Documentation
¶
Overview ¶
Package analyze contains the analyze-kind AI agent: operator- triggered, tool-using, single-incident investigation. The agent runs only via the admin endpoint `POST /api/admin/incidents/:id/analyze` and writes its output to the analyses storage blob — it never fans out to notification channels (that contract is enforced by the import-graph guard test).
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func BuildUserPrompt ¶
func BuildUserPrompt(s core.AnalyzeIncidentSnapshot) string
BuildUserPrompt renders the snapshot as the first user message sent to the model. The tool catalog is attached via Eino's ToolInfo schema, not in the user message.
func ParseFinding ¶
ParseFinding decodes the model's final assistant message into an AIFinding. Like the detect parser it tolerates leading prose and code fences. Analyze-specific fields (root_cause_hypotheses, evidence, related_pattern_ids, next_steps) are passed through because AIFinding already declares them.
func PromptOrder ¶
func PromptOrder() []string
PromptOrder returns the canonical fragment order so the admin endpoint can echo the list back to the UI alongside the assembled prompt.
func SystemPrompt ¶
func SystemPrompt() string
SystemPrompt returns the assembled system prompt sent on every analyze call. Exposed so the admin API can render it for operators.
Types ¶
type Agent ¶
type Agent struct {
// contains filtered or unexported fields
}
Agent is the analyze-kind AIAgent. The investigation loop, tool fan-out, and per-call audit all run on Eino's pre-built ReAct agent (flow/agent/react). The struct binds the resolved per-task config, the ReAct agent (which already owns the tool-calling chat model plus a compose.ToolsNode — sequential by default, concurrent when analyze.parallel_tools is set), and an in-memory registry of the read-only tools (kept for introspection / allow-list assertions).
The struct deliberately has NO Emitter / Notifier / Sender / Dispatcher field. The import-graph guard test in agent_test.go asserts this so future edits cannot silently turn analyze into a notification path.
func New ¶
func New(ctx context.Context, cfg config.AgentAIConfig, tools []core.AnalyzeTool, opts Options) (*Agent, error)
New constructs an analyze Agent. cfg must already be resolved for the analyze task (see config.AgentAIConfig.Resolve). Every tool in the supplied list is registered with the agent.
type Options ¶
type Options struct {
HTTPClient *http.Client
BaseURL string
Timeout time.Duration
// AuthKeyFunc is an OPTIONAL per-request Authorization override passed
// straight to the chat model's transport. Nil (the OSS default) leaves
// the YAML-keyed header untouched.
AuthKeyFunc func(ctx context.Context) (key string, ok bool)
// Runtime folds optional runtime overrides (provider / enabled / key
// state) into the model holder's rebuild signature. The zero value (the
// OSS default) pins the configured provider and builds the agent once.
// Ignored when ChatModel is set (a fixed override is never rebuilt).
Runtime einowrap.RuntimeAI
// ChatModel overrides the Eino tool-calling chat model. When
// non-nil the agent skips dialing OpenAI; tests pass a fake. The
// ReAct agent binds tools onto it via WithTools.
ChatModel model.ToolCallingChatModel
// ToolTimeout caps a single tool dispatch. Zero applies the built-in
// defaultToolTimeout; a negative value disables the per-tool cap.
ToolTimeout time.Duration
// ParallelTools runs multiple tool calls emitted in one model turn
// concurrently. False (the default) dispatches them sequentially.
ParallelTools bool
}
Options is the constructor-side bag for test plumbing.
Source Files
Directories