docs(agent): fix stale runtime comments about token injection

amStartTemplateAgent no longer claims OPENAI_API_KEY is uninjected or that
access_token is left empty for a future V2 path: the handler now mints +
injects both OPENAI_API_KEY and the per-agent AGENT_ACCESS_TOKEN (forwarded
as-is), and HM returns its own access_token to the client.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

heicode/controller/agent_template_runtime.go

@@ -160,9 +160,9 @@ func amStartTemplateAgent(ctx context.Context, args amStartArgs) (amStartResult,
 	}
 	env["MODEL_NAME"] = modelName
 	env["OPENAI_BASE_URL"] = publicV1BaseURL()
-	// NOTE: OPENAI_API_KEY is intentionally NOT injected here — the desktop
-	// client passes its own api_key per A2A request (CODING_A2A §7). A future
-	// V2-decrypt path may hand the agent a scoped key instead.
+	// NOTE: OPENAI_API_KEY and AGENT_ACCESS_TOKEN are already in args.Env (the
+	// handler mints + injects them before calling us) and are forwarded as-is by
+	// the copy loop above. We don't add model/auth keys here.
 
 	// env may carry plaintext secrets (db password, blob key…) — warn on http.
 	base := agentRuntimeClientConfigForMode(agentRuntimeModeAgent).BaseURL
@@ -188,8 +188,10 @@ func amStartTemplateAgent(ctx context.Context, args amStartArgs) (amStartResult,
 		return amStartResult{}, err
 	}
 	// AM exposes the agent via access_info.{domain,external_ip}; the instance id
-	// is name/namespace. access_token is not issued by AM (auth = per-request
-	// api_key); left empty until the V2-decrypt path lands.
+	// is name/namespace. AM's access_token (if any) is ignored — HM mints its own
+	// per-agent access token (AGENT_ACCESS_TOKEN) and returns that to the client;
+	// AM authorizes callers by comparing the X-Agent-Access-Token header to its
+	// env AGENT_ACCESS_TOKEN locally.
 	subdomain := stringFromMap(data, "subdomain", "address", "url")
 	if subdomain == "" {
 		if ai, ok := data["access_info"].(map[string]any); ok {