Documentation
¶
Overview ¶
Package mock ships two credential-free LLM providers and a recording wrapper that pair for offline testing of agent flows:
"echo" returns the user's last message as the model response. Zero config; useful for "does the binary boot?" smoke tests.
"scripted" replays a JSONL transcript turn-by-turn. Pair it with a recording captured against a real provider to exercise the agent loop without burning API quota.
NewRecorder wraps any model.LLM and appends each turn (request + response stream) to an io.Writer as JSONL. Recorded transcripts are consumed by the scripted provider via the same shared RecordedTurn type defined in format.go.
Tool execution at replay time uses the live environment, so the scripted provider faithfully replays the LLM side but not the wider tool surface — fine for testing prompt construction and loop shape, not for bit-exact session reproduction.
Index ¶
Constants ¶
const ( ProviderEcho = "echo" ProviderScripted = "scripted" )
Provider names registered by this package.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type Provider ¶
type Provider struct {
// contains filtered or unexported fields
}
Provider is the mock implementation of models.Provider. It hands back either one shared model.LLM (echo, NewScripted) or a freshly built one per call (NewScriptedPerCall) — the modelID argument is ignored either way, because mocks aren't model-specific.
func NewEcho ¶
func NewEcho() *Provider
NewEcho returns an echo Provider directly. Useful for library callers that want the mock without going through models.Resolve.
func NewScripted ¶
NewScripted returns a scripted Provider that plays back the JSONL transcript at path. strict toggles request-shape validation per turn (see scriptedLLM for details).
Every Model call gets the SAME replay, so the transcript is the whole conversation for one agent. For a fan-out where several agents each replay the transcript from its own first turn, use NewScriptedPerCall.
func NewScriptedPerCall ¶ added in v2.9.0
NewScriptedPerCall returns a scripted Provider that builds a NEW replay of the transcript at path on every Model call, so each caller starts at turn 0 with a cursor of its own.
This is the constructor for hermetic fan-out. A background.Manager asks its provider for one LLM per spawn, so with NewScripted three concurrent children share a single cursor and interleave through one script: child A gets recorded turn 0, child B turn 1, child C turn 2, and which is which depends on goroutine scheduling. The cursor is mutex-guarded, so this is not a data race — which makes it worse, not better: -race stays silent while the replay is nondeterministic. With NewScriptedPerCall, every child replays the same script from the top, which is almost always what a fan-out fixture means.
The transcript is read and parsed once, here, so a malformed one fails at construction rather than at the first spawn, and an edit to the file mid-run cannot change what a later spawn replays. Each Model call then re-parses those same bytes rather than sharing the decoded turns: recorded responses are pointers handed straight into the agent loop, and concurrent children must not be handed the same ones.
func (*Provider) Model ¶
Model returns an LLM to run against. The modelID is accepted for interface compatibility but ignored; mock providers don't differentiate by model name.
Whether repeated calls share one LLM depends on the constructor: NewEcho and NewScripted return the same instance every time, NewScriptedPerCall a new one per call.
Source Files