| import type Anthropic from "@anthropic-ai/sdk"; |
|
|
| export const GROUPING_SYSTEM_PROMPT = [ |
| "You decide whether a new error issue belongs to one of the open incidents in a project.", |
| "An incident represents one underlying root cause; multiple distinct error symptoms can belong to the same incident if they share that root cause (e.g. a database outage causing connection timeouts, query failures, and downstream HTTP 500s — all one incident).", |
| "Default to 'standalone'. Return 'join' only when there is positive evidence the new issue and the candidate incident share a single underlying root cause.", |
| "Surface-level similarity is NOT sufficient evidence. Examples that do NOT justify joining: same HTTP status code, same generic exception class with no shared context, similar wording in unrelated stack traces.", |
| "Never join across different known deployment.environment values. Never join when one side is clearly localhost/dev-worktree traffic and the other side is production/internal-host traffic.", |
| "Examples that DO justify joining: matching upstream dependency name in both stacks; both errors observed within the same trace tree; one error is the canonical fault and the other is a documented downstream symptom of it.", |
| "Identical trace ids are strong evidence that two symptoms happened in the same request path, but still explain why they share one root cause.", |
| "Start by calling list_incident_titles or list_incident_facets to orient yourself. Then use search_incidents and inspect_incident for any plausible join target.", |
| "inspect_incident can include linked issues, representative samples, and the latest agent investigation result. Prefer those structured details over title similarity.", |
| "Do not assume an unseen incident is a join target.", |
| "When you are done, call decide_grouping exactly once. Do not write a text reply.", |
| "When unsure, return 'standalone'.", |
| ].join("\n"); |
|
|
| const LIST_INCIDENT_TITLES_TOOL: Anthropic.Messages.Tool = { |
| name: "list_incident_titles", |
| description: |
| "Lists candidate incident titles and compact metadata so you can see whether any open " + |
| "incident is even close before inspecting details.", |
| input_schema: { |
| type: "object", |
| properties: { |
| service: { type: "string", description: "Optional exact service filter." }, |
| environment: { |
| type: "string", |
| description: "Optional exact deployment.environment filter.", |
| }, |
| limit: { |
| type: "number", |
| description: "Maximum rows. Defaults to all candidates; capped at 200.", |
| }, |
| }, |
| }, |
| }; |
|
|
| const LIST_INCIDENT_FACETS_TOOL: Anthropic.Messages.Tool = { |
| name: "list_incident_facets", |
| description: |
| "Returns counts for structured facets across candidate incidents: services, environments, " + |
| "exception types, endpoint hosts, endpoint kinds, and investigation states. Use this to " + |
| "slice the candidate set before inspecting individual incidents.", |
| input_schema: { |
| type: "object", |
| properties: {}, |
| }, |
| }; |
|
|
| const SEARCH_INCIDENTS_TOOL: Anthropic.Messages.Tool = { |
| name: "search_incidents", |
| description: |
| "Searches the open incident candidates for likely join targets. Use terms from the new issue " + |
| "such as dependency name, endpoint host, exception type, stack frame, service, trace id, or " + |
| "environment. Returns compact previews only.", |
| input_schema: { |
| type: "object", |
| properties: { |
| query: { |
| type: "string", |
| description: "Search terms, e.g. 'clickhouse ECONNRESET worker' or 'localhost 19019'.", |
| }, |
| service: { type: "string", description: "Optional exact service filter." }, |
| environment: { |
| type: "string", |
| description: "Optional exact deployment.environment filter.", |
| }, |
| limit: { |
| type: "number", |
| description: "Maximum results to return. Defaults to 10; capped at 25.", |
| }, |
| }, |
| required: ["query"], |
| }, |
| }; |
|
|
| const INSPECT_INCIDENT_TOOL: Anthropic.Messages.Tool = { |
| name: "inspect_incident", |
| description: |
| "Returns the full grouping context for one candidate incident by id. Use this before joining.", |
| input_schema: { |
| type: "object", |
| properties: { |
| incident_id: { |
| type: "string", |
| description: "Candidate incident id returned by search_incidents.", |
| }, |
| }, |
| required: ["incident_id"], |
| }, |
| }; |
|
|
| const DECIDE_GROUPING_TOOL: Anthropic.Messages.Tool = { |
| name: "decide_grouping", |
| description: |
| "Final answer. Join only when inspected evidence shows the new issue and candidate incident " + |
| "share a single underlying root cause. Otherwise choose standalone.", |
| input_schema: { |
| type: "object", |
| properties: { |
| decision: { type: "string", enum: ["join", "standalone"] }, |
| incidentId: { |
| type: "string", |
| description: "Required only when decision='join'. Must be one inspected candidate id.", |
| }, |
| evidence: { |
| type: "string", |
| description: |
| "For joins, >=20 chars explaining the shared root cause. For standalone, a short reason.", |
| }, |
| }, |
| required: ["decision"], |
| }, |
| }; |
|
|
| export const GROUPING_TOOLS: Anthropic.Messages.Tool[] = [ |
| LIST_INCIDENT_TITLES_TOOL, |
| LIST_INCIDENT_FACETS_TOOL, |
| SEARCH_INCIDENTS_TOOL, |
| INSPECT_INCIDENT_TOOL, |
| DECIDE_GROUPING_TOOL, |
| ]; |
|
|
