workspace agents: channels, DMs and named agents you talk to | 321 | return ok(toAgent(row!, new Date())); |
| 322 | } | |
| 323 | ||
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 324 | // ── The builder: describe it, try it, change it in words (./builder.ts) ── |
| 325 | ||
| 326 | /** Drafts a whole agent from a description, charged to the viewer. Nothing is saved. */ | |
| 327 | async draft(a: { workspace: string; viewer: User | null; description: unknown; scope?: unknown }): Promise<Result<AgentProposal>> { | |
| 328 | const allowed = await this.creatable(a?.workspace ?? "", a?.viewer ?? null, a?.scope); | |
| 329 | if (!allowed.ok) return allowed; | |
| 330 | const words = wordsOf(a.description, "description"); | |
| 331 | if (!words.ok) return fail("invalid", words.message); | |
| 332 | const { workspaceId, scope, policy } = allowed.value; | |
| 333 | const slug = a.workspace.toLowerCase(); | |
| 334 | const [answer, handles] = await Promise.all([ | |
| 335 | callBuilder(this.env, { | |
| 336 | slug, | |
| 337 | workspaceId, | |
| 338 | viewer: a.viewer!, | |
| 339 | system: draftSystem(slug, scope), | |
| 340 | messages: [{ role: "user", content: `What the agent should do:\n\n${words.value}` }], | |
| 341 | maxOutput: DRAFT_OUTPUT_TOKENS, | |
| 342 | }), | |
| 343 | this.db.prepare("SELECT handle FROM agents WHERE workspace_id = ? AND archived_at IS NULL").bind(workspaceId).all<{ handle: string }>(), | |
| 344 | ]); | |
| 345 | if (!answer.ok) return fail(answer.code, answer.message); | |
| 346 | const proposal = proposalFrom(jsonIn(answer.text), { | |
| 347 | scope, | |
| 348 | taken: new Set(handles.results.map((row) => row.handle)), | |
| 349 | budget: startingBudget(scope, policy.default_agent_monthly_micros), | |
| 350 | }); | |
| 351 | if (!proposal.ok) return fail("invalid", proposal.message); | |
| 352 | return ok({ ...proposal.value, charged_micros: answer.charged }); | |
| 353 | } | |
| 354 | ||
| 355 | /** Try it: the unsaved definition answers the conversation so far. Charged to the viewer; nothing is saved. */ | |
| 356 | async tryDraft(a: { workspace: string; viewer: User | null; definition: unknown; messages: unknown }): Promise<Result<DraftReply>> { | |
| 357 | const given = (a?.definition ?? {}) as NewWorkspaceAgent; | |
| 358 | const allowed = await this.creatable(a?.workspace ?? "", a?.viewer ?? null, given?.scope); | |
| 359 | if (!allowed.ok) return allowed; | |
| 360 | const { scope: _scope, ...rest } = given; | |
| 361 | // A handle someone has already taken is no reason not to try it. | |
| 362 | const checked = applyChanges(null, { ...rest, handle: rest.handle || "preview" }, TEMPLATE_IDS); | |
| 363 | if (!checked.ok) return fail("invalid", checked.message); | |
| 364 | const turns = tryTurns(a.messages); | |
| 365 | if (!turns.ok) return fail("invalid", turns.message); | |
| 366 | const viewer = a.viewer!; | |
| 367 | const answer = await callBuilder(this.env, { | |
| 368 | slug: a.workspace.toLowerCase(), | |
| 369 | workspaceId: allowed.value.workspaceId, | |
| 370 | viewer, | |
| 371 | system: trySystem({ workspace: a.workspace.toLowerCase(), definition: checked.value, asker: { username: viewer.username, display_name: viewer.display_username ?? null } }), | |
| 372 | messages: turns.value, | |
| 373 | routing: { floor: checked.value.routing.floor, ceiling: checked.value.routing.ceiling }, | |
| 374 | }); | |
| 375 | if (!answer.ok) return fail(answer.code, answer.message); | |
| 376 | return ok({ text: answer.text, charged_micros: answer.charged }); | |
| 377 | } | |
| 378 | ||
| 379 | /** Drafts changes from a request in words, for whoever may change the agent. Saved only through `update`. */ | |
| 380 | async redraft(a: { workspace: string; handle: string; viewer: User | null; request: unknown }): Promise<Result<AgentRedraft>> { | |
| 381 | const found = await this.visible(a?.workspace ?? "", a?.viewer ?? null, a?.handle); | |
| 382 | if (!found.ok) return found; | |
| 383 | const { workspaceId, row } = found.value; | |
| 384 | if (!canChange(a.viewer, a.workspace, row)) return fail("forbidden", isPersonal(row) ? NOT_YOURS_REFUSAL : MANAGE_REFUSAL); | |
| 385 | const words = wordsOf(a.request, "request"); | |
| 386 | if (!words.ok) return fail("invalid", words.message); | |
| 387 | const before = definitionOf(row); | |
| 388 | const slug = a.workspace.toLowerCase(); | |
| 389 | const answer = await callBuilder(this.env, { | |
| 390 | slug, | |
| 391 | workspaceId, | |
| 392 | viewer: a.viewer!, | |
| 393 | system: redraftSystem(slug, before, !!row.builtin), | |
| 394 | messages: [{ role: "user", content: `The change to make to @${row.handle}:\n\n${words.value}` }], | |
| 395 | maxOutput: DRAFT_OUTPUT_TOKENS, | |
| 396 | }); | |
| 397 | if (!answer.ok) return fail(answer.code, answer.message); | |
| 398 | const drafted = redraftFrom(jsonIn(answer.text), before, !!row.builtin); | |
| 399 | if (!drafted.ok) return fail("invalid", drafted.message); | |
| 400 | return ok({ ...drafted.value, from_version: row.version, charged_micros: answer.charged }); | |
| 401 | } | |
| 402 | ||
| 403 | /** | |
| 404 | * Owners make a personal agent a workspace agent: a new workspace agent | |
| 405 | * with the same handle, definition and every version, and one more | |
| 406 | * version saying so; the personal one is archived in the same step, its | |
| 407 | * memory and direct messages kept with it (owners can't read a member's | |
| 408 | * memory to choose from it, so none is carried). | |
| 409 | */ | |
| Each agent has an Abilities tab: g1t's built-ins, always on within the asker's access; its computer, coming; each connected integration's actions one row each, read, import, comment or resolve, with a level for each, alone, alone when the person asked for it, ask first or never, and whose connection it runs on; and MCP servers an owner adds. Reading is alone, writing inside g1t follows today's choices, anything that leaves g1t asks first, and production deploys can't go above ask. The agents service enforces every level: ask first posts a card to allow or deny and parks a session until it's answered, a refusal names its rule in the transcript and the audit log, and a missing ability posts a request to owners. The agent abilities guide says how. | 410 | // ── MCP servers (docs.g1t.sh/guides/agent-abilities/, "MCP servers") ───── |
| 411 | ||
| 412 | /** The agent, if the viewer may add servers to it: owners only, a workspace agent or a personal one alike. */ | |
| 413 | private async forMcp(a: { workspace: string; handle: string; viewer: User | null }): Promise<Result<{ row: Row; workspaceId: string }>> { | |
| 414 | const found = await this.visible(a?.workspace ?? "", a?.viewer ?? null, a?.handle); | |
| 415 | if (!found.ok) return found; | |
| 416 | if (!canManage(a.viewer, a.workspace)) return fail("forbidden", "Only the workspace's owners add MCP servers to an agent."); | |
| 417 | return ok({ row: found.value.row, workspaceId: found.value.workspaceId }); | |
| 418 | } | |
| 419 | ||
| 420 | /** Saves `definition` as the agent's next version, from the version read; what the agent is now. */ | |
| 421 | private async saveVersion(row: Row, definition: Definition, by: User, workspace: string, what: string): Promise<Result<WorkspaceAgent>> { | |
| 422 | const now = new Date().toISOString(); | |
| 423 | const version = row.version + 1; | |
| 424 | const [updated] = await this.db.batch([ | |
| 425 | updateAgent(this.db, row.id, row.version, definition, { version, updated_at: now }), | |
| 426 | this.db | |
| 427 | .prepare( | |
| 428 | `INSERT INTO agent_versions (agent_id, version, definition, changed_by, created_at) | |
| 429 | SELECT ?1, ?2, ?3, ?4, ?5 WHERE EXISTS (SELECT 1 FROM agents WHERE id = ?1 AND version = ?2 AND updated_at = ?5)`, | |
| 430 | ) | |
| 431 | .bind(row.id, version, JSON.stringify(definition), by.username, now), | |
| 432 | ]); | |
| 433 | if (!updated.meta.changes) return fail("conflict", `@${row.handle} was changed meanwhile. Reload it and try again.`); | |
| 434 | this.audit(by, workspace, "update_agent", row.handle, `${what} (version ${version})`, "agents", "owner"); | |
| 435 | const saved = await this.row(row.workspace_id, row.handle); | |
| 436 | return ok(toAgent(saved!, new Date())); | |
| 437 | } | |
| 438 | ||
| 439 | /** | |
| 440 | * Adds an MCP server to an agent: the name and address are checked, its | |
| 441 | * tools are listed from it now, and the agent gets a new version with | |
| 442 | * the server and its tools as abilities (each a write unless the server | |
| 443 | * says it only reads). Owners only. | |
| 444 | */ | |
| 445 | async addMcpServer(a: { workspace: string; handle: string; viewer: User | null; name: unknown; url: unknown }): Promise<Result<McpServer>> { | |
| 446 | const found = await this.forMcp(a); | |
| 447 | if (!found.ok) return found; | |
| 448 | const { row } = found.value; | |
| 449 | const name = checkMcpName(a.name); | |
| 450 | if (!name.ok) return fail("invalid", name.message); | |
| 451 | const url = checkMcpUrl(a.url); | |
| 452 | if (!url.ok) return fail("invalid", url.message); | |
| 453 | const before = definitionOf(row); | |
| 454 | const servers = before.abilities.mcp_servers ?? []; | |
| 455 | if (servers.length >= MAX_MCP_SERVERS) return fail("limit", `An agent has at most ${MAX_MCP_SERVERS} MCP servers.`); | |
| 456 | if (servers.some((server) => server.name === name.value)) return fail("conflict", `@${row.handle} already has a server called ${name.value}.`); | |
| 457 | if (servers.some((server) => server.url === url.value)) return fail("conflict", `@${row.handle} already has that server.`); | |
| 458 | const listed = await listMcpTools(url.value); | |
| 459 | const now = new Date().toISOString(); | |
| 460 | const server: McpServer = { | |
| 461 | id: newId("mcp"), | |
| 462 | name: name.value, | |
| 463 | url: url.value, | |
| 464 | tools: listed.ok ? listed.tools : [], | |
| 465 | added_by: a.viewer!.username, | |
| 466 | added_at: now, | |
| 467 | checked_at: now, | |
| 468 | problem: listed.ok ? null : listed.message, | |
| 469 | }; | |
| 470 | const checked = applyChanges(before, {}, TEMPLATE_IDS, { builtin: !!row.builtin, mcp_servers: [...servers, server] }); | |
| 471 | if (!checked.ok) return fail("invalid", checked.message); | |
| 472 | const saved = await this.saveVersion(row, checked.value, a.viewer!, a.workspace, `Added the MCP server ${server.name} (${server.tools.length} tools) to @${row.handle}`); | |
| 473 | return saved.ok ? ok(server) : saved; | |
| 474 | } | |
| 475 | ||
| 476 | /** Takes an MCP server off an agent, with its abilities' settings: a new version. Owners only. */ | |
| 477 | async removeMcpServer(a: { workspace: string; handle: string; viewer: User | null; id: unknown }): Promise<Result<null>> { | |
| 478 | const found = await this.forMcp(a); | |
| 479 | if (!found.ok) return found; | |
| 480 | const { row } = found.value; | |
| 481 | const before = definitionOf(row); | |
| 482 | const servers = before.abilities.mcp_servers ?? []; | |
| 483 | const server = servers.find((s) => s.id === a.id); | |
| 484 | if (!server) return fail("not_found", "There is no such server on this agent."); | |
| 485 | const checked = applyChanges(before, {}, TEMPLATE_IDS, { builtin: !!row.builtin, mcp_servers: servers.filter((s) => s.id !== server.id) }); | |
| 486 | if (!checked.ok) return fail("invalid", checked.message); | |
| 487 | const saved = await this.saveVersion(row, checked.value, a.viewer!, a.workspace, `Removed the MCP server ${server.name} from @${row.handle}`); | |
| 488 | return saved.ok ? ok(null) : saved; | |
| 489 | } | |
| 490 | ||
| 491 | /** Lists a server's tools again; a new version when they changed. Owners only. */ | |
| 492 | async refreshMcpServer(a: { workspace: string; handle: string; viewer: User | null; id: unknown }): Promise<Result<McpServer>> { | |
| 493 | const found = await this.forMcp(a); | |
| 494 | if (!found.ok) return found; | |
| 495 | const { row } = found.value; | |
| 496 | const before = definitionOf(row); | |
| 497 | const servers = before.abilities.mcp_servers ?? []; | |
| 498 | const server = servers.find((s) => s.id === a.id); | |
| 499 | if (!server) return fail("not_found", "There is no such server on this agent."); | |
| 500 | const listed = await listMcpTools(server.url); | |
| 501 | const now = new Date().toISOString(); | |
| 502 | const fresh: McpServer = { ...server, tools: listed.ok ? listed.tools : server.tools, checked_at: now, problem: listed.ok ? null : listed.message }; | |
| 503 | if (JSON.stringify(fresh.tools) === JSON.stringify(server.tools) && fresh.problem === server.problem) return ok(fresh); | |
| 504 | const checked = applyChanges(before, {}, TEMPLATE_IDS, { builtin: !!row.builtin, mcp_servers: servers.map((s) => (s.id === server.id ? fresh : s)) }); | |
| 505 | if (!checked.ok) return fail("invalid", checked.message); | |
| 506 | const saved = await this.saveVersion(row, checked.value, a.viewer!, a.workspace, `Listed the MCP server ${server.name}'s tools again for @${row.handle} (${fresh.tools.length} tools)`); | |
| 507 | return saved.ok ? ok(fresh) : saved; | |
| 508 | } | |
| 509 | ||
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 510 | async promote(a: { workspace: string; handle: string; viewer: User | null }): Promise<Result<WorkspaceAgent>> { |
| 511 | const managed = await this.managed(a?.workspace ?? "", a?.viewer ?? null); | |
| 512 | if (!managed.ok) return managed.error.code === "forbidden" ? fail("forbidden", "Only the workspace's owners promote personal agents.") : managed; | |
| 513 | const row = await this.row(managed.value, a.handle); | |
| 514 | if (!row) return fail("not_found", `There is no agent called @${a.handle}.`); | |
| 515 | if (!isPersonal(row)) return fail("invalid", `@${row.handle} is already a workspace agent.`); | |
| 516 | const versions = await this.db | |
| 517 | .prepare("SELECT version, definition, changed_by, created_at FROM agent_versions WHERE agent_id = ? ORDER BY version") | |
| 518 | .bind(row.id) | |
| 519 | .all<{ version: number; definition: string; changed_by: string; created_at: string }>(); | |
| 520 | const definition = definitionOf(row); | |
| 521 | const now = new Date().toISOString(); | |
| 522 | const id = newId("agt"); | |
| 523 | const version = row.version + 1; | |
| 524 | const by = a.viewer!.username; | |
| 525 | try { | |
| 526 | const [archived] = await this.db.batch([ | |
| 527 | // Archived first, from the version read, so the handle is free for the new one in the same step. | |
| An archived agent can be found, brought back, or deleted for good. The Agents page ends with Archived, folded, where owners see every archived agent and members their own personal ones, each with when and by whom it was archived and what deleting it would take; Restore brings it back, under a new handle when a live agent has taken its old one, and Delete permanently asks for the handle to be typed and then removes its sessions, replies, memories, routines, versions, drafts, skills and its computer's home, leaving its messages in chat under its saved name, the workspace's spend totals and the audit log. In Chat an archived agent is deactivated the way people expect: never offered as someone to message or mention, never woken, a conversation with it kept in the list, dimmed and marked Archived, open to read with a note in place of the composer and a way to restore it for owners; after deletion the conversation still reads, its messages marked as a deleted agent's. The agents, chat and agent budgets guides say so. | 528 | this.db.prepare("UPDATE agents SET archived_at = ?, archived_by = ?, updated_at = ? WHERE id = ? AND version = ? AND archived_at IS NULL").bind(now, by, now, row.id, row.version), |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 529 | insertAgent(this.db, id, managed.value, definition, { version, created_by: row.created_by, created_at: now, updated_at: now }), |
| 530 | ...versions.results.map((v) => | |
| 531 | this.db.prepare("INSERT INTO agent_versions (agent_id, version, definition, changed_by, created_at) VALUES (?, ?, ?, ?, ?)").bind(id, v.version, v.definition, v.changed_by, v.created_at), | |
| 532 | ), | |
| 533 | this.versionStatement(id, version, definition, by, now), | |
| 534 | ]); | |
| 535 | if (!archived.meta.changes) throw new Error("changed meanwhile"); | |
| 536 | } catch (error) { | |
| 537 | console.error("agents: a promotion failed", row.id, String(error)); | |
| 538 | return fail("conflict", `@${row.handle} was changed meanwhile. Reload it and try again.`); | |
| 539 | } | |
| 540 | this.audit(a.viewer!, a.workspace, "promote_agent", row.handle, `Promoted @${row.owner_username ?? "a member"}'s personal agent @${row.handle} to a workspace agent (version ${version})`); | |
| 541 | const saved = await this.row(managed.value, row.handle); | |
| 542 | return ok(toAgent(saved!, new Date())); | |
| 543 | } | |
| 544 | ||
| The Marketplace is where a workspace finds what to add, reached from Apps: the agent catalog of roles with who already does each job and Add to workspace; every integration with whether it's connected; g1t's own extensions, listed as Soon; and requests, where members ask and owners add or turn down, each told by a notification. Underneath, an extension is a manifest (publisher and tier, source and version from a tag, scopes, the domains its data goes to, its page on g1tusercontent.com, no price) and an install pins its version with an off switch, a budget and the audit log. Apps lists only what was added from the Marketplace, with pins; g1t's own pages stay in their sidebars, and AI Gateway joins Workspace's. The marketplace guide says how. | 545 | // ── The Marketplace's install requests (./installs.ts) ───────────────── |
| 546 | ||
| 547 | /** Requests as the viewer sees them: every one for an owner, their own for anyone else. */ | |
| 548 | async installRequests(a: { workspace: string; viewer: User | null }): Promise<Result<InstallRequests>> { | |
| 549 | if (a?.viewer && awaitsConfirmation(a.viewer)) return UNVERIFIED; | |
| 550 | const seen = await this.seen(a?.workspace ?? "", a?.viewer ?? null); | |
| 551 | if (!seen.ok) return seen; | |
| 552 | const owner = canManage(a.viewer, a.workspace); | |
| 553 | const requests = await listRequests(this.db, seen.value, a.viewer!, owner); | |
| 554 | return ok({ requests, can_resolve: owner }); | |
| 555 | } | |
| 556 | ||
| 557 | /** A member asks the owners to add something; each owner is notified. */ | |
| 558 | async requestInstall(a: { workspace: string; viewer: User | null; listing: unknown; note?: unknown }): Promise<Result<InstallRequest>> { | |
| 559 | if (a?.viewer && awaitsConfirmation(a.viewer)) return UNVERIFIED; | |
| 560 | const seen = await this.seen(a?.workspace ?? "", a?.viewer ?? null); | |
| 561 | if (!seen.ok) return seen; | |
| 562 | const viewer = a.viewer!; | |
| 563 | if ((viewer.kind ?? "user") !== "user") return fail("forbidden", "Only people ask the workspace's owners to add things."); | |
| 564 | if (canManage(viewer, a.workspace)) return fail("invalid", "You're an owner of this workspace: add it yourself."); | |
| 565 | const listing = findListing(a.listing); | |
| 566 | if (!listing) return fail("not_found", "That isn't something a workspace can add yet."); | |
| 567 | const opened = await openRequest(this.db, seen.value, newId("ins"), listing, viewer, cleanRequestNote(a.note)); | |
| 568 | if (!opened.ok) return opened; | |
| 569 | const slug = a.workspace.toLowerCase(); | |
| 570 | this.audit(viewer, slug, "request_install", listing.ref, `Asked the owners to add ${listing.name}`, "marketplace"); | |
| 571 | this.defer(this.tellOwners(slug, viewer, opened.value)); | |
| 572 | return opened; | |
| 573 | } | |
| 574 | ||
| 575 | /** An owner adds or turns down a request; whoever asked is told. */ | |
| 576 | async resolveInstallRequest(a: { workspace: string; viewer: User | null; id: unknown; status: unknown }): Promise<Result<InstallRequest>> { | |
| 577 | const managed = await this.managed(a?.workspace ?? "", a?.viewer ?? null); | |
| 578 | if (!managed.ok) { | |
| 579 | return managed.error.code === "forbidden" ? fail("forbidden", "Only the workspace's owners answer requests.") : managed; | |
| 580 | } | |
| 581 | if (a.status !== "done" && a.status !== "declined") return fail("invalid", "Answer a request with done or declined."); | |
| 582 | const answered = await resolveRequest(this.db, managed.value, String(a.id ?? ""), a.status, a.viewer!); | |
| 583 | if (!answered.ok) return answered; | |
| 584 | const slug = a.workspace.toLowerCase(); | |
| 585 | const { request } = answered.value; | |
| 586 | this.audit(a.viewer!, slug, "answer_install_request", request.listing, `${request.status === "done" ? "Added" : "Turned down"} ${request.name} for @${request.requested_by}`, "marketplace"); | |
| 587 | this.defer(this.tellRequester(slug, a.viewer!, answered.value)); | |
| 588 | return ok(request); | |
| 589 | } | |
| 590 | ||
| 591 | /** Marks every open request for `listing` added, and tells each person who asked. */ | |
| Merge branch 'customize' into main: Customize replaces the Marketplace, the Artifacts trash and spaces changes, the Security-free cleanup, with the desktop app's own title bar kept in the shell. | 592 | async answerListing(slug: string, workspaceId: string, listing: string, by: User): Promise<void> { |
| The Marketplace is where a workspace finds what to add, reached from Apps: the agent catalog of roles with who already does each job and Add to workspace; every integration with whether it's connected; g1t's own extensions, listed as Soon; and requests, where members ask and owners add or turn down, each told by a notification. Underneath, an extension is a manifest (publisher and tier, source and version from a tag, scopes, the domains its data goes to, its page on g1tusercontent.com, no price) and an install pins its version with an off switch, a budget and the audit log. Apps lists only what was added from the Marketplace, with pins; g1t's own pages stay in their sidebars, and AI Gateway joins Workspace's. The marketplace guide says how. | 593 | try { |
| 594 | const answered = await resolveListing(this.db, workspaceId, listing, by as Person); | |
| 595 | await Promise.all(answered.map((one) => this.tellRequester(slug.toLowerCase(), by, one))); | |
| 596 | } catch (error) { | |
| 597 | console.error("agents: requests for a listing were not answered", listing, String(error)); | |
| 598 | } | |
| 599 | } | |
| 600 | ||
| 601 | /** Every owner hears of a new request (at most 20 of them). */ | |
| 602 | private async tellOwners(slug: string, asker: User, request: InstallRequest): Promise<void> { | |
| 603 | if (!this.env.NOTIFY) return; | |
| 604 | try { | |
| 605 | const members = await identityClient(this.env.IDENTITY).listMembers(slug, asker); | |
| 606 | if (!members.ok) throw new Error(members.error.message); | |
| 607 | const owners = members.value.filter((member) => member.role === "owner").slice(0, 20); | |
| 608 | const notify = notifyClient(this.env.NOTIFY); | |
| 609 | await Promise.all( | |
| 610 | owners.map((owner) => | |
| 611 | notify | |
| 612 | .notify( | |
| 613 | { username: owner.username }, | |
| 614 | { | |
| 615 | id: `install-request:${request.id}:${owner.username}`, | |
| 616 | kind: "approval", | |
| 617 | workspace: slug, | |
| 618 | title: `@${asker.username} asks you to add ${request.name}`, | |
| The Marketplace adds what g1t doesn't do and connects what you already use: extensions (Mail, Support, CRM, Recruiting, On-call, and bridges to a helpdesk, CRM or ERP you run, all Soon), starter kits, and integrations connected by the workspace or by each person; agents aren't installed there. Agent templates live in Agents, as starting points you configure, at Agents → Templates, and the Marketplace's old agent addresses lead there. The marketplace and agents guides say how. | 619 | body: request.note ?? (request.kind === "extension" ? "An extension. Install it, or turn the request down." : "An integration. Connect it, or turn the request down."), |
| The Marketplace is where a workspace finds what to add, reached from Apps: the agent catalog of roles with who already does each job and Add to workspace; every integration with whether it's connected; g1t's own extensions, listed as Soon; and requests, where members ask and owners add or turn down, each told by a notification. Underneath, an extension is a manifest (publisher and tier, source and version from a tag, scopes, the domains its data goes to, its page on g1tusercontent.com, no price) and an install pins its version with an off switch, a budget and the audit log. Apps lists only what was added from the Marketplace, with pins; g1t's own pages stay in their sidebars, and AI Gateway joins Workspace's. The marketplace guide says how. | 620 | href: requestsPath(slug), |
| 621 | actor: { kind: "user", id: asker.id, name: asker.username, avatar: asker.avatar ?? null, avatar_seed: null }, | |
| 622 | created_at: new Date().toISOString(), | |
| 623 | }, | |
| 624 | ) | |
| 625 | .catch(() => undefined), | |
| 626 | ), | |
| 627 | ); | |
| 628 | } catch (error) { | |
| 629 | console.error("agents: owners were not told of an install request", request.id, String(error)); | |
| 630 | } | |
| 631 | } | |
| 632 | ||
| 633 | /** The person who asked hears their request was answered. */ | |
| 634 | private async tellRequester(slug: string, by: User, answered: Answered): Promise<void> { | |
| 635 | if (!this.env.NOTIFY) return; | |
| 636 | const { request } = answered; | |
| 637 | const line = answerLine(request, `@${by.username}`); | |
| 638 | await notifyClient(this.env.NOTIFY) | |
| 639 | .notify( | |
| 640 | { user_id: answered.requested_by_id }, | |
| 641 | { | |
| 642 | id: `install-answer:${request.id}`, | |
| 643 | kind: "inbox", | |
| 644 | workspace: slug, | |
| 645 | title: line.title, | |
| 646 | body: line.body, | |
| 647 | href: request.status === "done" ? listingPath(slug, request.listing) : requestsPath(slug), | |
| 648 | actor: { kind: "user", id: by.id, name: by.username, avatar: by.avatar ?? null, avatar_seed: null }, | |
| 649 | created_at: new Date().toISOString(), | |
| 650 | }, | |
| 651 | ) | |
| 652 | .catch((error: unknown) => console.error("agents: a requester was not told", request.id, String(error))); | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 653 | } |
| 654 | ||
| Chat and workspace agents: channels, DMs and named agents you talk to | 655 | async update(a: { workspace: string; handle: string; viewer: User | null; changes: Partial<NewWorkspaceAgent> }): Promise<Result<WorkspaceAgent>> { |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 656 | const found = await this.visible(a.workspace, a.viewer, a.handle); |
| 657 | if (!found.ok) return found; | |
| 658 | const { row } = found.value; | |
| 659 | // Owners change the workspace's agents; a member changes their own personal one. | |
| 660 | if (!canChange(a.viewer, a.workspace, row)) return fail("forbidden", isPersonal(row) ? NOT_YOURS_REFUSAL : MANAGE_REFUSAL); | |
| 661 | const workspaceId = found.value.workspaceId; | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 662 | const before = definitionOf(row); |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 663 | const { scope: _scope, ...asked } = (a.changes ?? {}) as Partial<NewWorkspaceAgent>; |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 664 | // The built-in @g1t keeps who it is and its job; the rest is the workspace's. |
| An agent is on teams the way a person is: through team membership, with no team or department of its own. Add it to teams when you create it, from its profile or from the team's page; the Agents sidebar groups agents by the teams they're on, with those on none last; what an agent is told about its teammates, team budgets, skills reach, Spend's By team and the People pages all read memberships. Agents with an old team name matching a team are moved onto it once; the rest are left off a team. Templates name a role only. The agents, teams and people guides say how. | 665 | const allowed = row.builtin ? builtinChanges(before, asked) : { ok: true as const, value: asked }; |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 666 | if (!allowed.ok) return fail("invalid", allowed.message); |
| 667 | const checked = applyChanges(before, allowed.value, TEMPLATE_IDS, { builtin: !!row.builtin }); | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 668 | if (!checked.ok) return fail("invalid", checked.message); |
| 669 | const definition = checked.value; | |
| 670 | if (JSON.stringify(definition) === JSON.stringify(before)) return ok(toAgent(row, new Date())); | |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 671 | if (definition.handle !== before.handle && (await this.handleTaken(workspaceId, definition.handle, row.id))) { |
| Chat and workspace agents: channels, DMs and named agents you talk to | 672 | return fail("conflict", `${a.workspace} already has an agent called @${definition.handle}.`); |
| 673 | } | |
| 674 | const now = new Date().toISOString(); | |
| 675 | const version = row.version + 1; | |
| 676 | const by = a.viewer!.username; | |
| 677 | try { | |
| 678 | const [updated] = await this.db.batch([ | |
| 679 | // Only from the version read: two owners saving at once never lose | |
| 680 | // one's change silently; the second is told to look again. | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 681 | updateAgent(this.db, row.id, row.version, definition, { version, updated_at: now }), |
| Chat and workspace agents: channels, DMs and named agents you talk to | 682 | this.db |
| 683 | .prepare( | |
| 684 | `INSERT INTO agent_versions (agent_id, version, definition, changed_by, created_at) | |
| 685 | SELECT ?1, ?2, ?3, ?4, ?5 WHERE EXISTS (SELECT 1 FROM agents WHERE id = ?1 AND version = ?2 AND updated_at = ?5)`, | |
| 686 | ) | |
| 687 | .bind(row.id, version, JSON.stringify(definition), by, now), | |
| 688 | ]); | |
| 689 | if (!updated.meta.changes) return fail("conflict", `@${before.handle} was changed meanwhile. Reload it and try again.`); | |
| 690 | } catch (error) { | |
| 691 | if (String(error).includes("UNIQUE")) return fail("conflict", `${a.workspace} already has an agent called @${definition.handle}.`); | |
| 692 | throw error; | |
| 693 | } | |
| 694 | const renamed = definition.handle !== before.handle ? ` (was @${before.handle})` : ""; | |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 695 | this.audit(a.viewer!, a.workspace, "update_agent", definition.handle, `Changed @${definition.handle}${renamed} to version ${version}`, "agents", isPersonal(row) ? "member" : "owner"); |
| 696 | const saved = await this.row(workspaceId, definition.handle); | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 697 | return ok(toAgent(saved!, new Date())); |
| 698 | } | |
| 699 | ||
| Pricing says it plainly: models at the provider's price, the agent rate for what g1t runs around every model call (the gateway, secrets, routing, context and pass-through to your own provider, so your own keys too), everything else at cost plus 20%, your own runners free, no seats; Security and quality comes with the plan with no separate fee, and live activations end. Each agent has an effort setting, Auto to Max, with what a typical task has cost at each level, and Spend's Spend less, keep quality suggests a lower level only when the agent's own past work shows quality held, to apply or dismiss. The pricing, spend and agents guides say how. | 700 | /** What each effort level has cost one agent, from its own finished sessions. */ |
| 701 | async effortCosts(a: { workspace: string; handle: string; viewer: User | null }): Promise<Result<AgentEffortCosts>> { | |
| 702 | if (a?.viewer && awaitsConfirmation(a.viewer)) return UNVERIFIED; | |
| 703 | const seen = await this.seen(a?.workspace ?? "", a?.viewer ?? null); | |
| 704 | if (!seen.ok) return seen; | |
| 705 | const row = await this.row(seen.value, a.handle); | |
| 706 | if (!row) return fail("not_found", `There is no agent called @${a.handle}.`); | |
| 707 | const outcomes = await outcomesSince(this.db, seen.value, sinceWindow(), row.id); | |
| 708 | return ok(effortCostsOf(row.handle, effortOf(definitionOf(row).routing), outcomes.get(row.id) ?? [])); | |
| 709 | } | |
| 710 | ||
| 711 | /** Ways to spend less, checked against past work: every agent's, or one's. */ | |
| 712 | async recommendations(a: { workspace: string; viewer: User | null; handle?: string | null }): Promise<Result<AgentRecommendations>> { | |
| 713 | if (a?.viewer && awaitsConfirmation(a.viewer)) return UNVERIFIED; | |
| 714 | const seen = await this.seen(a?.workspace ?? "", a?.viewer ?? null); | |
| 715 | if (!seen.ok) return seen; | |
| 716 | let agentId: string | null = null; | |
| 717 | if (a.handle) { | |
| 718 | const row = await this.row(seen.value, a.handle); | |
| 719 | if (!row) return fail("not_found", `There is no agent called @${a.handle}.`); | |
| 720 | agentId = row.id; | |
| 721 | } | |
| 722 | return ok(await readRecommendations(this.db, seen.value, agentId)); | |
| 723 | } | |
| 724 | ||
| 725 | /** | |
| 726 | * Owners apply a suggestion (the agent's effort changes as a new version | |
| 727 | * of it, and the audit log says so) or dismiss it. Only an open one, and | |
| 728 | * only while the agent's setting is still the one it was made for. | |
| 729 | */ | |
| 730 | async resolveRecommendation(a: { workspace: string; viewer: User | null; id: string; action: string }): Promise<Result<AgentRecommendation>> { | |
| 731 | const managed = await this.managed(a?.workspace ?? "", a?.viewer ?? null); | |
| 732 | if (!managed.ok) return managed; | |
| 733 | if (a.action !== "apply" && a.action !== "dismiss") return fail("invalid", "Apply or dismiss."); | |
| 734 | const found = await recommendationRow(this.db, managed.value, String(a.id ?? "")); | |
| 735 | if (!found) return fail("not_found", "There is no such suggestion."); | |
| 736 | if (found.status !== "open") return fail("conflict", "This suggestion was already decided, or is no longer current."); | |
| 737 | const by = a.viewer!.username; | |
| 738 | if (a.action === "dismiss") { | |
| 739 | if (!(await markResolved(this.db, found.id, "dismissed", by))) return fail("conflict", "This suggestion was decided meanwhile."); | |
| 740 | this.audit(a.viewer!, a.workspace, "dismiss_recommendation", found.handle ?? found.agent_id, `Dismissed "${found.title}"`); | |
| 741 | } else { | |
| 742 | const agent = await this.row(managed.value, found.handle); | |
| 743 | if (!agent || agent.id !== found.agent_id) return fail("not_found", "That agent is gone."); | |
| 744 | if (effortOf(definitionOf(agent).routing) !== found.from_effort) { | |
| 745 | await this.db.prepare("UPDATE agent_recommendations SET status = 'stale' WHERE id = ? AND status = 'open'").bind(found.id).run(); | |
| 746 | return fail("conflict", `@${agent.handle}'s effort was changed since this was suggested. The next check looks again.`); | |
| 747 | } | |
| 748 | // Claimed first, so two owners pressing Apply change the agent once. | |
| 749 | if (!(await markResolved(this.db, found.id, "applied", by))) return fail("conflict", "This suggestion was decided meanwhile."); | |
| 750 | const changed = await this.update({ workspace: a.workspace, handle: agent.handle, viewer: a.viewer, changes: { routing: { effort: found.to_effort as never } } }); | |
| 751 | if (!changed.ok) { | |
| 752 | await this.db.prepare("UPDATE agent_recommendations SET status = 'open', resolved_by = NULL, resolved_at = NULL WHERE id = ?").bind(found.id).run(); | |
| 753 | return changed; | |
| 754 | } | |
| 755 | this.audit( | |
| 756 | a.viewer!, | |
| 757 | a.workspace, | |
| 758 | "apply_recommendation", | |
| 759 | agent.handle, | |
| 760 | `Applied "${found.title}": @${agent.handle}'s effort from ${EFFORT_NAMES[found.from_effort as keyof typeof EFFORT_NAMES] ?? found.from_effort} to ${EFFORT_NAMES[found.to_effort as keyof typeof EFFORT_NAMES] ?? found.to_effort}`, | |
| 761 | ); | |
| 762 | } | |
| 763 | const after = await recommendationRow(this.db, managed.value, found.id); | |
| 764 | return ok(toRecommendation(after!)); | |
| 765 | } | |
| 766 | ||
| Chat and workspace agents: channels, DMs and named agents you talk to | 767 | async archive(a: { workspace: string; handle: string; viewer: User | null }): Promise<Result<null>> { |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 768 | const found = await this.visible(a.workspace, a.viewer, a.handle); |
| 769 | if (!found.ok) return found; | |
| 770 | const { row } = found.value; | |
| 771 | // Owners archive any agent; a member their own personal one. | |
| 772 | if (!canArchive(a.viewer, a.workspace, row)) return fail("forbidden", MANAGE_REFUSAL); | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 773 | if (row.builtin) return fail("invalid", "@g1t is every workspace's orchestrator and can't be archived. You can set its budget to limit it."); |
| Chat and workspace agents: channels, DMs and named agents you talk to | 774 | const now = new Date().toISOString(); |
| An archived agent can be found, brought back, or deleted for good. The Agents page ends with Archived, folded, where owners see every archived agent and members their own personal ones, each with when and by whom it was archived and what deleting it would take; Restore brings it back, under a new handle when a live agent has taken its old one, and Delete permanently asks for the handle to be typed and then removes its sessions, replies, memories, routines, versions, drafts, skills and its computer's home, leaving its messages in chat under its saved name, the workspace's spend totals and the audit log. In Chat an archived agent is deactivated the way people expect: never offered as someone to message or mention, never woken, a conversation with it kept in the list, dimmed and marked Archived, open to read with a note in place of the composer and a way to restore it for owners; after deletion the conversation still reads, its messages marked as a deleted agent's. The agents, chat and agent budgets guides say so. | 775 | await this.db.prepare("UPDATE agents SET archived_at = ?, archived_by = ?, updated_at = ? WHERE id = ? AND archived_at IS NULL").bind(now, a.viewer!.username, now, row.id).run(); |
| Chat and workspace agents: channels, DMs and named agents you talk to | 776 | this.audit(a.viewer!, a.workspace, "archive_agent", row.handle, `Archived @${row.handle}`); |
| Every agent can have its own computer. A session that needs one wakes it: a home of its own on g1t cloud, one per agent and never shared, where it runs commands, reads and writes files and keeps what it made, with each session working in its own folder under a shared home; after ten idle minutes it sleeps, its home kept as a snapshot and restored when it wakes, and Reset wipes the home while memory and artifacts stay. Its shell and files are abilities with the usual choices, Alone, Alone when asked, Ask first or Never, offered only inside sessions and never to a chat reply; every command shows on the session with its output, and the agent's new Computer tab shows the state, the disk used of the five gigabytes included, the recent commands, and Wake, Put to sleep and Reset. Machine time counts only while it is awake, on the sandbox lines of the ledger that name the agent and who asked, held to the same spend caps as the session; the disk itself costs nothing in this version. The runner gained a long-lived supervisor that answers the computer's requests inside the container, and the runner service a computer per agent that keeps its snapshot in the agent homes bucket when one is attached, and says so when none is. The REST API and the agent tool can read a computer, wake it, put it to sleep and reset it. The agents, abilities, sessions, runners, billing and deploy guides say how it works and what an operator sets up; pinning a computer to your own runner, its browser and take-over come next. | 777 | // Its computer goes to sleep now; its disk is kept 30 days, then deleted (computer.ts). |
| 778 | this.defer(forgetComputer(this.env, row.id)); | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 779 | return ok(null); |
| 780 | } | |
| 781 | ||
| An archived agent can be found, brought back, or deleted for good. The Agents page ends with Archived, folded, where owners see every archived agent and members their own personal ones, each with when and by whom it was archived and what deleting it would take; Restore brings it back, under a new handle when a live agent has taken its old one, and Delete permanently asks for the handle to be typed and then removes its sessions, replies, memories, routines, versions, drafts, skills and its computer's home, leaving its messages in chat under its saved name, the workspace's spend totals and the audit log. In Chat an archived agent is deactivated the way people expect: never offered as someone to message or mention, never woken, a conversation with it kept in the list, dimmed and marked Archived, open to read with a note in place of the composer and a way to restore it for owners; after deletion the conversation still reads, its messages marked as a deleted agent's. The agents, chat and agent budgets guides say so. | 782 | // ── Archived agents (./archived.ts) ────────────────────────────────────── |
| 783 | ||
| 784 | /** The archived agents the viewer may see: every one for an owner, their own archived personal agents for anyone else. */ | |
| 785 | async listArchived(a: { workspace: string; viewer: User | null }): Promise<Result<ArchivedAgent[]>> { | |
| 786 | if (a?.viewer && awaitsConfirmation(a.viewer)) return UNVERIFIED; | |
| 787 | const seen = await this.seen(a?.workspace ?? "", a?.viewer ?? null); | |
| 788 | if (!seen.ok) return seen; | |
| 789 | const viewer = a.viewer!; | |
| 790 | return ok(await listArchived(this.db, seen.value, { id: (viewer.kind ?? "user") === "user" ? viewer.id : "", owner: canManage(viewer, a.workspace) })); | |
| 791 | } | |
| 792 | ||
| 793 | /** | |
| 794 | * An archived agent the viewer may see, by id, else the one most recently | |
| 795 | * archived under `handle` (a handle is unique only among live agents). | |
| 796 | */ | |
| 797 | private async archivedRow(workspace: string, viewer: User | null, a: { id?: unknown; handle?: unknown }): Promise<Result<{ workspaceId: string; row: Row }>> { | |
| 798 | if (awaitsConfirmation(viewer)) return UNVERIFIED; | |
| 799 | const seen = await this.seen(workspace, viewer); | |
| 800 | if (!seen.ok) return seen; | |
| 801 | const now = new Date(); | |
| 802 | const handle = typeof a.handle === "string" ? a.handle.trim().replace(/^@/, "").toLowerCase() : ""; | |
| 803 | const row = | |
| 804 | typeof a.id === "string" && a.id | |
| 805 | ? await this.db.prepare(selectAgents("a.workspace_id = ?3 AND a.id = ?4 AND a.archived_at IS NOT NULL")).bind(...periods(now), seen.value, a.id).first<Row>() | |
| 806 | : handle | |
| 807 | ? await this.db | |
| 808 | .prepare(`${selectAgents("a.workspace_id = ?3 AND a.handle = ?4 AND a.archived_at IS NOT NULL")} ORDER BY a.archived_at DESC LIMIT 1`) | |
| 809 | .bind(...periods(now), seen.value, handle) | |
| 810 | .first<Row>() | |
| 811 | : null; | |
| 812 | if (!row || !canSeeAgent(viewer, workspace, row)) return fail("not_found", `There is no archived agent called @${handle || String(a.id ?? "")}.`); | |
| 813 | return ok({ workspaceId: seen.value, row }); | |
| 814 | } | |
| 815 | ||
| 816 | /** | |
| 817 | * Brings an archived agent back, live again under its handle, or under | |
| 818 | * `new_handle` when a live agent took the old one (`conflict` says so). | |
| 819 | * Owners any agent, a member their own personal one. A new handle is a | |
| 820 | * new version, as any rename is. Its computer's disk is kept after all. | |
| 821 | */ | |
| 822 | async restore(a: { workspace: string; handle?: string; id?: string; viewer: User | null; new_handle?: unknown }): Promise<Result<WorkspaceAgent>> { | |
| 823 | const found = await this.archivedRow(a?.workspace ?? "", a?.viewer ?? null, a ?? {}); | |
| 824 | if (!found.ok) return found; | |
| 825 | const { workspaceId, row } = found.value; | |
| 826 | const refused = restoreRefusal(a.viewer, a.workspace, row); | |
| 827 | if (refused) return { ok: false, error: refused }; | |
| 828 | const live = await this.db | |
| 829 | .prepare("SELECT handle FROM agents WHERE workspace_id = ? AND archived_at IS NULL") | |
| 830 | .bind(workspaceId) | |
| 831 | .all<{ handle: string }>(); | |
| 832 | const taken = new Set(live.results.map((r) => r.handle)); | |
| 833 | const handle = restoreHandle(row, a.new_handle, (h) => taken.has(h)); | |
| 834 | if (!handle.ok) return handle; | |
| 835 | const now = new Date().toISOString(); | |
| 836 | const renamed = handle.handle !== row.handle; | |
| 837 | const version = renamed ? row.version + 1 : row.version; | |
| 838 | try { | |
| 839 | const [restored] = await this.db.batch([ | |
| 840 | this.db | |
| 841 | .prepare("UPDATE agents SET archived_at = NULL, archived_by = NULL, handle = ?, version = ?, updated_at = ? WHERE id = ? AND archived_at IS NOT NULL") | |
| 842 | .bind(handle.handle, version, now, row.id), | |
| 843 | ...(renamed ? [this.versionStatement(row.id, version, { ...definitionOf(row), handle: handle.handle }, a.viewer!.username, now)] : []), | |
| 844 | ]); | |
| 845 | if (!restored.meta.changes) return fail("not_found", `@${row.handle} was restored or deleted meanwhile.`); | |
| 846 | } catch (error) { | |
| 847 | if (String(error).includes("UNIQUE")) return fail("conflict", `A live agent is called @${handle.handle} now. Rename that one first, or restore this one under a new handle.`); | |
| 848 | throw error; | |
| 849 | } | |
| 850 | const was = renamed ? ` (was @${row.handle})` : ""; | |
| 851 | this.audit(a.viewer!, a.workspace, "restore_agent", handle.handle, `Restored @${handle.handle}${was}`, "agents", isPersonal(row) ? "member" : "owner"); | |
| 852 | this.defer(keepComputer(this.env, row.id)); | |
| 853 | const saved = await this.row(workspaceId, handle.handle); | |
| 854 | return ok(toAgent(saved!, new Date())); | |
| 855 | } | |
| 856 | ||
| 857 | /** | |
| 858 | * Deletes an archived agent for good (archived.ts `DESTROYED`): every row | |
| 859 | * the service keeps about it, in one batch; then its computer on the | |
| 860 | * runner, its chat memberships (its messages stay under the name it | |
| 861 | * had) and its team memberships. Owners any agent, a member their own | |
| 862 | * personal one; `confirm` is its handle typed back. A live agent is | |
| 863 | * archived first. A second call finds nothing. | |
| 864 | */ | |
| 865 | async destroy(a: { workspace: string; handle?: string; id?: string; viewer: User | null; confirm?: unknown }): Promise<Result<null>> { | |
| 866 | const found = await this.archivedRow(a?.workspace ?? "", a?.viewer ?? null, a ?? {}); | |
| 867 | if (!found.ok) { | |
| 868 | // A live agent of that handle: say what to do, rather than "no such archived agent". | |
| 869 | if (found.error.code === "not_found" && typeof a?.handle === "string") { | |
| 870 | const live = await this.visible(a.workspace, a.viewer, a.handle); | |
| 871 | if (live.ok) { | |
| 872 | const refused = destroyRefusal(a.viewer, a.workspace, live.value.row, a.confirm); | |
| 873 | return { ok: false, error: refused ?? found.error }; | |
| 874 | } | |
| 875 | } | |
| 876 | return found; | |
| 877 | } | |
| 878 | const { workspaceId, row } = found.value; | |
| 879 | const refused = destroyRefusal(a.viewer, a.workspace, row, a.confirm); | |
| 880 | if (refused) return { ok: false, error: refused }; | |
| 881 | const results = await this.db.batch(destroyStatements(this.db, row.id)); | |
| 882 | if (!results[results.length - 1]!.meta.changes) return fail("not_found", `@${row.handle} was restored or deleted meanwhile.`); | |
| 883 | this.audit(a.viewer!, a.workspace, "delete_agent", row.handle, `Deleted the archived agent @${row.handle} for good`, "agents", isPersonal(row) ? "member" : "owner"); | |
| 884 | const slug = a.workspace.toLowerCase(); | |
| 885 | const definition = definitionOf(row); | |
| 886 | this.defer(eraseComputer(this.env, row.id)); | |
| 887 | this.defer( | |
| 888 | chatClient(this.env.CHAT) | |
| 889 | .agentDestroyed({ workspace_id: workspaceId, agent_id: row.id, handle: row.handle, display_name: row.display_name, avatar_seed: definition.avatar_seed, look: definition.look }) | |
| 890 | .then((told) => (told.ok ? null : console.error("agents: chat was not told of a deleted agent", row.id, told.error.message))) | |
| 891 | .catch((error: unknown) => console.error("agents: chat was not told of a deleted agent", row.id, String(error))), | |
| 892 | ); | |
| 893 | this.defer(this.leaveTeams(slug, a.viewer!, row.id)); | |
| 894 | return ok(null); | |
| 895 | } | |
| 896 | ||
| 897 | /** A deleted agent comes off every team it was on (memberships live in identity). */ | |
| 898 | private async leaveTeams(slug: string, by: User, agentId: string): Promise<void> { | |
| 899 | try { | |
| 900 | const identity = identityClient(this.env.IDENTITY); | |
| 901 | const teams = await identity.agentTeams(slug, agentId); | |
| 902 | await Promise.all( | |
| 903 | teams.map((team) => identity.removeTeamAgent(by, slug, team.slug, agentId).catch((error: unknown) => console.error("agents: a deleted agent was not taken off its team", agentId, team.slug, String(error)))), | |
| 904 | ); | |
| 905 | } catch (error) { | |
| 906 | console.error("agents: a deleted agent's teams were not read", agentId, String(error)); | |
| 907 | } | |
| 908 | } | |
| 909 | ||
| Chat and workspace agents: channels, DMs and named agents you talk to | 910 | /** |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 911 | * Makes the workspace's built-in @g1t if it does not exist yet: an |
| 912 | * ordinary agent row, marked builtin, at version 1. Safe to call on | |
| 913 | * every request; once seen, an isolate does not write again. | |
| 914 | */ | |
| 915 | private async ensureBuiltin(workspaceId: string): Promise<void> { | |
| 916 | await ensureBuiltin(this.db, workspaceId); | |
| 917 | } | |
| 918 | ||
| 919 | /** Internal: the workspace's @g1t, made if need be, for the chat service. */ | |
| 920 | async builtin(a: { workspace: string; workspace_id: string }): Promise<Result<WorkspaceAgent>> { | |
| 921 | if (typeof a?.workspace_id !== "string" || !a.workspace_id) return fail("invalid", "Name the workspace by id."); | |
| 922 | await this.ensureBuiltin(a.workspace_id); | |
| 923 | const now = new Date(); | |
| 924 | const row = await this.db | |
| 925 | .prepare(selectAgents("a.workspace_id = ?3 AND a.builtin = 1")) | |
| 926 | .bind(...periods(now), a.workspace_id) | |
| 927 | .first<Row>(); | |
| 928 | return row ? ok(toAgent(row, now)) : fail("not_found", "This workspace's @g1t could not be made."); | |
| 929 | } | |
| 930 | ||
| 931 | /** | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 932 | * Hands a message to the agent's desk, which answers it in the |
| 933 | * background. Returns as soon as the desk holds it. | |
| 934 | */ | |
| 935 | async deliver(delivery: AgentDelivery): Promise<Result<null>> { | |
| 936 | const fields = ["workspace", "workspace_id", "channel_id", "agent_id", "message_id", "asked_by"] as const; | |
| 937 | if (!delivery || fields.some((field) => typeof delivery[field] !== "string" || !delivery[field])) { | |
| 938 | return fail("invalid", "A delivery names the workspace, channel, agent, message and who asked."); | |
| 939 | } | |
| 940 | if (delivery.channel_kind !== "channel" && delivery.channel_kind !== "dm") return fail("invalid", "channel_kind is channel or dm."); | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 941 | await this.ensureBuiltin(delivery.workspace_id); |
| Chat and workspace agents: channels, DMs and named agents you talk to | 942 | const desk = this.env.DESKS.get(this.env.DESKS.idFromName(delivery.agent_id)); |
| 943 | await desk.take({ ...delivery, hops: Math.max(0, Math.floor(Number(delivery.hops) || 0)), thread_root: delivery.thread_root ?? null }); | |
| 944 | return ok(null); | |
| 945 | } | |
| 946 | ||
| 947 | private versionStatement(agentId: string, version: number, d: Definition, by: string, at: string): D1PreparedStatement { | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 948 | return versionStatement(this.db, agentId, version, d, by, at); |
| Chat and workspace agents: channels, DMs and named agents you talk to | 949 | } |
| 950 | ||
| 951 | /** | |
| 952 | * Records a change to an agent in the workspace's audit log, after the | |
| 953 | * answer. Never fails the change: a log that cannot be written is logged. | |
| 954 | * The events service's audit contract speaks camelCase (Rust's | |
| 955 | * `NewAuditEntry`). | |
| 956 | */ | |
| Agents → Skills is the workspace's skill library in the open SKILL.md format: write a skill in g1t, import a SKILL.md or zip, link a repository whose .g1t/skills folders publish on every push, or save a finished session as a skill once a person reviews it; attach a skill to an agent, a team or the whole workspace, each attachment pinned to a version. Agents see each skill's name and when to use it, and read the rest with use_skill when a request needs it; a skill never adds a permission, and scripts wait for an agent's own computer. The agent skills guide says how. | 957 | /** A change to the skill library, in the audit log as `agents/skills/<name>`. */ |
| 958 | auditSkill(actor: User, workspace: string, action: string, name: string, message: string): void { | |
| 959 | this.audit(actor, workspace, action, `skills/${name}`, message); | |
| 960 | } | |
| 961 | ||
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 962 | private audit( |
| 963 | actor: User, | |
| 964 | workspace: string, | |
| 965 | action: string, | |
| 966 | handle: string, | |
| 967 | message: string, | |
| 968 | area: "agents" | "marketplace" = "agents", | |
| 969 | rule: "owner" | "member" | null = null, | |
| 970 | ): void { | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 971 | const kind = actor.kind === "workspace" ? "workspace" : actor.kind === "agent" ? "agent" : actor.kind === "system" ? "system" : "person"; |
| 972 | const entry = { | |
| 973 | actorKind: kind, | |
| 974 | actor: actor.username, | |
| 975 | actorId: actor.id, | |
| 976 | agent: null, | |
| 977 | onBehalfOf: null, | |
| 978 | runId: null, | |
| 979 | runKind: null, | |
| 980 | credentialId: null, | |
| 981 | action, | |
| 982 | surface: "web", | |
| 983 | workspace: workspace.toLowerCase(), | |
| 984 | repo: null, | |
| 985 | number: null, | |
| 986 | gitRef: null, | |
| The Marketplace is where a workspace finds what to add, reached from Apps: the agent catalog of roles with who already does each job and Add to workspace; every integration with whether it's connected; g1t's own extensions, listed as Soon; and requests, where members ask and owners add or turn down, each told by a notification. Underneath, an extension is a manifest (publisher and tier, source and version from a tag, scopes, the domains its data goes to, its page on g1tusercontent.com, no price) and an install pins its version with an off switch, a budget and the audit log. Apps lists only what was added from the Marketplace, with pins; g1t's own pages stay in their sidebars, and AI Gateway joins Workspace's. The marketplace guide says how. | 987 | path: `${area}/${handle}`, |
| Chat and workspace agents: channels, DMs and named agents you talk to | 988 | outcome: "allowed", |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 989 | rule: rule ?? (area === "marketplace" && action === "request_install" ? "member" : "owner"), |
| Chat and workspace agents: channels, DMs and named agents you talk to | 990 | result: "ok", |
| 991 | message, | |
| 992 | requestId: `req_${crypto.randomUUID()}`, | |
| 993 | }; | |
| 994 | this.defer( | |
| 995 | this.env.EVENTS.fetch("https://service/rpc/audit_record", { | |
| 996 | method: "POST", | |
| 997 | headers: { "content-type": "application/json" }, | |
| 998 | body: JSON.stringify({ entries: [entry] }), | |
| 999 | }) | |
| 1000 | .then((response) => { | |
| 1001 | if (!response.ok) throw new Error(`status ${response.status}`); | |
| 1002 | }) | |
| 1003 | .catch((error: unknown) => console.error("agents: audit entry not recorded", action, handle, String(error))), | |
| 1004 | ); | |
| 1005 | } | |
| 1006 | } | |
| 1007 | ||
| 1008 | /** One RPC method's answer. */ | |
| 1009 | async function answer(service: Agents, method: string, args: any): Promise<Response> { | |
| Agents → Skills is the workspace's skill library in the open SKILL.md format: write a skill in g1t, import a SKILL.md or zip, link a repository whose .g1t/skills folders publish on every push, or save a finished session as a skill once a person reviews it; attach a skill to an agent, a team or the whole workspace, each attachment pinned to a version. Agents see each skill's name and when to use it, and read the rest with use_skill when a request needs it; a skill never adds a permission, and scripts wait for an agent's own computer. The agent skills guide says how. | 1010 | // The skill library (./skill-rpc.ts). |
| 1011 | const skill = await skillRpc(method, args, (a, run) => service.view(a, run), (viewer, workspace, action, name, message) => service.auditSkill(viewer, workspace, action, name, message)); | |
| 1012 | if (skill) return Response.json(skill); | |
| Merge branch 'customize' into main: Customize replaces the Marketplace, the Artifacts trash and spaces changes, the Security-free cleanup, with the desktop app's own title bar kept in the shell. | 1013 | // Customize: marketplaces, plugins and connectors (./customize-rpc.ts). |
| 1014 | const customize = await customizeRpc(method, args, (a, run) => service.view(a, run)); | |
| 1015 | if (customize) return Response.json(customize); | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 1016 | switch (method) { |
| 1017 | case "list": | |
| 1018 | return Response.json(await service.list(args)); | |
| 1019 | case "get": | |
| 1020 | return Response.json(await service.get(args)); | |
| 1021 | case "by_ids": | |
| 1022 | return Response.json(await service.byIds(args)); | |
| 1023 | case "create": | |
| 1024 | return Response.json(await service.create(args)); | |
| 1025 | case "update": | |
| 1026 | return Response.json(await service.update(args)); | |
| 1027 | case "archive": | |
| 1028 | return Response.json(await service.archive(args)); | |
| An archived agent can be found, brought back, or deleted for good. The Agents page ends with Archived, folded, where owners see every archived agent and members their own personal ones, each with when and by whom it was archived and what deleting it would take; Restore brings it back, under a new handle when a live agent has taken its old one, and Delete permanently asks for the handle to be typed and then removes its sessions, replies, memories, routines, versions, drafts, skills and its computer's home, leaving its messages in chat under its saved name, the workspace's spend totals and the audit log. In Chat an archived agent is deactivated the way people expect: never offered as someone to message or mention, never woken, a conversation with it kept in the list, dimmed and marked Archived, open to read with a note in place of the composer and a way to restore it for owners; after deletion the conversation still reads, its messages marked as a deleted agent's. The agents, chat and agent budgets guides say so. | 1029 | case "list_archived": |
| 1030 | return Response.json(await service.listArchived(args)); | |
| 1031 | case "restore": | |
| 1032 | return Response.json(await service.restore(args)); | |
| 1033 | case "destroy": | |
| 1034 | return Response.json(await service.destroy(args)); | |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 1035 | case "draft": |
| 1036 | return Response.json(await service.draft(args)); | |
| 1037 | case "try_draft": | |
| 1038 | return Response.json(await service.tryDraft(args)); | |
| 1039 | case "redraft": | |
| 1040 | return Response.json(await service.redraft(args)); | |
| Each agent has an Abilities tab: g1t's built-ins, always on within the asker's access; its computer, coming; each connected integration's actions one row each, read, import, comment or resolve, with a level for each, alone, alone when the person asked for it, ask first or never, and whose connection it runs on; and MCP servers an owner adds. Reading is alone, writing inside g1t follows today's choices, anything that leaves g1t asks first, and production deploys can't go above ask. The agents service enforces every level: ask first posts a card to allow or deny and parks a session until it's answered, a refusal names its rule in the transcript and the audit log, and a missing ability posts a request to owners. The agent abilities guide says how. | 1041 | case "add_mcp_server": |
| 1042 | return Response.json(await service.addMcpServer(args)); | |
| 1043 | case "remove_mcp_server": | |
| 1044 | return Response.json(await service.removeMcpServer(args)); | |
| 1045 | case "refresh_mcp_server": | |
| 1046 | return Response.json(await service.refreshMcpServer(args)); | |
| Anyone makes an agent by describing it: one box drafts the whole agent (name, job, responsibilities, voice, skills to keep, integrations it needs, model limits and budget) as a card to change and try in a chat beside it before creating it, and an agent changes later by saying what to change, shown as a diff before it's saved. Members can make personal agents, which only they talk to and which spend from their own budget, unless owners turn that off; owners promote one to a workspace agent. The agents guide says how. | 1047 | case "promote": |
| 1048 | return Response.json(await service.promote(args)); | |
| Merge the workspace shell: navigation and phone shell, g1t as orchestrator, agents in roles with audience-checked reads, reactions and custom emoji, live notifications and browser push, the homepage tour (agents 0002, chat 0002) | 1049 | case "builtin": |
| 1050 | return Response.json(await service.builtin(args)); | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 1051 | case "templates": |
| 1052 | return Response.json(TEMPLATES); | |
| 1053 | case "deliver": | |
| 1054 | return Response.json(await service.deliver(args)); | |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 1055 | case "overview": |
| 1056 | return Response.json(await service.view(args, (ctx) => views.overview(ctx))); | |
| 1057 | case "sessions": | |
| 1058 | return Response.json(await service.view(args, (ctx) => views.listSessions(ctx, args))); | |
| 1059 | case "session": | |
| 1060 | return Response.json(await service.view(args, (ctx) => views.sessionDetail(ctx, args.id))); | |
| 1061 | case "stop_session": | |
| 1062 | return Response.json(await service.view(args, (ctx) => views.stopSession(ctx, args.id))); | |
| 1063 | case "approve_session": | |
| 1064 | return Response.json(await service.view(args, (ctx) => views.approveSession(ctx, args.id, args.cap_micros))); | |
| 1065 | case "steer_session": | |
| 1066 | return Response.json(await service.view(args, (ctx) => views.steerSession(ctx, args.id, args.body))); | |
| Every agent can have its own computer. A session that needs one wakes it: a home of its own on g1t cloud, one per agent and never shared, where it runs commands, reads and writes files and keeps what it made, with each session working in its own folder under a shared home; after ten idle minutes it sleeps, its home kept as a snapshot and restored when it wakes, and Reset wipes the home while memory and artifacts stay. Its shell and files are abilities with the usual choices, Alone, Alone when asked, Ask first or Never, offered only inside sessions and never to a chat reply; every command shows on the session with its output, and the agent's new Computer tab shows the state, the disk used of the five gigabytes included, the recent commands, and Wake, Put to sleep and Reset. Machine time counts only while it is awake, on the sandbox lines of the ledger that name the agent and who asked, held to the same spend caps as the session; the disk itself costs nothing in this version. The runner gained a long-lived supervisor that answers the computer's requests inside the container, and the runner service a computer per agent that keeps its snapshot in the agent homes bucket when one is attached, and says so when none is. The REST API and the agent tool can read a computer, wake it, put it to sleep and reset it. The agents, abilities, sessions, runners, billing and deploy guides say how it works and what an operator sets up; pinning a computer to your own runner, its browser and take-over come next. | 1067 | // The agent's own computer (./computer.ts). |
| 1068 | case "computer": | |
| 1069 | return Response.json(await service.view(args, (ctx) => computerView(ctx, args.handle))); | |
| 1070 | case "computer_wake": | |
| 1071 | return Response.json(await service.view(args, (ctx) => wakeComputer(ctx, args.handle))); | |
| 1072 | case "computer_sleep": | |
| 1073 | return Response.json(await service.view(args, (ctx) => sleepComputer(ctx, args.handle))); | |
| 1074 | case "computer_reset": | |
| 1075 | return Response.json(await service.view(args, (ctx) => resetComputer(ctx, args.handle))); | |
| 1076 | case "computer_commands": | |
| 1077 | return Response.json(await service.view(args, (ctx) => computerCommands(ctx, args.handle, args.session_id ?? null))); | |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 1078 | case "memories": |
| 1079 | return Response.json(await service.view(args, (ctx) => views.memories(ctx, args.handle))); | |
| 1080 | case "remember": | |
| 1081 | return Response.json(await service.view(args, (ctx) => views.remember(ctx, args.handle, args.input))); | |
| 1082 | case "update_memory": | |
| 1083 | return Response.json(await service.view(args, (ctx) => views.updateMemory(ctx, args.handle, args.id, args.changes))); | |
| 1084 | case "forget": | |
| 1085 | return Response.json(await service.view(args, (ctx) => views.forget(ctx, args.handle, args.id))); | |
| 1086 | case "routines": | |
| 1087 | return Response.json(await service.view(args, (ctx) => views.routines(ctx, args.handle))); | |
| 1088 | case "save_routine": | |
| 1089 | return Response.json(await service.view(args, (ctx) => views.saveRoutine(ctx, args.handle, args.input, args.id ?? null))); | |
| 1090 | case "delete_routine": | |
| 1091 | return Response.json(await service.view(args, (ctx) => views.deleteRoutine(ctx, args.handle, args.id))); | |
| 1092 | case "run_routine": | |
| 1093 | return Response.json(await service.view(args, (ctx) => views.runRoutineNow(ctx, args.handle, args.id))); | |
| 1094 | case "spend": | |
| The Spend page says what a workspace spent, where it went and the budgets that hold it, in Workspace under Money: spent at price, agents' spend and the spend limit; by day; by agent, person, channel, model, kind of work and product; budgets from the workspace down to each person, agent and task, with what happens at 100%; the costliest tasks, each with a receipt of every session at the provider's price and what was charged; and how it's priced. Everyone sees what agents spent for them, and owners and billing managers the whole workspace; each person can have a monthly budget agents keep to, and the spend pill in the top bar shows your month or the workspace's against its limit. The spend guide says how. | 1095 | return Response.json(await service.view(args, (ctx) => views.spend(ctx, args.handle ?? null, { period: args.period, person: args.person }))); |
| 1096 | case "person_budgets": | |
| 1097 | return Response.json(await service.view(args, (ctx) => views.personBudgetsView(ctx))); | |
| 1098 | case "set_person_budget": | |
| 1099 | return Response.json(await service.view(args, (ctx) => views.setPersonBudget(ctx, args.username, args.monthly_micros))); | |
| People and teams are front and centre: one directory of people and agents with presence, local time, titles, teams and what each owns; profiles with manager and reports and the agents they work with; an org chart with each team's agents beside the person who leads it; and teams of any mix, with a lead, a channel, a budget agents keep to and the agents on them. Every agent is told its teams each turn (who leads, who owns what, who's around and who to page), and the team page shows exactly what. Member management is Members and invites; the people and teams guide says how. | 1100 | case "team_context": |
| 1101 | return Response.json(await service.view(args, (ctx) => views.teamContext(ctx, args.handle))); | |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 1102 | case "activity": |
| 1103 | return Response.json(await service.view(args, (ctx) => views.activity(ctx, args.handle))); | |
| 1104 | case "versions": | |
| 1105 | return Response.json(await service.view(args, (ctx) => views.versions(ctx, args.handle))); | |
| Cards you act on in chat; agents comment and review as themselves; names shown cleanly; commits on the calendar | 1106 | case "card_action": |
| 1107 | return Response.json(await service.cardAction(args)); | |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 1108 | case "policy": |
| 1109 | return Response.json(await service.view(args, (ctx) => views.policy(ctx))); | |
| The Marketplace is where a workspace finds what to add, reached from Apps: the agent catalog of roles with who already does each job and Add to workspace; every integration with whether it's connected; g1t's own extensions, listed as Soon; and requests, where members ask and owners add or turn down, each told by a notification. Underneath, an extension is a manifest (publisher and tier, source and version from a tag, scopes, the domains its data goes to, its page on g1tusercontent.com, no price) and an install pins its version with an off switch, a budget and the audit log. Apps lists only what was added from the Marketplace, with pins; g1t's own pages stay in their sidebars, and AI Gateway joins Workspace's. The marketplace guide says how. | 1110 | case "install_requests": |
| 1111 | return Response.json(await service.installRequests(args)); | |
| 1112 | case "request_install": | |
| 1113 | return Response.json(await service.requestInstall(args)); | |
| 1114 | case "resolve_install_request": | |
| 1115 | return Response.json(await service.resolveInstallRequest(args)); | |
| Pricing says it plainly: models at the provider's price, the agent rate for what g1t runs around every model call (the gateway, secrets, routing, context and pass-through to your own provider, so your own keys too), everything else at cost plus 20%, your own runners free, no seats; Security and quality comes with the plan with no separate fee, and live activations end. Each agent has an effort setting, Auto to Max, with what a typical task has cost at each level, and Spend's Spend less, keep quality suggests a lower level only when the agent's own past work shows quality held, to apply or dismiss. The pricing, spend and agents guides say how. | 1116 | case "effort_costs": |
| 1117 | return Response.json(await service.effortCosts(args)); | |
| 1118 | case "recommendations": | |
| 1119 | return Response.json(await service.recommendations(args)); | |
| 1120 | case "resolve_recommendation": | |
| 1121 | return Response.json(await service.resolveRecommendation(args)); | |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 1122 | case "set_policy": |
| 1123 | return Response.json(await service.view(args, (ctx) => views.setPolicy(ctx, args.policy))); | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 1124 | default: |
| 1125 | return new Response("Unknown method\n", { status: 404 }); | |
| 1126 | } | |
| 1127 | } | |
| 1128 | ||
| 1129 | export default { | |
| 1130 | async fetch(request: Request, env: Env, ctx: ExecutionContext): Promise<Response> { | |
| 1131 | const match = new URL(request.url).pathname.match(/^\/rpc\/([a-z_]+)$/); | |
| 1132 | if (request.method !== "POST" || !match) return new Response("Not found\n", { status: 404 }); | |
| 1133 | // A replica near the caller when it asks for one (@g1t/contracts d1.ts). | |
| 1134 | const opened = openD1(env.DB, request); | |
| 1135 | const service = new Agents(Object.create(env, { DB: { value: opened.db } }) as Env, (work) => ctx.waitUntil(work)); | |
| 1136 | const args = (await request.json().catch(() => ({}))) as any; | |
| 1137 | return opened.finish(await answer(service, match[1], args)); | |
| 1138 | }, | |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 1139 | |
| @g1t welcomes you when you join a workspace, and that is the one introduction any agent makes. Identity now says when a person becomes a member, workspace.member_joined, by invitation, by an invite code, or by creating the workspace, and webhooks can hear it; the agents service hears it too, and for anyone who joined, never the creator, the workspace's @g1t opens your direct message and says in its own voice who it is, what to ask it, how agents answer, and where things are, once per person per workspace and charged to @g1t like any reply, with a fixed welcome when no model can be used. The agents, workspaces, authentication and webhooks guides say so. | 1140 | /** Events routines run on, and people joining whom @g1t welcomes, from the events service (SUBSCRIBER_AGENTS). */ |
| Routines run when something happens: a pull request ready for review or merged, checks or a deploy failing, an issue opened | 1141 | async queue(batch: MessageBatch<unknown>, env: Env): Promise<void> { |
| Agents → Skills is the workspace's skill library in the open SKILL.md format: write a skill in g1t, import a SKILL.md or zip, link a repository whose .g1t/skills folders publish on every push, or save a finished session as a skill once a person reviews it; attach a skill to an agent, a team or the whole workspace, each attachment pinned to a version. Agents see each skill's name and when to use it, and read the rest with use_skill when a request needs it; a skill never adds a permission, and scripts wait for an agent's own computer. The agent skills guide says how. | 1142 | const events = batch.messages.map((message) => message.body as G1tEvent); |
| 1143 | // A push to a default branch: skill libraries that follow that repository read it again (./skill-library.ts). | |
| 1144 | const pushes = events.flatMap((event) => (event?.type === "git.push" && event.data?.defaultBranch && event.data.repoId ? [{ repoId: event.data.repoId }] : [])); | |
| @g1t welcomes you when you join a workspace, and that is the one introduction any agent makes. Identity now says when a person becomes a member, workspace.member_joined, by invitation, by an invite code, or by creating the workspace, and webhooks can hear it; the agents service hears it too, and for anyone who joined, never the creator, the workspace's @g1t opens your direct message and says in its own voice who it is, what to ask it, how agents answer, and where things are, once per person per workspace and charged to @g1t like any reply, with a fixed welcome when no model can be used. The agents, workspaces, authentication and webhooks guides say so. | 1145 | await Promise.all([ |
| 1146 | onEvents(env as unknown as SessionEnv, events.filter((event) => event?.type !== "git.push")), | |
| 1147 | pushes.length ? skillPushes(env as unknown as SessionEnv, pushes) : null, | |
| 1148 | // A person joined a workspace: its @g1t says welcome, once (./welcome.ts). | |
| 1149 | welcome(env as unknown as SessionEnv, events).catch((error: unknown) => console.error("agents: welcomes were not sent", String(error))), | |
| 1150 | ]); | |
| Routines run when something happens: a pull request ready for review or merged, checks or a deploy failing, an issue opened | 1151 | batch.ackAll(); |
| 1152 | }, | |
| 1153 | ||
| An agent is on teams the way a person is: through team membership, with no team or department of its own. Add it to teams when you create it, from its profile or from the team's page; the Agents sidebar groups agents by the teams they're on, with those on none last; what an agent is told about its teammates, team budgets, skills reach, Spend's By team and the People pages all read memberships. Agents with an old team name matching a team are moved onto it once; the rest are left off a team. Templates name a role only. The agents, teams and people guides say how. | 1154 | /** Every few minutes: routines that are due, session steps a desk lost, the spend check, and the one-time team move. */ |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 1155 | async scheduled(_controller: ScheduledController, env: Env, ctx: ExecutionContext): Promise<void> { |
| 1156 | const sessions = env as unknown as SessionEnv; | |
| 1157 | ctx.waitUntil( | |
| 1158 | Promise.all([ | |
| 1159 | runDue(sessions).catch((error: unknown) => console.error("agents: routines did not run", String(error))), | |
| 1160 | sweep(sessions).catch((error: unknown) => console.error("agents: the session sweep failed", String(error))), | |
| Pricing says it plainly: models at the provider's price, the agent rate for what g1t runs around every model call (the gateway, secrets, routing, context and pass-through to your own provider, so your own keys too), everything else at cost plus 20%, your own runners free, no seats; Security and quality comes with the plan with no separate fee, and live activations end. Each agent has an effort setting, Auto to Max, with what a typical task has cost at each level, and Spend's Spend less, keep quality suggests a lower level only when the agent's own past work shows quality held, to apply or dismiss. The pricing, spend and agents guides say how. | 1161 | // Spend less, keep quality: each workspace checked weekly (src/recommend.ts). |
| 1162 | checkDue(env.DB).catch((error: unknown) => console.error("agents: the spend check failed", String(error))), | |
| An agent is on teams the way a person is: through team membership, with no team or department of its own. Add it to teams when you create it, from its profile or from the team's page; the Agents sidebar groups agents by the teams they're on, with those on none last; what an agent is told about its teammates, team budgets, skills reach, Spend's By team and the People pages all read memberships. Agents with an old team name matching a team are moved onto it once; the rest are left off a team. Templates name a role only. The agents, teams and people guides say how. | 1163 | // Once: agents' own old teams become team memberships (src/team-move.ts); after that, one cheap read. |
| 1164 | moveAgentTeams(env.DB, (id, agents) => identityClient(env.IDENTITY).adoptAgentTeams(id, agents)).catch((error: unknown) => console.error("agents: moving agents' teams failed", String(error))), | |
| Agents work in sessions: bounded, visible, steerable work spun off from chat, with subagents and colleagues in a tree paid by its root; memory with sources and scopes; routines; a workspace budget for every agent; agents file issues for whoever asked | 1165 | ]), |
| 1166 | ); | |
| 1167 | }, | |
| Chat and workspace agents: channels, DMs and named agents you talk to | 1168 | } satisfies ExportedHandler<Env>; |
This file's history is long; its oldest lines are credited to the oldest commit read.