Skip to content
POST/api/v1/debug-agent

Start a debug-agent session

Analyses ONE failing trace with an LLM and returns a proposed fix — not an interactive session. Requires orgId + projectId (both verified, and the project must belong to the org) and `traceId`; context comes from the stored trace (`traces`, falling back to `gateway_logs`) plus its `eval_results` scorer rows, or from an `inlineContext` body when the trace lives outside EvalGuard. 400 NO_FAILURES if nothing failed. Opens a `debug_sessions` row, makes a single OpenAI chat-completion call — `analyzerProvider` must be `openai`, `analyzerModel` defaults to gpt-4o-mini, 30s timeout — billed to the org's BYOK OpenAI key when stored, else the server key, and 502 ANALYZER_FAILED on error. Returns 201 with `{sessionId, fixKind, confidence, rationale, suggestedFix, analyzerModel, analyzerCostUsd}`. requiredRole: member. Rate limit 10/min.

Authentication

Send Authorization: Bearer YOUR_API_KEY on every request. Generate API keys at /dashboard/settings/api-keys.

Response

202 example

{
  "session_id": "00000000-0000-0000-0000-000000000000"
}

All status codes

201Created.
202Session started.
400(no description)
401(no description)
403Forbidden — the credential is valid but lacks the API-key scope, member role, or plan entitlement this operation requires.
404Not Found — PROJECT_NOT_FOUND, TRACE_CONTEXT_FAILED.
409Conflict — CONFLICT.
429(no description)
500Internal Server Error — DB_ERROR.
502Bad Gateway — ANALYZER_FAILED.

Code samples

cURL

curl -X POST \
  https://evalguard.ai/api/v1/debug-agent \
  -H "Authorization: Bearer $EVALGUARD_API_KEY"

TypeScript

// The TypeScript SDK (@evalguard/sdk) exposes TYPED methods — runEval,
// getEval, runSecurityScan, checkFirewall, … — not a generic request().
// For an arbitrary endpoint, call it directly:

const res = await fetch("https://evalguard.ai/api/v1/debug-agent", {
  method: "POST",
  headers: { Authorization: `Bearer ${process.env.EVALGUARD_API_KEY}` },
});
console.log(res.status, await res.json());

Python

# The Python SDK (pip install evalguardai) exposes TYPED methods on
# EvalGuardClient — run_eval, get_eval, … — not a generic request().
# For an arbitrary endpoint, call it directly:

import os
import requests

headers = {"Authorization": f"Bearer {os.environ['EVALGUARD_API_KEY']}"}

response = requests.request("POST", "https://evalguard.ai/api/v1/debug-agent", headers=headers)
print(response.status_code, response.json())

Go

package main

import (
	"context"
	"fmt"
	"net/http"
	"os"
)

func main() {
	req, _ := http.NewRequestWithContext(context.Background(), "POST", "https://evalguard.ai/api/v1/debug-agent", nil)
	req.Header.Set("Authorization", "Bearer "+os.Getenv("EVALGUARD_API_KEY"))
	resp, err := http.DefaultClient.Do(req)
	if err != nil { panic(err) }
	defer resp.Body.Close()
	fmt.Println(resp.Status)
}

Errors

400401403404409429500502

Other Evals endpoints