Documentation
¶
Overview ¶
Package prompt renders the agent's system prompt at run time from the agent's live registrations (tools, connections, MCPs, topics, webhooks, crons, routes) plus the small slice of platform data Airlock supplies at sync (dashboard/route URLs, the sibling address book). The same template was previously rendered server-side in airlock/prompt; moving it here lets the per-run caller filter (caller access level, per-user sibling visibility) take effect without exploding the on-wire payload into N variants.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func Render ¶
Render renders the system prompt for the given caller access. Pass caller as "admin", "user", or "public"; empty means unfiltered admin variant (used by tests). The caller value is also stamped onto data.CallerAccess so templates can branch on tier (e.g. `{{ if .CallerAccess.IsPublic }}…{{ end }}`).
Types ¶
type AgentData ¶
type AgentData struct {
AgentDashboardURL string
AgentRouteURL string
CallerAccess Caller
Capabilities Capabilities
SupportedModalities Modalities
Tools []ToolInfo
Connections []ConnInfo
Topics []TopicInfo
Webhooks []WebhookInfo
Crons []CronInfo
Routes []RouteInfo
MCPServers []MCPServerStatus
Siblings []SiblingInfo
Directories []DirInfo
ExecEndpoints []ExecEndpointInfo
// Per-turn environment, rendered into the <env> block. Each is set
// explicitly by the dispatch path (never inferred) and omitted from
// the block when empty.
Date string
Platform string // web | telegram | discord | a2a
UserName string
UserEmail string
Conversation string
// DirectTools is true when the run exposes each capability as its
// own typed LLM tool instead of one `run_js` binding. The template
// branches on this to skip JS-flavoured sections (the TypeScript
// manifest, `## JavaScript environment`, `var` / `let` guidance, the
// namespaced binding docs) and emit a short tool-listing block
// instead. Schemas reach the model via the tool-calling protocol,
// not the system prompt.
DirectTools bool
}
AgentData is the template input. Lists are pre-filtered for the caller's access level by Render before the template runs, so the {{ range }} loops don't have to think about access. The CallerAccess field is also exposed so the template CAN branch on it when the difference isn't "include/exclude an item" but "write a different sentence" — e.g. a public-tier preamble or admin-only guidance.
type Caller ¶
type Caller string
Caller wraps the raw access string with template-friendly predicate methods so authors write `{{ if .CallerAccess.IsUser }}` instead of `{{ if eq .CallerAccess "user" }}`.
Empty Caller (zero value) is treated as Admin — matches the unfiltered "render everything" mode Render uses when caller is passed as "".
func (Caller) AtLeastUser ¶
AtLeastUser reports user-or-above. Useful for "everyone except public" branches.
type Capabilities ¶
type Capabilities struct {
Vision bool
Transcription bool
Speech bool
Embedding bool
Image bool
Search bool
}
Capabilities mirrors agentsdk.Capabilities — duplicated here to avoid an import cycle. buildPromptData copies field-for-field.
type DirInfo ¶
type DirInfo struct {
Path string
Description string
LLMHint string
Read string
Write string
List string
// Scope, when non-empty, indicates the directory is per-context.
// Values: "user", "conv", "run". The template renders an extra
// note so the LLM understands paths it returns will include a
// scope key segment (e.g. `<dir>/user-<id>/...`).
Scope string
}
DirInfo describes one registered storage directory for the prompt. Each cap is shown independently — a caller may, for example, have list+read but not write. The template uses these to tell the LLM which S3 paths it can touch on this run.
type ExecEndpointInfo ¶
ExecEndpointInfo carries one registered exec endpoint for the prompt. Bind-time access gating happens in vm.go; the prompt block also access-filters via filterExecEndpoints below so a non-admin caller never sees an admin-only endpoint listed.
type MCPServerStatus ¶
type MCPServerStatus struct {
Slug string
Name string
Status string // "connected, 5 tools" or "requires authentication"
Access string
// Description is the server-level usage hint the remote MCP server
// advertised via initialize `instructions`. Empty when unset.
Description string
Tools []ToolInfo
}
MCPServerStatus carries the per-server status line + (when authorized) the discovered tool schemas the template renders into a typed `declare const mcp_{slug}: {...}` block.
type Modalities ¶
type Modalities []string
Modalities is the list of input types the chat model accepts. Wrapped in a type so the template can ask `.SupportedModalities.HasImage` instead of doing string comparisons.
func (Modalities) Has ¶
func (m Modalities) Has(modality string) bool
func (Modalities) HasAudio ¶
func (m Modalities) HasAudio() bool
func (Modalities) HasImage ¶
func (m Modalities) HasImage() bool
HasImage / HasPDF / HasAudio / HasVideo are template-friendly predicates for the modality strings emitted by sol's provider.GetModalities. "text" is always implied so there's no matching predicate for it.
func (Modalities) HasPDF ¶
func (m Modalities) HasPDF() bool
func (Modalities) HasVideo ¶
func (m Modalities) HasVideo() bool
type SiblingInfo ¶
type SiblingInfo struct {
ID uuid.UUID
Slug string
Name string
Description string
Tools []ToolInfo
}
SiblingInfo carries one sibling agent record for prompt rendering. ID is the canonical, rename-safe identifier; Slug is the human-readable binding name (`agent_<slug>`). Tools is the sibling's published tool schemas (synced server-side).
type ToolInfo ¶
type ToolInfo struct {
Name string
Description string
LLMHint string
Access string
InputSchema json.RawMessage
OutputSchema json.RawMessage
}
ToolInfo carries the hydrated tool record for prompt rendering.
Source Files